Glossary

users or cases are split at random between two versions to measure a difference.

From Evaluating in production: shadow mode, A/B tests and user feedback

the open protocol SAP uses to let Joule call an agent you host.

From Agent frameworks compared

an SAP API meant for integration with other applications, such as API_SALES_ORDER_SRV.

From Calling your first SAP API

SAP's programming language for business applications.

From Set up for Unit 13: ABAP development tools and a coding agent

SAP's modern ABAP development model, which SAP calls the clean-core variant of the language.

From Coding agents for SAP developers

SAP's tool that runs code checks, including S/4HANA readiness checks.

From AI in S/4HANA migration projects

the assistant declines to answer, for example "I can't find that in the documents".

From Measuring hallucinations and groundedness

the share of AI proposals users keep unchanged.

From Embedded AI across SAP applications

a filter that limits results to what the user is allowed to see.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

a tag on a copied document, such as sales organization or confidentiality, used to filter search results.

From Grounding on SAP data without breaking authorizations: live queries, access labels and the late check

a rule saying which tasks, alone or together, are risky; SAP Access Control distinguishes SoD, critical action and critical permission risks.

From Roles and authorization design across modules

a cost or revenue booked in the period it belongs to, before the invoice arrives or is issued.

From Record to report, end to end

share of all cases the model got right.

From Classification and the metrics that matter

A custom operation on a service, such as "explain this order"

From CAP and side-by-side extensions for AI

a wrapper around an OData API so Joule skills can call it.

From An agent that calls a RAP service

the "twist" in each unit that lets a network draw bends and corners. ReLU is the most common.

From Neural networks from scratch

the date goods issue was actually posted.

From Predicting late deliveries with SAP data

the connector that lets a flow talk to a system. The AI adapter talks to the generative AI hub or OpenAI.

From Calling AI from SAP Integration Suite

the part of the metric the proposed solution can actually change.

From Problem framing: from 'we need AI' to a measurable problem

people using a new tool in real work, well enough for the business result to appear.

From Adoption and change management for AI

SAP's tools for writing ABAP in Eclipse or VS Code.

From Set up for Unit 13: ABAP development tools and a coding agent

an AI system that can take actions through tools, not just answer questions.

From What an SAP forward deployed engineer does

a short public description of an agent: what it does and how to reach it.

From Joule agents and agent orchestration in SAP

a library that supplies the agent loop, tool definitions, state and often approvals and tracing.

From Agent frameworks compared

SAP's governed entry point through which outside agents call Joule agents over A2A.

From Joule agents and agent orchestration in SAP

decide, act, observe, repeat. Also called the run loop or the agentic loop.

From Agents from first principles

splitting a task across agents, running them in order and combining the results.

From Joule agents and agent orchestration in SAP

one unit of agent work that SAP counts for billing.

From Joule agents and agent orchestration in SAP

an open protocol that lets one AI agent hand a task to another.

From What SAP TechEd 2026 changed for AI builders

how often the judge's verdict matches the expert's.

From Set up for Unit 8: evaluation tools

Regulation (EU) 2024/1689, the EU law on AI systems.

From The EU AI Act: what SAP customers need to do and when

a checkpoint in front of models, tools and agents that authenticates, authorizes, limits, meters, filters, routes and logs every call.

From AI gateways: governing model and agent traffic

the decisions, roles and records that control which AI a company runs and how.

From AI governance for SAP programs

an assistant in SAP AI Agent Hub that drafts EU AI Act assessments for human review.

From What SAP TechEd 2026 changed for AI builders

the duty to take measures that help staff who work with AI understand it.

From The EU AI Act: what SAP customers need to do and when

the company-wide set of policies and processes for AI, as defined by ISO/IEC 42001.

From AI governance for SAP programs

the AI backlog placed in quarters, under capacity, dependency and date constraints.

From An AI roadmap and operating model for an SAP customer

the SuccessFactors admin tool where AI features are switched on.

From Embedded AI across SAP applications

SAP's prepaid currency for premium AI capabilities.

From The SAP Business AI landscape: where every piece sits

reviewers stop checking carefully because most flags are noise.

From Anomaly detection in finance and procurement

sharing a cost center's costs out to others by a key, such as floor space or energy use.

From Record to report, end to end

the models an administrator permits; anything else is refused.

From Choosing and calling LLMs

a list of what is allowed; everything not on it is left out.

From Set up for Unit 14: capstone environment and a public portfolio

a CDS view that defines a report's measures, filters and layout.

From Reporting and analytics architecture: what lives where in an SAP landscape

using an index to check only part of the data; faster, may miss matches.

From Vector databases explained: indexes, trade-offs and how to choose

a label a server puts on a tool, such as "read-only". A hint, not a guarantee.

From The Model Context Protocol (MCP)

a record that differs from most others. Not necessarily an error.

From Anomaly detection in finance and procurement

removing the link to a person for good.

From Data security and PII in AI systems

masking that can't be reversed / masking that can, so real values return in the answer.

From Guardrails and content filtering

a hand-written, reviewed version of an artifact used to score AI drafts of the same kind.

From AI-accelerated delivery on SAP programs

a door one program offers to others to request data or actions.

From Python and APIs for AI engineers

a secret string that identifies the caller.

From Python and APIs for AI engineers

accepting the AI when it is right and overriding it when it is wrong.

From Adoption and change management for AI

a point where the agent stops and a person with the right role decides before data changes.

From Multi-step agents: planning, state and recovery

code that checks every proposed action and queues it for a person, or refuses it.

From One order-exception agent, built three ways

the highest amount a person may approve.

From Capstone: an enterprise exception management platform

the table of which roles must approve, by tier.

From AI governance for SAP programs

a signed record that a named person approved one exact action, until a set time.

From Designing human approval into AI workflows

a short record of one decision: context, options, decision, consequences.

From The integration platform and clean core: a target architecture for a hybrid landscape

a listed use case is not high-risk if the system only performs a narrow or preparatory task and doesn't materially influence the decision. Never available for profiling.

From The EU AI Act: what SAP customers need to do and when

the one decision or action you want from this audience.

From Technical storytelling: one architecture, five audiences

a configured link between two org units, such as a plant to a company code.

From How an SAP enterprise is wired: organizational structure

the person who must make an assumption true, by a date.

From Scoping and estimating an AI engagement

the sender hands over a message and continues; delivery happens later.

From How SAP systems talk: integration styles and when to use each

every event is delivered, but some may be delivered more than once.

From Event-driven agents: triggering AI from SAP business events

the quantity that can still be promised to new orders.

From Order to cash end to end

the stored set of attack prompts, grouped by what they try to achieve.

From Red-teaming your AI system

the share of attack attempts that got what the attacker wanted.

From Red-teaming your AI system

a bot that runs on demand or with a person present during the run.

From Agents that work SAP screens: computer use, RPA and when APIs are missing

a step where each word gives weights to other words and blends in their information.

From Attention, explained

one learned "question" a model asks; models run many heads at once.

From Attention, explained

how much one word draws on another; one word's weights add up to 100%.

From Attention, explained

a tamper-resistant record of who did what, when.

From SAP tools for agents

a random share of decisions checked after the fact by a reviewer.

From Adoption and change management for AI

the logged record of who changed or read what, and when.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

SAP's rule for what a user may do with which data, for example display orders of sales organization 1010.

From Grounding on SAP data without breaking authorizations: live queries, access labels and the late check

the OAuth flow where a person logs on in the browser and the app receives a token for that person.

From Secrets, API authentication and OAuth

rules that choose G/L accounts from organizational units, master data and the type of transaction.

From Document flow and how logistics becomes finance

following a system's suggestion so readily that its errors go unnoticed.

From Evaluating in production: shadow mode, A/B tests and user feedback

the share of a month a system is up, as defined and measured by the contract.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

the check that decides what quantity a sales order can be promised, and when.

From One exception, traced across the enterprise

a separate data center within a region / a geographic area with several zones.

From Running AI on SAP BTP in production

work the API accepts now and finishes later; the caller checks its status.

From REST API design for AI services

the list of assessed AI ideas, each with value, effort, run cost, risk tier and dependencies.

From An AI roadmap and operating model for an SAP customer

re-running the availability check for many orders after supply or demand changes.

From One exception, traced across the enterprise

the method that works out how to nudge every weight in every layer.

From Neural networks from scratch

replaying the past, forecasting from older data and comparing with what happened next.

From Time-series forecasting for SAP planning

a business function in an SAP system, and the remote call used to run it.

From How SAP systems talk: integration styles and when to use each

AI capabilities included in standard SAP cloud subscriptions.

From Embedded AI across SAP applications

SAP's split between AI included in cloud subscriptions and AI paid for with AI Units.

From Buy vs. build for SAP AI

an AI feature included with the SAP product license.

From Where AI creates value in S/4HANA processes

the model you start from.

From When to fine-tune (and when not to)

the measured state of a process before a change, so improvement can be proven.

From Where AI creates value in S/4HANA processes

the rows used to compute one step. All rows, one row, or a small group (mini-batch).

From Gradient descent, by hand

the time when heavy background jobs can run without hurting users.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

a token that works for whoever holds it, like a cinema ticket.

From Secrets, API authentication and OAuth

comparing actual results with the plan after go-live.

From The AI business case

a normal question run alongside attacks, to check the defences don't block real work.

From Red-teaming your AI system

whether errors lean one way, consistently over- or under-forecasting.

From Time-series forecasting for SAP planning

the list of components and quantities that make a product.

From Master data that every module depends on

a reason set on an order that stops invoicing or shipping until someone removes it.

From Order to cash end to end

a WBS element that receives the sales order item and its revenue.

From Quality, projects and service: the processes between the big four

the average storage per weight, including the scales. Q8_0 averages 8.5, not 8.

From Quantization and local models

one unit of the stack, made of an attention step and a feed-forward step.

From The transformer architecture

restricting access to data whose purpose has ended but which must be kept for legal retention.

From Data security and PII in AI systems

the standard keyword ranking formula; it rewards rare words and repeated matches, with limits.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

a named line of work; main is the agreed, working version.

From Git and GitHub for AI engineers: branches, pull requests and review

SAP's pattern for connecting a pro-code agent to Joule.

From Agent frameworks compared

an AI agent that works inside web pages, often by reading the page's structure as well as screenshots.

From Agents that work SAP screens: computer use, RPA and when APIs are missing

a hard limit on steps, calls, tokens or time for one run.

From Multi-step agents: planning, state and recovery

the platform tool that recognizes your language and installs what the app needs.

From Deploying AI apps on SAP BTP

a document that compares options on cost, benefit, risk and timing to support a decision.

From The AI business case

a group of apps and the authorizations behind them, used to build business roles.

From Roles and authorization design across modules

a message from a business system saying something happened to a business object, such as "sales order changed".

From Event-driven agents: triggering AI from SAP business events

a WHERE condition on a business column, such as company code, that narrows a search.

From SAP HANA Cloud vector engine: embeddings next to business data, searched with SQL

in RAP, data plus its behavior: operations, actions, checks and authorization.

From ABAP Cloud and RAP for AI builders

the single object for customers, suppliers and contacts in SAP S/4HANA, with roles for each use.

From Master data that every module depends on

the access for one job, such as buyer. In Public Edition, a business role is built from business catalogs; in PFCG systems, often a composite of single roles.

From Roles and authorization design across modules

the stored result was found and used / not found, so the work was done.

From Caching for LLM applications

everything used to decide whether a stored answer matches a new request.

From Caching for LLM applications

whether a score of 0.7 really means about 70% of such cases are positive.

From Classification and the metrics that matter

a small share of users or traffic that gets a new version first, compared with a control group.

From Shipping AI changes safely: CI/CD for prompts, models and agents

the passages the first, fast search passes on.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

SAP Cloud Application Programming Model: a framework for business services on BTP

From CAP and side-by-side extensions for AI

SAP BTP's billing unit; generative AI hub tokens are converted into them.

From Token economics

the end-of-unit project that combines what the course taught into one working solution.

From Set up for Unit 14: capstone environment and a public portfolio

a router that asks a cheap model first and escalates to a stronger one when needed.

From Model selection and routing

one instance of the process, such as one sales order.

From Process mining to find AI opportunities

problem, users, approach, human approval, results with sources, decision and lessons for one project.

From Your FDE portfolio and career path

a list of standard codes, grouped by type (object parts, damage, causes, activities, tasks).

From Maintain to settle: plant maintenance and assets

Core Data Services: the modeling language CAP uses to describe data and services

From CAP and side-by-side extensions for AI

a named, business-friendly view over tables, defined with SAP's Core Data Services.

From The SAP data model for AI engineers: tables, CDS views and OData

one block of code in a notebook, run on its own.

From Set up for Unit 2: data science tools

a central team that sets standards, runs intake and builds shared assets.

From An AI roadmap and operating model for an SAP customer

a respected user in a team who helps colleagues and passes feedback on.

From Adoption and change management for AI

recording and approving every change before it reaches production.

From AI governance for SAP programs

saving that state after a step, so the run can continue from there.

From Multi-step agents: planning, state and recovery

a store a framework writes state to after every step, so a run can resume.

From Agent frameworks compared

an open-source vector store that runs on a laptop inside Python.

From Set up for Unit 7: vector search: HANA Cloud and a local store

a passage cut from a document, small enough to retrieve and paste into a prompt.

From RAG fundamentals: the architecture end to end

the maximum length of a chunk, in characters or tokens.

From Chunking and document preparation: from PDF to retrievable passages

a service that runs the tests automatically on every change, such as GitHub Actions or SAP Continuous Integration and Delivery.

From Testing AI applications: unit and integration tests

continuous integration and delivery; every change is built and tested automatically, and passing changes are released through a pipeline.

From Shipping AI changes safely: CI/CD for prompts, models and agents

a switch that stops calling a failing model for a while, so users don't wait on it.

From Model selection and routing

a pointer from a claim in the answer to the passage that supports it.

From RAG fundamentals: the architecture end to end

share of citations that are needed to support their claim.

From Measuring hallucinations and groundedness

share of claims whose cited passages actually support them.

From Measuring hallucinations and groundedness

one checkable statement in an answer; long answers are split into many.

From Measuring hallucinations and groundedness

one outcome is much rarer than the other.

From Classification and the metrics that matter

predicting one category from a fixed set, such as late or on time.

From Classification and the metrics that matter

keeping SAP's standard code unmodified and building extensions outside or through released interfaces.

From What an SAP forward deployed engineer does

SAP's A-to-D rating of how safely an extension survives upgrades.

From Fit-to-standard and extension choices: where each requirement should live

SAP's scale from A (released interfaces only) to D (modifications or writing to SAP tables).

From AI in S/4HANA migration projects

fixing, merging or excluding bad records, ideally in the source system.

From Data migration: strategy, objects and reconciliation

a self-contained unit inside one SAP system with its own business data and settings, identified by three digits.

From Landscape and deployment: editions, systems and the migration approach

the OAuth flow where a system swaps its own client ID and secret for a token. No person involved.

From Secrets, API authentication and OAuth

the plan of tasks by working day (WD1, WD2) after the month ends.

From Record to report, end to end

software inside the company network that links on-premise systems to BTP.

From SAP BTP foundations for AI builders

an open-source platform for running apps, offered as an environment in SAP BTP. You push code; it runs it.

From Deploying AI apps on SAP BTP

an open standard for the envelope of an event (its ID, source, type and time).

From Event-driven agents: triggering AI from SAP business events

an AI assistant that plans, edits code, runs tools and checks its own work in a loop.

From Coding agents for SAP developers

a named set of vectors that share one embedding model and settings.

From Vector databases explained: indexes, trade-offs and how to choose

one saved snapshot of the files, with a message saying why.

From Git and GitHub for AI engineers: branches, pull requests and review

the S/4HANA Cloud configuration that opens an API to a specific external system.

From Python and APIs for AI engineers

SAP's way of opening a specific API to a specific external system with a specific technical identity.

From Calling your first SAP API

SAP's objects that allow an outside system to call an inbound service, with a defined user and URL.

From An agent that calls a RAP service

a technical user that an outside system signs in with.

From An agent that calls a RAP service

replacing a long conversation with a short summary built from the saved state.

From Multi-step agents: planning, state and recovery

a legal entity with its own set of books.

From How an SAP enterprise is wired: organizational structure

checking every request in the system that owns the data, not trusting the caller.

From Agent permissions and SAP authorizations

an AI model that looks at screenshots and decides which mouse and keyboard actions to take next.

From Agents that work SAP screens: computer use, RPA and when APIs are missing

a stored price, discount or surcharge with its key and validity.

From Master data that every module depends on

SAP's rule mechanism for finding prices, discounts and surcharges.

From Order to cash end to end

a range that likely holds the true rate, given how many cases you have.

From Evaluating in production: shadow mode, A/B tests and user feedback

the system's own estimate, per field, that a value is right.

From Document AI, built and bought: from supplier invoice PDF to checked SAP data

the confidence below which a cascade escalates.

From Model selection and routing

settings that change how standard SAP behaves, without code.

From Fit-to-standard and extension choices: where each requirement should live

the shop floor's report of yield, scrap and time for an operation.

From Plan to produce end to end

the setting that says which confirmations a supplier must send.

From Procure to pay end to end

comparing what happened with what should have happened.

From Process mining to find AI opportunities

the table of caught, false alarms, missed and correctly left alone.

From Classification and the metrics that matter

one tool that does a whole job, even if it makes several API calls inside.

From Tool design for agents

combining several companies' figures into group figures, without what they sold to each other.

From Record to report, end to end

a limit the solution must respect, such as read-only access or a person approving every change.

From Problem framing: from 'we need AI' to a measurable problem

Your own, smaller view of a remote API's entity, with the fields renamed to your terms

From CAP and side-by-side extensions for AI

a sealed, portable package of a program and everything it needs.

From Set up for Unit 10: observability and containers

the software that runs containers, such as Docker or Podman.

From Set up for Unit 10: observability and containers

a classifier that scores text for harm categories such as hate or violence.

From Guardrails and content filtering

checking input and output against harm categories, with a strictness threshold for each.

From SAP Generative AI Hub and the orchestration service

everything the model sees in one call: instructions, examples, data, earlier messages and the question.

From Prompt and context engineering

example rows where the answer is known, sent with the request.

From Tabular AI and SAP's predictive models

the most text a model can take in one request.

From Attention, explained

a structured kind of shadowing in which the user is the expert and the researcher asks questions while the work happens.

From Discovery: interviewing the customer

money held for known risks, sized from the range rather than a flat percentage.

From Scoping and estimating an AI engagement

the agreed request fields, response fields and error codes of an API.

From Building an AI API

a test that checks a promise between two systems still holds: that callers still get the fields they rely on, or that a stand-in still behaves like the real system.

From Testing AI applications: unit and integration tests

the source says otherwise, versus the source says nothing about it.

From Measuring hallucinations and groundedness

how much one field pushed one record's score up or down from the average.

From Explaining model predictions to the business

cases handled without the AI in the same period, used for a fair comparison.

From Defining success metrics with the business

a count or amount from the source that a report must match.

From Reporting and analytics architecture: what lives where in an SAP landscape

the point where more steps no longer reduce the loss.

From Gradient descent, by hand

the usual score, based on the angle between two vectors. 1 means same direction; 0 means unrelated.

From Embeddings and semantic similarity

the cost of one complete unit of business work, across all its model calls.

From Token economics

the share of eligible work items where the tool was used.

From Adoption and change management for AI

the computer's normal processor.

From Set up for Unit 4: PyTorch and an optional free GPU

what a customer already owes or has ordered, measured against their credit limit.

From Order to cash end to end

a re-ranker that reads the question and one passage together and outputs a relevance score.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

a short-lived value SAP services expect before they accept a change request.

From Calling your first SAP API

programs and objects a company wrote itself, usually with names starting with Z or Y.

From AI in S/4HANA migration projects

the mechanism that keeps business partners and the customer and supplier records in step.

From Master data that every module depends on

the planned window when the old system stops, the final load runs and the new system goes live.

From Data migration: strategy, objects and reconciliation

the hour-by-hour task list for the switch from the old system to the new one.

From AI-accelerated delivery on SAP programs

a public ID for one vulnerability, such as CVE-2026-59822.

From MCP security in 2026: lessons from the first exploited vulnerabilities

a plain-language page saying what each field means, its type, and the rules for reading it.

From The SAP data model for AI engineers: tables, CDS views and OData

replacing personal data with placeholders before the model sees it. Anonymization can't be reversed; pseudonymization can, so names can be put back into the answer.

From SAP Generative AI Hub and the orchestration service

a GDPR principle that personal data must be limited to what the purpose needs.

From Is your SAP data ready for AI? A readiness check you can run

sending or storing only the personal data a purpose needs.

From Hire-to-retire: where people data meets logistics and finance

a packaged, described dataset, such as S/4HANA sales orders, ready for analytics and AI.

From Reporting and analytics architecture: what lives where in an SAP landscape

a check one field or record must pass, such as "a weight has a weight unit".

From Is your SAP data ready for AI? A readiness check you can run

the person who reviews, approves and corrects master data for an area.

From Master data that every module depends on

the person the data is about.

From Data security and PII in AI systems

a pandas table with named columns, the main thing you work with in Unit 2.

From Set up for Unit 2: data science tools

a holding place for messages that can't be processed, so they don't block the rest.

From Calling AI from SAP Integration Suite

a tool that pauses a running program at a breakpoint so you can inspect its values and step through it line by line.

From Errors, logging and debugging

the list of a program's architecture decisions, each with context, options, decision and consequences.

From The solution architect on an SAP program

the rule that turns the model's probabilities into one chosen token.

From How LLMs generate text, and what that means for enterprise use

an interview round where you break a vague real-world problem into parts, usually with little or no code.

From Your FDE portfolio and career path

neural networks with many layers.

From Neural networks from scratch

an agent acts for a user, and both are known to the systems it calls.

From Agent permissions and SAP authorizations

something standard SAP doesn't do that the business needs.

From Fit-to-standard and extension choices: where each requirement should live

a package your project needs. Packages your dependencies need are transitive dependencies.

From Python projects done right: structure, virtual environments and packages

checking the libraries you use against lists of known security holes.

From Set up for Unit 11: security testing tools

the organization that uses an AI system under its authority, for example a company using an SAP AI feature.

From The EU AI Act: what SAP customers need to do and when

moving an app to the environment where it will run for its users.

From Deploying AI apps on SAP BTP

where the system runs and who operates it: public cloud, private cloud or on-premise.

From Landscape and deployment: editions, systems and the migration approach

a regular, small meeting with the right to review and record significant design decisions for the program.

From The solution architect on an SAP program

a stored connection to another system: URL, proxy type and logon method.

From SAP BTP foundations for AI builders

a step only some cases take, such as a credit block.

From Process mining to find AI opportunities

code inside S/4HANA restricted to SAP's released objects.

From Fit-to-standard and extension choices: where each requirement should live

how much technical detail a diagram shows. SAP's guideline runs from an executive overview to detailed technical views.

From Technical storytelling: one architecture, five audiences

software that shows guidance inside other applications and measures their use. WalkMe is one.

From Adoption and change management for AI

Regulation (EU) 2026/1744, in force 27 July 2026, which amended the AI Act's dates and some duties.

From The EU AI Act: what SAP customers need to do and when

one position in the vector. A 384-dimensional embedding has 384 numbers.

From Embeddings and semantic similarity

the attack comes from the user's own prompt.

From Prompt injection and tool poisoning

the time before building, spent learning how the process works today and what limits a solution.

From Discovery: interviewing the customer

the rule for "close", for example cosine similarity or straight-line (Euclidean) distance.

From Vector databases explained: indexes, trade-offs and how to choose

the loss grows instead of shrinking, usually because steps are too big.

From Gradient descent, by hand

a file and command that start several containers together.

From Set up for Unit 10: observability and containers

software that turns documents into structured data and checks it.

From Document AI, built and bought: from supplier invoice PDF to checked SAP data

the links between the documents of one transaction, each pointing to its predecessor and successors.

From Document flow and how logistics becomes finance

turning files into clean, structured text before chunking.

From Chunking and document preparation: from PDF to retrievable passages

four patterns for how much a person is involved, from most to least.

From Designing human approval into AI workflows

the patterns in the data change after training, for example a route gets slower.

From Predicting late deliveries with SAP data

the average number of days customers take to pay.

From Order to cash end to end

S/4HANA's comparison of a new invoice with existing ones by supplier, currency and configured fields.

From Anomaly detection in finance and procurement

a run that survives a restart because its state is saved outside the process.

From Agent frameworks compared

a private file on a developer's computer that holds environment variables for one project. Never committed to Git.

From Secrets, API authentication and OAuth

a limited SAP programme that gives selected customers access before general availability.

From AI in S/4HANA migration projects

An SAP program that gives selected customers access before general availability.

From The SAP Business AI landscape: where every piece sits

a limited program in which selected customers use a feature before broad release.

From What SAP TechEd 2026 changed for AI builders

a runtime of Integration Suite that runs in your own environment while flows are designed in the cloud.

From The integration platform and clean core: a target architecture for a hybrid landscape

electronic data interchange, structured business messages such as orders and invoices sent between systems.

From Intercompany and multi-company scenarios

removing intercompany revenue, cost, balances and profit so the group reports only deals with outsiders.

From Intercompany and multi-company scenarios

AI features SAP builds into its applications and maintains for you.

From Embedded AI across SAP applications

S/4HANA's built-in analytics on its own live data, through CDS views and analytical queries.

From Reporting and analytics architecture: what lives where in an SAP landscape

a list of numbers (a vector) that represents the meaning of a text, image or record.

From Embeddings and semantic similarity

stores the vectors computed for text.

From Caching for LLM applications

the neural network that turns text into embeddings.

From Embeddings and semantic similarity

temporary, logged elevated access for urgent fixes.

From Roles and authorization design across modules

S/4HANA's record of a person as an employee, used by logistics and finance processes.

From Hire-to-retire: where people data meets logistics and finance

a block stack that sees the whole text, or one that sees only earlier words.

From The transformer architecture

the point where data is no longer needed for its original business purpose.

From Data security and PII in AI systems

the address of one API resource, such as sales orders.

From Python and APIs for AI engineers

matching extracted values to SAP master data, such as finding the supplier number.

From Document AI, built and bought: from supplier invoice PDF to checked SAP data

a separate platform that combines data from many systems with history.

From Reporting and analytics architecture: what lives where in an SAP landscape

the right to use a service in a subaccount; the basis of billing in a paid account.

From Set up for Unit 3: neural network tools and an SAP BTP trial

the right to use a service plan, and how much of it.

From SAP BTP foundations for AI builders

a collection of records you can ask for, such as A_SalesOrder.

From Calling your first SAP API

a named value the operating system hands to a program, such as SAP_API_KEY. Code reads the name; the value stays outside the code.

From Secrets, API authentication and OAuth

one full pass over all training examples.

From Gradient descent, by hand

an individual physical object maintained as a unit, with its own history and costs.

From Maintain to settle: plant maintenance and assets

the named team that fixes and reprocesses a failed message.

From The integration lead's toolkit: inventory, specs, testing and cutover

handing a case to a person because no rule or agent covers it.

From One order-exception agent, built three ways

the share of requests passed to the stronger model.

From Model selection and routing

a version stamp of a record. A change must quote it, so a change to an outdated version is refused.

From An agent that calls a RAP service

measuring how good AI answers are across a set of examples (see LLM evaluation fundamentals).

From Testing AI applications: unit and integration tests

a pipeline step that blocks a release when evaluation scores fall below agreed limits.

From Shipping AI changes safely: CI/CD for prompts, models and agents

the machinery that runs a fixed set of test cases on every change, records the results and applies a gate.

From Building an evaluation harness

the saved record of every case, answer and score in a run.

From Set up for Unit 8: evaluation tools

fixed test cases with the expected result, re-run after every change.

From One order-exception agent, built three ways

a message saying something happened, delivered to subscribers through a broker.

From How SAP systems talk: integration styles and when to use each

the service that receives events from publishers and delivers them to subscribers.

From Event-driven agents: triggering AI from SAP business events

records of what happened, each with a case ID, an activity name and a timestamp.

From Process mining to find AI opportunities

comparing the question with every stored vector. Always correct, slower at scale.

From Vector databases explained: indexes, trade-offs and how to choose

an item that leaves the happy path and needs a person, such as a blocked order.

From Where AI creates value in S/4HANA processes

MRP's flag on a situation it can't solve, such as reschedule in, reschedule out or cancel.

From Plan to produce end to end

OWASP's name for an AI system with more functions, permissions or autonomy than its job needs.

From SAP tools for agents

the share of test questions whose query returns the same result as a known-correct query.

From Asking SAP data questions in plain language: text-to-SQL and its limits

getting data out of a system to somewhere the attacker can read it.

From Prompt injection and tool poisoning

the conditions a mock must meet before the next step.

From Data migration: strategy, objects and reconciliation

the estimated value multiplied by the confidence that it will be reached.

From An AI roadmap and operating model for an SAP customer

how often two experts make the same decision on the same cases. Low agreement means the task is not defined yet.

From AI opportunity assessment: AI, automation or neither?

an account of why a model produced a score, in terms of its input fields.

From Explaining model predictions to the business

feedback the user gives on purpose, such as a thumb or a comment.

From Evaluating in production: shadow mode, A/B tests and user feedback

take data out of the old system, reshape it, load it into the new one.

From Data migration: strategy, objects and reconciliation

one number that balances precision and recall.

From Classification and the metrics that matter

when information is missing, deny rather than allow.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

a second model the system uses automatically when the first one fails or isn't available.

From Choosing and calling LLMs

declining when the answer was available.

From Measuring hallucinations and groundedness

a good request that a guardrail blocks by mistake.

From Guardrails and content filtering

a Python library for building web APIs, with automatic documentation pages.

From Set up for Unit 6: app tooling: Node.js, CAP and Cloud Foundry

one input column the model uses, such as lead time or weight.

From Machine learning in plain terms

a ranking of input fields by how much the model depends on them.

From Explaining model predictions to the business

putting input columns on similar scales so steps behave well.

From Gradient descent, by hand

a small network applied to each word on its own, after attention.

From The transformer architecture

a handful of worked input and answer pairs placed in the prompt.

From Prompt and context engineering

a prompt that includes a few worked examples.

From When to fine-tune (and when not to)

the share of records where a needed field actually holds a value.

From Rapid prototyping: one day, one week, two weeks

a rule on structured fields, such as product group, applied before ranking.

From Semantic search over SAP master data

one confirmed weakness, with severity, evidence, owner and status.

From Red-teaming your AI system

further training of a pretrained model on a smaller, specific data set.

From Build a tiny GPT

a short code computed from a file's contents; any edit changes it.

From Shipping AI changes safely: CI/CD for prompts, models and agents

the list of requirements with their decision, rationale and owner.

From Fit-to-standard and extension choices: where each requirement should live

the SAP Activate step that compares standard SAP processes with the business's needs, so only real gaps are built.

From Fit-to-standard and extension choices: where each requirement should live

an SAP Activate workshop where the business sees the standard process and records the gaps (delta requirements).

From AI-accelerated delivery on SAP programs

an upstream reason for the work, such as master data, settings or policy, that can be removed rather than handled.

From AI opportunity assessment: AI, automation or neither?

one agreed price for a defined scope.

From Scoping and estimating an AI engagement

a tool the agent must never call in this context. Any attempt counts as a failure.

From Evaluating agents: tasks, trajectories and tool calls

an engineer embedded with a customer who owns an outcome end to end.

From What an SAP forward deployed engineer does

the contract that sets a carrier's rates.

From Warehouse and transportation execution

the planned shipment, with the carrier, stops, dates and charges.

From Warehouse and transportation execution

the document that sends the freight charge to purchasing and finance.

From Warehouse and transportation execution

the smallest unit that transportation management plans.

From Warehouse and transportation execution

the model asks your application to run a named action with given inputs.

From Structured outputs and function calling

a place in a plant's technical structure, arranged as a hierarchy, where equipment is installed.

From Maintain to settle: plant maintenance and assets

what an interface does in business terms: trigger, content, rules, errors.

From The integration lead's toolkit: inventory, specs, testing and cutover

an assessment some deployers must do before first use, including deployers of consumer credit and life or health insurance scoring.

From The EU AI Act: what SAP customers need to do and when

the rule that merges two ranked lists into one.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

the rules that decide whether a change may go live.

From Building an evaluation harness

a product SAP sells and supports for production use.

From What SAP TechEd 2026 changed for AI builders

a model that produces text or other content, often from unstructured input.

From AI opportunity assessment: AI, automation or neither?

The part of SAP AI Core that provides access to many language models and orchestration tools.

From The SAP Business AI landscape: where every piece sits

a single model file format used by llama.cpp and Ollama, holding the weights and metadata.

From Quantization and local models

GitHub's free hosting for static web pages from a repository.

From Your FDE portfolio and career path

the contract with SAP; holds subaccounts, entitlements and quotas.

From SAP BTP foundations for AI builders

what the model relies on across many records.

From Explaining model predictions to the business

written definitions of business terms ("open order") in terms of fields and values.

From Asking SAP data questions in plain language: text-to-SQL and its limits

the three gate outcomes: continue, repeat once with a named fix, or end.

From Rapid prototyping: one day, one week, two weeks

cases with answers experts agreed are right; the fixed exam.

From LLM evaluation fundamentals

a free, browser-based notebook service from Google, with optional GPUs when available.

From Set up for Unit 4: PyTorch and an optional free GPU

a decoder-only transformer trained to predict the next token.

From Build a tiny GPT

a chip that runs many calculations at once; it speeds up neural network training.

From Set up for Unit 4: PyTorch and an optional free GPU

a temporary account between goods receipt and supplier invoice.

From Document flow and how logistics becomes finance

the direction and steepness of the loss for each weight; it says which way to nudge.

From Gradient descent, by hand

always pick the most likely token.

From How LLMs generate text, and what that means for enterprise use

an answer based on retrieved sources rather than the model's general knowledge.

From Observability for AI systems

whether each claim in an answer is supported by the retrieved passages.

From Measuring hallucinations and groundedness

anchoring an answer in trusted data. RAG is one way to ground.

From RAG fundamentals: the architecture end to end

the currency the group reports in; Calder's is USD.

From Intercompany and multi-company scenarios

a check before or after the model that can stop, change or log a request.

From Guardrails and content filtering

a number that must not get worse, with an agreed limit.

From Defining success metrics with the business

fluent output that is not true, such as an invented quantity.

From Build a tiny GPT

one agent passing a task to another.

From One order-exception agent, built three ways

a package with its contents, such as a pallet with two pumps, with its own number.

From Warehouse and transportation execution

the normal flow of a process when nothing goes wrong.

From Where AI creates value in S/4HANA processes

one request and response shape across many model providers.

From SAP Generative AI Hub and the orchestration service

SAP's Python driver for connecting to SAP HANA.

From Set up for Unit 7: vector search: HANA Cloud and a local store

one record per document and one per line of it.

From The SAP data model for AI engineers: tables, CDS views and OData

the first sentence of an update, stating the answer and any bad news.

From Technical storytelling: one architecture, five audiences

a regular test the platform runs to see if the app still answers. A failed check leads to a restart.

From Deploying AI apps on SAP BTP

a layer between the inputs and the output that builds its own intermediate signals.

From Neural networks from scratch

a system in a listed use case (Annex III) or a safety component of a regulated product (Annex I).

From The EU AI Act: what SAP customers need to do and when

the end-to-end process for employees, from recruiting to exit.

From Hire-to-retire: where people data meets logistics and finance

the share of requests answered from the cache.

From Caching for LLM applications

the share of test searches with a right answer in the top three results.

From Semantic search over SAP master data

the share of test questions whose answer is in one of the top k retrieved chunks.

From Chunking and document preparation: from PDF to retrievable passages

share of questions with at least one relevant document in the top k.

From Evaluating retrieval

a common ANN index that links vectors into layered graphs of neighbours.

From Vector databases explained: indexes, trade-offs and how to choose

a graph-based index that makes similarity search fast on large collections, with a small chance of missing a match.

From SAP HANA Cloud vector engine: embeddings next to business data, searched with SQL

how far ahead a forecast must be right to be useful.

From Time-series forecasting for SAP planning

the AI application. It holds the conversation, connects to servers and enforces policy.

From The Model Context Protocol (MCP)

the protocol web browsers and APIs use to send requests and answers.

From Python and APIs for AI engineers

a tool that sends a request to an API and shows the status code, headers and body of the answer.

From Set up for Unit 5B: integration tools

the verb of a request: GET reads, POST creates, PUT replaces, PATCH changes, DELETE removes.

From REST API design for AI services

a central team for platform and rules, with builders placed in the business areas.

From An AI roadmap and operating model for an SAP customer

a design where a person approves or corrects the AI's action before it takes effect.

From What an SAP forward deployed engineer does

the people and steps that check, and can overrule, what the AI produces.

From Writing a solution design document for an AI use case

an expert grading answers; the reference for every other method.

From LLM evaluation fundamentals

combines keyword and semantic results into one list.

From Semantic search over SAP master data

the period of close support right after go-live.

From The integration lead's toolkit: inventory, specs, testing and cutover

S/4HANA Cloud authorization objects that give users access to an app or feature.

From Embedded AI across SAP applications

processing the same message twice has the same effect as once.

From Calling AI from SAP Integration Suite

a unique value the caller sends so a retried request isn't processed twice.

From REST API design for AI services

a call that has the same effect whether it runs once or several times. Reading data is idempotent; creating a sales order usually isn't.

From Errors, logging and debugging

SAP's intermediate document format for exchanging business documents.

From How SAP systems talk: integration styles and when to use each

the template a container starts from; a container is a running image.

From Set up for Unit 10: observability and containers

feedback read from behavior, such as accepting, editing or overriding a suggestion.

From Evaluating in production: shadow mode, A/B tests and user feedback

learning only from the examples included in the request, with no change to the model.

From Tabular AI and SAP's predictive models

the process for a product the customer sends in to be repaired.

From Quality, projects and service: the processes between the big four

an event where an AI use case causes, or could cause, harm or a wrong business result.

From AI governance for SAP programs

the list of fields missing for a later step.

From Order to cash end to end

the attack hides in content the model reads, such as a document, email or record.

From Prompt injection and tool poisoning

SAP's term for use of its system through other software, devices or bots instead of a person logging on directly.

From Agents that work SAP screens: computer use, RPA and when APIs are missing

using a trained model to predict for a new case.

From Machine learning in plain terms

the prompt and the answer, counted separately for billing.

From How LLMs generate text, and what that means for enterprise use

what you send: instructions, data, retrieved passages, the question.

From Token economics

an open-source evaluation framework from the UK AI Security Institute.

From Set up for Unit 8: evaluation tools

the quality management record of an inspection, created for example at goods receipt.

From Procure to pay end to end

one running database you created.

From Set up for Unit 7: vector search: HANA Cloud and a local store

a file in the project, such as AGENTS.md, that tells coding agents the team's rules.

From Coding agents for SAP developers

a sequence of steps in SAP Cloud Integration that receives, transforms and sends messages.

From Calling AI from SAP Integration Suite

a cloud service that translates, routes and monitors messages between systems.

From The integration platform and clean core: a target architecture for a hybrid landscape

SAP's integration platform on SAP BTP.

From Set up for Unit 5B: integration tools

a test that uses a real external system, such as the SAP sandbox.

From Testing AI applications: unit and integration tests

a business flow between two company codes of the same group, such as a sale from one to the other.

From How an SAP enterprise is wired: organizational structure

one defined exchange between two systems.

From How SAP systems talk: integration styles and when to use each

the single list of all interfaces with owner, pattern, volume, service level, error handling, monitoring and status.

From The integration lead's toolkit: inventory, specs, testing and cutover

demand with many zero periods, typical of spare parts.

From Time-series forecasting for SAP planning

the invoice the delivering company sends the selling company in an intercompany sale (billing type IV).

From Intercompany and multi-company scenarios

the agent may do only what both the user and the agent are allowed to do.

From Agent permissions and SAP authorizations

removing stored answers that are no longer true.

From Caching for LLM applications

SAP's Integration Solution Advisory Methodology, a way to classify interfaces and choose technology.

From The integration platform and clean core: a target architecture for a hybrid landscape

Intelligent Scenario Lifecycle Management, the S/4HANA tool for managing machine learning scenarios.

From Tabular AI and SAP's predictive models

SAP's way to manage AI scenarios in ABAP systems and call generative AI hub models from ABAP code.

From ABAP Cloud and RAP for AI builders

a model that isolates records with random splits; records that are easy to isolate are unusual.

From Anomaly detection in finance and procurement

SAP's AI assistant in SAP applications.

From The SAP Business AI landscape: where every piece sits

An AI agent that performs a task in a business function.

From The SAP Business AI landscape: where every piece sits

A role-based assistant that coordinates several agents for a person.

From The SAP Business AI landscape: where every piece sits

included capabilities versus advanced capabilities and agent actions paid in AI Units.

From Joule and Joule Studio

a fixed, rule-based job Joule can run, such as reading one record.

From Joule and Joule Studio

SAP's environment for building custom agents, apps and workflows.

From The SAP Business AI landscape: where every piece sits

SAP's managed service that runs what you build in the new Joule Studio.

From Joule and Joule Studio

the finance record of a posting, with debit and credit lines that add up to zero.

From Document flow and how logistics becomes finance

a plain-text format for data, built from lists and name-value pairs.

From Python and APIs for AI engineers

a setting that guarantees valid JSON, but not any particular fields.

From Structured outputs and function calling

the rule set, small model or LLM that decides whether a claim is supported.

From Measuring hallucinations and groundedness

the natural-language query feature of SAP Analytics Cloud, also used by Joule for analytical questions.

From Asking SAP data questions in plain language: text-to-SQL and its limits

the running Python behind a notebook. It remembers what earlier cells did.

From Set up for Unit 2: data science tools

CISA's list of vulnerabilities known to be exploited in real attacks.

From MCP security in 2026: lessons from the first exploited vulnerabilities

SAP's term for a recorded design decision, used when a deviation from standard is reviewed.

From The solution architect on an SAP program

the table that links each old number to its new number in the target.

From Data migration: strategy, objects and reconciliation

SAP apps that let trained business users add fields, rules, forms and reports.

From Fit-to-standard and extension choices: where each requirement should live

ranking by the words a document shares with the question.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

finds records that share words with the query. BM25 is a common way to rank them.

From Semantic search over SAP master data

a store of facts as linked things and relationships, such as "order 5004 is sold to customer 10100001".

From SAP Knowledge Graph and Business Data Cloud: giving AI a map of the business and data it can trust

an unusual but explained pattern, approved by a reviewer, that the detector stops flagging.

From Anomaly detection in finance and procurement

key performance indicator, a metric the business already tracks and reports.

From Defining success metrics with the business

SAP's managed Kubernetes runtime on BTP, for container-based apps.

From Deploying AI apps on SAP BTP

the known answer in past examples, such as "late".

From Machine learning in plain terms

past items with the correct label, used to train or test.

From Classifying text: LLMs, embeddings or a trained model?

test questions with the documents an expert marked as relevant, sometimes with grades such as 2 = answers it, 1 = helps.

From Evaluating retrieval

the set of systems behind one ERP, such as development, quality and production.

From Landscape and deployment: editions, systems and the migration approach

how long an answer takes to come back.

From Choosing and calling LLMs

the time a use case may take, as a percentile, split across its steps.

From Latency budgets and performance

rescaling a word's numbers to a standard range before each step.

From The transformer architecture

leading metrics move quickly (notes opened); lagging ones move later (overdue receivables).

From Defining success metrics with the business

a question that suggests its own answer, such as "Wouldn't AI make this easier?"

From Discovery: interviewing the customer

an automated test in which users with different roles ask the same questions and nobody receives a source they may not see.

From Grounding on SAP data without breaking authorizations: live queries, access labels and the late check

training with information that would not exist when the prediction is made.

From Predicting late deliveries with SAP data

how big each step is.

From Gradient descent, by hand

giving each identity only the permissions its job needs.

From Agent permissions and SAP authorizations

private data, untrusted content and a way to send data out, all in one agent.

From Prompt injection and tool poisoning

the usual amount, its slow drift, its repeating shape and the random rest.

From Time-series forecasting for SAP planning

ready-made code you install with pip.

From Set up for Unit 2: data science tools

the terms under which others may use your code.

From Set up for Unit 14: capstone environment and a public portfolio

a model that reads and writes text, called over the internet and paid per use.

From Choosing and calling LLMs

a general language model asked to put numbered passages in order.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

a model that grades answers against a written rubric.

From LLM evaluation fundamentals

the order objects load in, so every record finds what it points to.

From Data migration: strategy, objects and reconciliation

why one record got its score.

From Explaining model predictions to the business

a language model that runs on a computer you control, not in a provider's cloud.

From Set up for Unit 12: local models

a file that pins every package, transitive ones included, so installs are repeatable.

From Python projects done right: structure, virtual environments and packages

how serious a log entry is: DEBUG, INFO, WARNING, ERROR or CRITICAL.

From Errors, logging and debugging

a cheap way to fine-tune that trains small add-on matrices instead of the whole model.

From When to fine-tune (and when not to)

a single number measuring how wrong the model is on the training examples. Lower is better.

From Gradient descent, by hand

a Node.js release line that gets fixes for a long, published period.

From Set up for Unit 6: app tooling: Node.js, CAP and Cloud Foundry

a score that weighs every label equally, so rare labels count.

From Classifying text: LLMs, embeddings or a trained model?

mean absolute error, in units such as pieces.

From Time-series forecasting for SAP planning

SAP's standard support period, and a paid extension after it.

From AI for SAP ECC customers before they migrate

the standard support period SAP includes in maintenance fees, with fixes and legal changes.

From Landscape and deployment: editions, systems and the migration approach

the plan and cost collector for a piece of maintenance work.

From Maintain to settle: plant maintenance and assets

the rule that creates orders or notifications on a cycle, a counter or a condition.

From Maintain to settle: plant maintenance and assets

a small file that tells Cloud Foundry how to run the app: name, memory, start command, health check.

From Deploying AI apps on SAP BTP

mean absolute percentage error; undefined when actual demand is zero.

From Time-series forecasting for SAP planning

a field-by-field table from source to target, with the rule for each field.

From The integration lead's toolkit: inventory, specs, testing and cutover

S/4HANA's profitability reporting by product, customer and other characteristics.

From Record to report, end to end

mean absolute scaled error; below 1 means better than the naive benchmark it is scaled by.

From Time-series forecasting for SAP planning

long-lived records that many documents refer to, such as customers, suppliers and materials.

From Master data that every module depends on

long-lived records (customers, materials) vs. business documents (orders, invoices, journal entries).

From The SAP data model for AI engineers: tables, CDS views and OData

the logistics record of a goods movement, such as a goods issue or goods receipt.

From Document flow and how logistics becomes finance

the record for a product or material, with views for each department and org level.

From Master data that every module depends on

the most output tokens a call may produce; the answer stops there.

From How LLMs generate text, and what that means for enterprise use

an open standard for connecting AI applications to tools and data.

From The Model Context Protocol (MCP)

the program that connects to a server and calls its tools; in production, the AI application.

From Set up for Unit 9: MCP and agent tooling

SAP's managed service in Integration Suite for exposing APIs as MCP tools.

From The Model Context Protocol (MCP)

a free tool for testing an MCP server by hand.

From Set up for Unit 9: MCP and agent tooling

a program that offers tools to an AI agent through the Model Context Protocol.

From Coding agents for SAP developers

where readings are recorded; a counter adds up, such as operating hours.

From Maintain to settle: plant maintenance and assets

the model reproduces training examples word for word.

From Build a tiny GPT

notes kept outside one run and shown to the agent in later runs.

From Multi-step agents: planning, state and recovery

two changes to the same lines that Git can't combine on its own; a person decides.

From Git and GitHub for AI engineers: branches, pull requests and review

a text format for diagrams, written in a code block and drawn by tools that support it.

From Set up for Unit 1B: process tracing tools

the machine-readable contract of an OData API: records, fields, types and links.

From Calling your first SAP API

a condition on labels such as company code or process, applied during the search.

From Vector databases explained: indexes, trade-offs and how to choose

a number aggregated over many requests, such as error rate or tokens per minute.

From Observability for AI systems

one kind of data the tool loads, such as customers, open purchase orders or receivable open items, with its own template and rules.

From Data migration: strategy, objects and reconciliation

invoicing a percentage of the price when a project milestone is reached.

From Quality, projects and service: the processes between the big four

a tool that copies one API call one to one, with the API's names and results.

From Tool design for agents

a documented check that reduces a risk you can't remove.

From Roles and authorization design across modules

a small program that imitates a real system so you can test against it safely.

From Set up for Unit 5B: integration tools

a full rehearsal of the load into a test system.

From Data migration: strategy, objects and reconciliation

a stand-in for a real tool that records calls and returns test data, so nothing real changes.

From Evaluating agents: tasks, trajectories and tool calls

the learned logic that turns inputs into a prediction.

From Machine learning in plain terms

the screen in SAP AI Launchpad for exploring models, benchmarks and costs.

From Choosing and calling LLMs

the date after which a model version is no longer served.

From Shipping AI changes safely: CI/CD for prompts, models and agents

choosing which model answers a call, so apps ask for "fast" or "strong" rather than a provider's model name.

From AI gateways: governing model and agent traffic

a program that loads a model file and answers prompts. Ollama is one.

From Set up for Unit 12: local models

choosing the model for a task from evaluation results against a quality, speed and cost target.

From Model selection and routing

a dated release of a model; versions are deprecated and then retired.

From Choosing and calling LLMs

the date after which a model version is no longer served.

From Running AI on SAP BTP in production

one step in that pipeline, such as templating, data masking or output filtering.

From SAP Generative AI Hub and the orchestration service

the value of a measure, such as how often users override answers, that triggers action.

From AI governance for SAP programs

the run that compares requirements with stock and receipts and proposes what to make or buy.

From Plan to produce end to end

average of 1 divided by the position of the first relevant result.

From Evaluating retrieval

SAP's way to package an app with several parts, such as a service, a database and sign-in, and deploy them together.

From Deploying AI apps on SAP BTP

a question whose answer needs several linked facts, such as order to customer to invoices.

From SAP Knowledge Graph and Business Data Cloud: giving AI a map of the business and data it can trust

the classic network of fully connected layers, the kind built in this topic.

From Neural networks from scratch

searching with several versions of the question and merging the results.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

a case that must be correct every time, whatever the averages say.

From Building an evaluation harness

next period equals the last period. Seasonal naive: next period equals the same period one cycle ago.

From Time-series forecasting for SAP planning

starting tool names with the object or system, such as sales_order_ or customer_, so similar tools are easy to tell apart.

From Tool design for agents

asking a data question in everyday words and getting an answer from a database.

From Asking SAP data questions in plain language: text-to-SQL and its limits

a named link from one record to related records, such as to_Item from an order to its items.

From Calling your first SAP API

a score from 0 to 1 that rewards putting the most relevant documents first.

From Evaluating retrieval

the stored vectors closest to a question's vector.

From Vector databases explained: indexes, trade-offs and how to choose

future benefits minus future costs, each reduced for how far in the future it falls. Positive means the option beats doing nothing.

From The AI business case

a model made of layers of simple units, trained by gradient descent.

From Neural networks from scratch

the three ways to reach S/4HANA from an older ERP.

From Landscape and deployment: editions, systems and the migration approach

an instance setting that SAP's documentation names as needed for in-database embeddings and re-ranking.

From SAP HANA Cloud vector engine: embeddings next to business data, searched with SQL

a program that runs JavaScript outside a browser; CAP's Node.js flavour runs on it.

From Set up for Unit 6: app tooling: Node.js, CAP and Cloud Foundry

a measurable "how much, how fast, how reliably" target, with an owner.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

something the team states it will not do in this release.

From Writing a solution design document for an AI use case

a page of code cells, each followed by its output.

From Set up for Unit 2: data science tools

the report of a malfunction or a request for work, with codes for object part, damage and cause.

From Maintain to settle: plant maintenance and assets

the package installer that comes with Node.js.

From Set up for Unit 6: app tooling: Node.js, CAP and Cloud Foundry

a standard for handing out access tokens so an app can call an API without sending a password each time.

From Secrets, API authentication and OAuth

analysis where one event can belong to several objects, such as an order and a delivery.

From Process mining to find AI opportunities

being able to answer new questions about a running system from the data it records.

From Observability for AI systems

the result of a tool, which the model reads before its next decision.

From Agents from first principles

turning an image of text, such as a scan, into text.

From Chunking and document preparation: from PDF to retrievable passages

the query protocol behind many SAP APIs.

From Python and APIs for AI engineers

SAP's standard web API format for business objects, described by its metadata.

From The SAP data model for AI engineers: tables, CDS views and OData

testing on a fixed set before release.

From LLM evaluation fundamentals

measuring quality on real use after release, through ratings, reviews and monitoring.

From LLM evaluation fundamentals

the agreed list of business concepts and allowed relationships in a graph.

From SAP Knowledge Graph and Business Data Cloud: giving AI a map of the business and data it can trust

a beta tool in SAP Business Data Cloud for defining business views on SAP Knowledge Graph.

From What SAP TechEd 2026 changed for AI builders

you may use documentation, SAP Community and AI tools during the exam.

From Preparing for SAP's AI certifications

a model whose trained parameters you may download and run, under its licence.

From Set up for Unit 12: local models

a machine-readable file describing every endpoint, field and error of an API.

From REST API design for AI services

the open standard for recording and sending traces, metrics and logs; SAP Cloud Logging accepts it.

From Observability for AI systems

the teams, roles, funding and reporting that run the roadmap.

From An AI roadmap and operating model for an SAP customer

reports on current business, read and acted on in the transactional system.

From Reporting and analytics architecture: what lives where in an SAP landscape

the decision, per framed problem, between fixing the cause, standard SAP, rules automation, AI and doing nothing.

From AI opportunity assessment: AI, automation or neither?

the tendency of appraisals to be too optimistic about costs, timing and benefits.

From The AI business case

deciding what runs next. In SAP's language it is also the name of a generative AI hub service; here it means the framework's job.

From Agent frameworks compared

in SAP's generative AI hub, the saved setup of prompt template, model and modules that an application calls.

From Building an evaluation harness

the part of SAP's generative AI hub that runs a model call through a fixed pipeline of modules.

From SAP Generative AI Hub and the orchestration service

the component that picks which agent handles a request. In SAP, the Joule Orchestrator.

From Joule agents and agent orchestration in SAP

the end-to-end process from customer request to cleared payment.

From Order to cash end to end

Cloud Foundry's folders for apps; a trial has one org and a space named dev.

From Set up for Unit 6: app tooling: Node.js, CAP and Cloud Foundry

a configured code, such as a company code or plant, that business documents refer to.

From How an SAP enterprise is wired: organizational structure

planned work to help people move to a new way of working.

From Adoption and change management for AI

a field in a role that takes an org unit value, such as a company code, to limit what the role can do.

From How an SAP enterprise is wired: organizational structure

OpenTelemetry's protocol for sending that data.

From Set up for Unit 10: observability and containers

what the project will not do, written down to stop scope creep.

From Problem framing: from 'we need AI' to a measurable problem

the S/4HANA document that says what to ship to whom and when.

From Warehouse and transportation execution

the list of approved actions waiting to be carried out in SAP.

From Capstone: an enterprise exception management platform

the end state in the system after a trial, such as what is in the approval queue.

From Evaluating agents: tasks, trajectories and tool calls

the word pieces the model writes. Each one takes time, so they drive both latency and cost.

From Latency budgets and performance

accepting wrong suggestions / overriding right ones.

From Adoption and change management for AI

doing well on training examples but badly on new ones.

From Machine learning in plain terms

text repeated between neighbouring chunks.

From Chunking and document preparation: from PDF to retrievable passages

the user decides something different from what the AI suggested.

From Evaluating in production: shadow mode, A/B tests and user feedback

the cost that half, or four in five, of the simulated outcomes stay under.

From Scoping and estimating an AI engagement

a library you can install, such as requests or sap-ai-sdk-gen.

From Python projects done right: structure, virtual environments and packages

ready-made code you install with pip.

From Set up your computer for this course

splitting a long list into pages the caller fetches one by one.

From REST API design for AI services

reading a large result a slice at a time.

From Calling your first SAP API

a judge picks the better of two answers to the same case.

From LLM evaluation fundamentals

SAP HANA's libraries for training classic machine learning models inside the database.

From Tabular AI and SAP's predictive models

one learned number in the model; model size is counted in parameters.

From The transformer architecture

the numbers a model learned. "0.6b" means about 0.6 billion. More usually means better and slower.

From Set up for Unit 12: local models

the chance that k runs of the same task all succeed. Use it when reliability matters.

From Evaluating agents: tasks, trajectories and tool calls

how long until total benefits catch up with total costs.

From The AI business case

the scheduled job that pays due supplier items and clears them.

From Procure to pay end to end

the summary a pay run sends to finance: wages, taxes and payables by account and cost assignment.

From Hire-to-retire: where people data meets logistics and finance

the busiest period's volume divided by the average period's.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

the link from a receipt to the requirement it covers, used to find which customer a late receipt affects.

From One exception, traced across the enterprise

p95 is the response time that 95 in 100 requests beat. p50 is the median.

From Observability for AI systems

the tasks that finish a month's accounting so the figures can be reported.

From Record to report, end to end

a failure that will repeat until someone changes something, such as a refused key or a wrong address.

From Errors, logging and debugging

importance measured by shuffling one field and seeing how much the model's accuracy drops.

From Explaining model predictions to the business

information about an identified or identifiable person. PII (personally identifiable information) is the common US term.

From Data security and PII in AI systems

the ID of an employee's employment in SAP HR and on time records.

From Hire-to-retire: where people data meets logistics and finance

a stretch of work with one question to answer, ending at a gate.

From Scoping and estimating an AI engagement

SAP's term for general availability that reaches customers in stages rather than all at once.

From Joule agents and agent orchestration in SAP

an open-source trace viewer built for AI applications.

From Set up for Unit 10: observability and containers

data that identifies a person, such as a name, email or phone number.

From Set up for Unit 11: security testing tools

real users doing real work with the tool for a set period, measured.

From Rapid prototyping: one day, one week, two weeks

some users work with the feature, some without, so the difference shows the effect.

From Embedded AI across SAP applications

naming one exact version, such as python-dotenv==1.2.4.

From Python projects done right: structure, virtual environments and packages

the automated steps a change goes through, such as test, gate and deploy.

From Shipping AI changes safely: CI/CD for prompts, models and agents

a marked gap in a template, such as {{?order}}, filled in at run time.

From Prompt and context engineering

the agent's to-do list for a task, kept up to date as it works.

From Multi-step agents: planning, state and recovery

the date goods are scheduled to leave the warehouse.

From Predicting late deliveries with SAP data

expected demand for a material in a plant, from a forecast or entered by hand.

From Plan to produce end to end

planned demand in S/4HANA that MRP plans against.

From Time-series forecasting for SAP planning

MRP's proposal to make something; it becomes a production order.

From Plan to produce end to end

the rule for how PIRs and sales orders work together, for example make-to-stock or make-to-order.

From Plan to produce end to end

a site where goods are produced, stored or planned.

From How an SAP enterprise is wired: organizational structure

a short session where you show interviewees what you heard and saw, so they can correct it.

From Discovery: interviewing the customer

two systems wired directly, each pair built and run separately.

From The integration platform and clean core: a target architecture for a hybrid landscape

collecting the top results from several methods so a labeler judges all of them.

From Evaluating retrieval

a public set of finished projects, each with a case study, a run path and a demo.

From Your FDE portfolio and career path

a public GitHub repository holding the work you choose to show.

From Set up for Unit 14: capstone environment and a public portfolio

a judge's preference for an answer because of where it appears.

From LLM evaluation fundamentals

numbers added at the bottom of the stack so the model knows word order.

From The transformer architecture

SAP's performance-based format, in which you complete tasks instead of answering multiple-choice questions.

From Preparing for SAP's AI certifications

share of flagged cases that were really positive.

From Classification and the metrics that matter

whether the right passage comes first.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

a filter that narrows results to what is relevant.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

share of the top k results that are relevant.

From Evaluating retrieval

a model that learns from past outcomes to predict a number or category.

From AI opportunity assessment: AI, automation or neither?

advanced AI capabilities paid for in AI Units, per use or per user.

From Embedded AI across SAP applications

an AI feature that consumes AI Units.

From Where AI creates value in S/4HANA processes

the first, expensive training on a huge general text collection.

From Build a tiny GPT

the one business number the project is judged on. Also called the north star or overall evaluation criterion.

From Defining success metrics with the business

forwarding the logged-on user's identity to the target system, so its own authorizations apply.

From Secrets, API authentication and OAuth

in garak, one family of attack prompts aimed at one weakness.

From Set up for Unit 11: security testing tools

a standard JSON shape for errors (RFC 9457), so all errors look alike.

From REST API design for AI services

writing a business problem as an actor, a process step, a metric, a baseline, a target, a date, guardrails and constraints.

From Problem framing: from 'we need AI' to a measurable problem

the person accountable for a business process and its results.

From Discovery: interviewing the customer

the end-to-end process from a need to a paid supplier.

From Procure to pay end to end

the order to make a quantity of a product, with components, operations and costs.

From Plan to produce end to end

the difference between an order's actual costs and the standard value of what it delivered.

From Plan to produce end to end

the instructions and input sent to a model in one call.

From Prompt and context engineering

reusing the processed start of a prompt across calls, to answer faster.

From Prompt and context engineering

input that changes a model's behaviour in ways its owner didn't intend.

From Prompt injection and tool poisoning

the most tokens each part of the prompt may use.

From Token economics

automatically rewriting a prompt to score better on a labelled dataset, without changing the model.

From When to fine-tune (and when not to)

SAP AI Core's store for versioned prompt templates.

From Prompt and context engineering

an input check for text that tries to override the app's instructions.

From SAP Generative AI Hub and the orchestration service

a reusable prompt with placeholders, stored and versioned.

From Prompt and context engineering

a name for the exact prompt in use, so answers can be traced to it.

From Building an AI API

the platform's suggested next step, with an explanation and evidence.

From Capstone: an enterprise exception management platform

where an agent's suggested changes wait for a person to decide.

From An agent that calls a RAP service

the organization that develops an AI system and places it on the market or puts it into service under its name.

From The EU AI Act: what SAP customers need to do and when

a field the model uses because it moves with the real cause, not because it is the cause.

From Explaining model predictions to the business

replacing identifiers so the link can be restored only with extra information kept apart.

From Data security and PII in AI systems

publishers send events to topics; subscribers say which topics they want.

From Event-driven agents: triggering AI from SAP business events

a request to merge one branch into another, with review.

From Git and GitHub for AI engineers: branches, pull requests and review

an internal request to buy, created by MRP or a person.

From Procure to pay end to end

a supplier's terms for one material, used as defaults in purchase orders.

From Master data that every module depends on

send your commits to the remote / bring others' commits down.

From Git and GitHub for AI engineers: branches, pull requests and review

a GitHub feature that blocks a push when it finds a known secret format.

From Set up for Unit 14: capstone environment and a public portfolio

the Python Package Index, the public website pip and uv download packages from.

From Python projects done right: structure, virtual environments and packages

the standard file that describes a Python project and what it needs.

From Python projects done right: structure, virtual environments and packages

a checkpoint at a phase transition where deliverables are reviewed before the program moves on.

From The solution architect on an SAP program

quality control data for one supplier and material, including a procurement block.

From Quality, projects and service: the processes between the big four

the record of a quality problem, with defects, causes, tasks and activities.

From Quality, projects and service: the processes between the big four

storing weights with fewer bits, plus a few scale values to map them back.

From Quantization and local models

adding related terms or synonyms to the question.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

changing the question before searching, for example into the words documents use.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

rows where the answer is missing and should be predicted. SAP-RPT marks them with [PREDICT].

From Tabular AI and SAP's predictive models

a place in the broker that holds events for one consumer until it confirms it has processed them.

From Event-driven agents: triggering AI from SAP business events

an offer to a customer, with prices and dates, which an order can copy.

From Order to cash end to end

retrieve relevant passages, add them to the prompt, generate an answer from them.

From RAG fundamentals: the architecture end to end

an open-source library for retrieval metrics.

From Set up for Unit 8: evaluation tools

SAP's model for building transactional services and apps in ABAP.

From ABAP Cloud and RAP for AI builders

the maximum number of calls one caller may make in a time window.

From Building an AI API

open standards for storing graph facts and for querying them.

From SAP Knowledge Graph and Business Data Cloud: giving AI a map of the business and data it can trust

a model that reorders the candidates for one question.

From Re-ranking and query rewriting: a second look at the candidates, a better question for the search

the share of intended users who used the tool in a period.

From Adoption and change management for AI

SAP's record of who read sensitive data and when.

From Data security and PII in AI systems

tools that look up, prepare a request for approval, or change data.

From SAP tools for agents

a tool that reads data and changes nothing.

From Set up for Unit 9: MCP and agent tooling

ready, ready after fixes, or not ready, for one named use case.

From Is your SAP data ready for AI? A readiness check you can run

the share of a theoretical saving that turns into real value.

From The AI business case

column types for embeddings; half vectors use half the space per number.

From SAP HANA Cloud vector engine: embeddings next to business data, searched with SQL

a short, fixed reason chosen when overriding, such as "customer payment expected".

From Designing human approval into AI workflows

the short sentence a user sees next to a score.

From Explaining model predictions to the business

share of real positive cases that were flagged.

From Classification and the metrics that matter

the share of the true top k matches that the index actually returns.

From Vector databases explained: indexes, trade-offs and how to choose

share of all relevant documents that appear in the top k.

From Evaluating retrieval

proving that counts and amounts in the target match the source.

From Data migration: strategy, objects and reconciliation

model answers saved once and reused in tests, so tests run without calling the model.

From Shipping AI changes safely: CI/CD for prompts, models and agents

attacking your own system on purpose to find weaknesses before others do.

From Red-teaming your AI system

a published, reusable design pattern, such as those in the SAP Architecture Center.

From Writing a solution design document for an AI use case

the agreed message a user sees for each kind of block.

From Guardrails and content filtering

something that used to work and broke after a change.

From Testing AI applications: unit and integration tests

an automatic check that fails a release if any attack does better than before.

From Red-teaming your AI system

rerunning known cases after each change to catch regressions.

From LLM evaluation fundamentals

the exact combination of prompt, model, settings and tools proposed for release.

From Shipping AI changes safely: CI/CD for prompts, models and agents

SAP promises stability for the object and lists it as usable.

From The SAP data model for AI engineers: tables, CDS views and OData

an interface SAP publishes and keeps stable across releases.

From The integration platform and clean core: a target architecture for a hybrid landscape

one SAP has published and keeps stable across upgrades.

From How SAP systems talk: integration styles and when to use each

changing custom code so it works, and stays supportable, on the new system.

From AI in S/4HANA migration projects

a copy of the repository on another computer, such as GitHub.

From Git and GitHub for AI engineers: branches, pull requests and review

copying records from the system of record to another system, on a schedule or on change.

From Hire-to-retire: where people data meets logistics and finance

a folder whose full change history Git tracks.

From Git and GitHub for AI engineers: branches, pull requests and review

running a failed message again after the cause is fixed.

From The integration lead's toolkit: inventory, specs, testing and cutover

how many of several tries of the same attack worked, such as 3 out of 5.

From Red-teaming your AI system

a unique label on each call, used to find it in logs later.

From Building an AI API

how long a document stays in the database before it can be archived.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

each step adds its result to the word's vector instead of replacing it.

From The transformer architecture

what stays open on a customer's account after a short payment.

From Order to cash end to end

a thing the API offers, such as a blocked order or an explanation, with its own address.

From REST API design for AI services

a separate workspace inside one SAP AI Core tenant, holding its own deployments and configurations.

From Running AI on SAP BTP in production

in SAP AI Core, the setting that picks the machine size, including GPU nodes.

From Set up for Unit 4: PyTorch and an optional free GPU

stores whole answers for identical requests.

From Caching for LLM applications

how long a record must be kept at all, set by law or policy.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

the date after which an exam can no longer be booked; valid holders get 12 more months.

From Preparing for SAP's AI certifications

the search step of RAG that picks passages for the model.

From Evaluating retrieval

scores for search results, such as recall (did we find the right items?) and MRR (how high was the first right item?).

From Set up for Unit 8: evaluation tools

trying again after a wait that grows each time. Jitter adds randomness so many clients don't retry at the same moment.

From Errors, logging and debugging

net benefit divided by cost.

From The AI business case

how easily a change can be undone, and at what cost.

From Designing human approval into AI workflows

a step that happens more than once in the same case.

From Process mining to find AI opportunities

SAP's subscription offering for the private cloud edition, bundling software, infrastructure and technical operations.

From Landscape and deployment: editions, systems and the migration approach

in SAP's AI ethics approach, standard, high-risk or red line; it decides how much review a use case needs.

From Writing a solution design document for an AI use case

a level, such as low, medium or high, that sets how much checking a use case needs.

From AI governance for SAP programs

distance from the median, scaled by the median absolute deviation, so outliers don't distort the yardstick.

From Anomaly detection in finance and procurement

how well the scores rank positives above negatives; 0.5 is random, 1.0 perfect.

From Classification and the metrics that matter

A named set of permissions, such as "order clerk", assigned to users

From CAP and side-by-side extensions for AI

a bundle of roles assigned to users or groups.

From SAP BTP foundations for AI builders

returning a deployment to its last configuration that worked.

From Running AI on SAP BTP in production

replacing running copies one at a time, so the service stays available.

From Deploying AI apps on SAP BTP

a group of teams that goes live together.

From Adoption and change management for AI

replacing a secret with a new one on a schedule or after a leak, and retiring the old one.

From Secrets, API authentication and OAuth

the web address the platform gives the app.

From Deploying AI apps on SAP BTP

logic that chooses a model for each request.

From Model selection and routing

the sequence of operations to make a product, each at a work center.

From Master data that every module depends on

software that repeats recorded clicks and keystrokes on screens.

From Agents that work SAP screens: computer use, RPA and when APIs are missing

how much recent data you may lose in a disaster, in minutes.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

a fusion rule that uses only each document's rank in each list.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

how long it may take to restore service after a disaster.

From Non-functional design for SAP: volumes, batch windows, availability, retention and audit

the written criteria a grader applies.

From LLM evaluation fundamentals

a tool that changes its description or behaviour after it was approved.

From Prompt injection and tool poisoning

software that applies written-down rules to structured inputs, such as decision tables in a workflow.

From AI opportunity assessment: AI, automation or neither?

the written agreement on what may be attacked, when, and by whom.

From Red-teaming your AI system

what a live AI solution costs every month: platform, people and model usage.

From Scoping and estimating an AI engagement

the written steps the team follows when an alert fires.

From Running AI on SAP BTP in production

where apps run: Cloud Foundry, Kyma or ABAP.

From SAP BTP foundations for AI builders

a sales organization, a distribution channel and a division together. Customer sales data is kept per sales area.

From How an SAP enterprise is wired: organizational structure

the set of made-up Calder data files the Unit 1B scripts read when run with --sample.

From Set up for Unit 1B: process tracing tools

draw a token at random, weighted by its probability.

From How LLMs generate text, and what that means for enterprise use

a test system with made-up data, safe for learning.

From Python and APIs for AI engineers

the SAP BTP service that runs and serves AI models, including the generative AI hub.

From Running AI on SAP BTP in production

the web screen for working with SAP AI Core, including trying prompts.

From Set up for Unit 5: SAP generative AI hub access

SAP Business Technology Platform, where SAP customers build and run extensions and AI services.

From Set up for Unit 3: neural network tools and an SAP BTP trial

an ABAP system that SAP runs in the cloud on SAP BTP.

From Set up for Unit 13: ABAP development tools and a coding agent

SAP's catalog of APIs, with documentation, a sandbox and Try Out (api.sap.com).

From Calling your first SAP API

SAP's umbrella, since May 2026, for its build, context and governance layers.

From The SAP Business AI landscape: where every piece sits

SAP's data foundation for AI and analytics.

From The SAP Business AI landscape: where every piece sits

SAP's managed observability service on BTP.

From Set up for Unit 10: observability and containers

SAP's own SDK for reaching generative AI hub models from Java, Python and TypeScript.

From Agent frameworks compared

SAP's ERP system before S/4HANA, part of SAP Business Suite 7.

From AI for SAP ECC customers before they migrate

an announced SAP control point for authentication, policy, routing, monitoring and auditing of all AI traffic.

From What SAP TechEd 2026 changed for AI builders

the scripting interface of SAP GUI for Windows that lets a program read and drive its screens.

From Agents that work SAP screens: computer use, RPA and when APIs are missing

SAP's cloud database on SAP BTP.

From Set up for Unit 7: vector search: HANA Cloud and a local store

the web tool for creating, starting and stopping SAP HANA Cloud instances.

From Set up for Unit 7: vector search: HANA Cloud and a local store

SAP's integration platform on SAP BTP.

From The integration platform and clean core: a target architecture for a hybrid landscape

SAP's paid learning subscription that includes exam attempts, practice systems and renewals.

From Preparing for SAP's AI certifications

a route SAP names for agentic access to its APIs, such as the MCP Gateway or Joule agents.

From SAP tools for agents

SAP's family of tabular foundation models ("relational pretrained transformer"), sold through the generative AI hub.

From Tabular AI and SAP's predictive models

a timed roleplay with an AI avatar, or a recorded video, graded on how well you advise or pitch.

From Preparing for SAP's AI certifications

the part of an order item with a confirmed date and quantity.

From Order to cash end to end

a description of the fields an answer must have, their types and their allowed values.

From Structured outputs and function calling

checking that an answer has the expected fields and types.

From Guardrails and content filtering

a named permission written into a token, such as "read orders".

From Secrets, API authentication and OAuth

the model's number for "how likely is the positive outcome". Often between 0 and 1.

From Classification and the metrics that matter

a software kit developers use to build MCP clients and servers.

From MCP security in 2026: lessons from the first exploited vulnerabilities

anything that proves identity: API key, password, client secret, private key, token.

From Secrets, API authentication and OAuth

a managed service that keeps secrets encrypted, controls access and logs use. SAP Credential Store is SAP BTP's.

From Secrets, API authentication and OAuth

splitting tasks so no one person controls a whole risky transaction.

From Roles and authorization design across modules

attention among the words of the same text.

From Attention, explained

returns a stored answer when a new question is similar enough in meaning.

From Caching for LLM applications

the agreed definitions of measures and dimensions that every report and AI tool uses.

From Reporting and analytics architecture: what lives where in an SAP landscape

finding the stored texts whose embeddings are closest to the embedding of a question.

From Embeddings and semantic similarity

closeness in meaning, measured between two embeddings.

From Embeddings and semantic similarity

the rule that the person who makes a change can't approve it.

From AI governance for SAP programs

a program that offers tools, resources and prompts.

From The Model Context Protocol (MCP)

a standard way for a server to send a stream of small messages over one HTTP response.

From REST API design for AI services

how the platform hands an app the credentials of a service, such as a database or SAP AI Core, without putting them in code.

From Deploying AI apps on SAP BTP

the record of the time, parts and expenses actually used.

From Quality, projects and service: the processes between the big four

the record that a service was performed, which works like a goods receipt for services.

From Procure to pay end to end

one copy of a service, created in a subaccount.

From SAP BTP foundations for AI builders

a small file of access details (an ID, a secret and two web addresses) for one service instance.

From Set up for Unit 5: SAP generative AI hub access

credentials for a service instance; a key for outside callers, a binding for a deployed app.

From SAP BTP foundations for AI builders

an agreed target for a signal, such as "95 in 100 answers within 3 seconds".

From Observability for AI systems

the document that plans service at a customer site and collects its costs and revenue.

From Quality, projects and service: the processes between the big four

the contract level of a service. For generative AI in production, SAP AI Core's extended plan.

From Running AI on SAP BTP in production

moving an order's costs to their final receiver, such as a cost center or a fixed asset.

From Maintain to settle: plant maintenance and assets

the AI runs on real inputs but its output is only logged, never shown.

From Evaluating in production: shadow mode, A/B tests and user feedback

watching someone do their real work, in their place, and noting each step and its time.

From Discovery: interviewing the customer

work many use cases need, such as model access, identity, logging and a use-case register.

From An AI roadmap and operating model for an SAP customer

an application that runs next to the ERP, usually on SAP BTP, and talks to it through interfaces.

From AI for SAP ECC customers before they migrate

an app outside S/4HANA, usually on SAP BTP, that uses S/4HANA through APIs or events.

From Fit-to-standard and extension choices: where each requirement should live

a wrong value that passed every check.

From Document AI, built and bought: from supplier invoice PDF to checked SAP data

a query that runs without error but answers a different question than the one asked.

From Asking SAP data questions in plain language: text-to-SQL and its limits

finding the rows whose vectors are closest to a question's vector.

From SAP HANA Cloud vector engine: embeddings next to business data, searched with SQL

a documented change in S/4HANA that can break old code or processes.

From AI in S/4HANA migration projects

one model choosing tools in a loop until the job is done.

From One order-exception agent, built three ways

in PFCG, a single role holds authorizations; a composite role bundles single roles; a derived role copies a master role and sets its own org values.

From Roles and authorization design across modules

one place that holds the solution's facts and numbers. Every version of the story draws from it.

From Technical storytelling: one architecture, five audiences

SAP S/4HANA's framework for notifying users about issues that need attention.

From Capstone: an enterprise exception management platform

a group of cases that share a trait, such as all credit-block orders. Scores per slice show where a change hurts.

From Building an evaluation harness

the person accountable for how a program's decisions fit together into one working, upgradeable solution.

From The solution architect on an SAP program

the agreed, written plan for a solution: problem, metric, options, design, risks, cost and decisions.

From Writing a solution design document for an AI use case

the set of documents an architect brings to a design review: summary, landscape, process trace, interfaces, risks and AI opportunities.

From One exception, traced across the enterprise

the group that approves and records deviations from standard.

From Fit-to-standard and extension choices: where each requirement should live

the allowed sources of supply for a material in a plant, with validity dates.

From Master data that every module depends on

the five options (use, configure, extend, build, partner), ordered from least to most you own.

From Buy vs. build for SAP AI

one step inside a trace, with a start, an end and labelled facts.

From Set up for Unit 10: observability and containers

an agent with one narrow job and only the tools for it.

From Joule agents and agent orchestration in SAP

a small group gets the change first and is compared with the rest in the same period.

From Evaluating in production: shadow mode, A/B tests and user feedback

everything a run knows so far: the plan, the facts it read, the actions it queued, and the budget used.

From Multi-step agents: planning, state and recovery

the contract document listing scope, deliverables, phases, exit criteria, assumptions, pricing and change control.

From Scoping and estimating an AI engagement

evidence that a difference between two options is unlikely to be luck.

From Evaluating retrieval

a three-digit number in every answer; 200 means OK, 401 means "who are you?", 404 means "not found".

From Python and APIs for AI engineers

the short yearly test that keeps a certificate valid; included with SAP Learning Hub.

From Preparing for SAP's AI certifications

the plainest way a client and server talk, through the server process's own input and output.

From Set up for Unit 9: MCP and agent tooling

one update of the weights.

From Gradient descent, by hand

the most steps or tool calls a task may take before it counts as a failure.

From Evaluating agents: tasks, trajectories and tool calls

goods that have left one plant and not reached the other.

From Intercompany and multi-company scenarios

the result, agreed in advance, that means stop or rethink.

From Problem framing: from 'we need AI' to a measurable problem

what ends the loop: a final answer, a step limit, an error or a repeated request.

From Agents from first principles

documents that pass every check and need no person.

From Document AI, built and bought: from supplier invoice PDF to checked SAP data

sending the answer in pieces as it is generated.

From How LLMs generate text, and what that means for enterprise use

a setting that makes the answer follow the schema exactly.

From Structured outputs and function calling

a test of several connected steps or systems in a chain, short of the whole process.

From The integration lead's toolkit: inventory, specs, testing and cutover

cutting at headings and keeping tables and lists whole.

From Chunking and document preparation: from PDF to retrievable passages

log entries written as data (for example one JSON object per line), so tools can search by field.

From Errors, logging and debugging

a model answer that follows a schema you supply.

From Structured outputs and function calling

a space in one region where apps run and services are used.

From SAP BTP foundations for AI builders

the fixed set of test cases. Also called a golden set or test set.

From Building an evaluation harness

the supplier's answer to a purchase order, with the quantity and date it commits to.

From One exception, traced across the enterprise

a number that helps explain the result but doesn't decide it.

From Defining success metrics with the business

the value an assumption must reach for the decision to flip.

From The AI business case

the caller waits for the answer.

From How SAP systems talk: integration styles and when to use each

made-up data shaped like SAP data; results from it show the method, not a business outcome.

From Your FDE portfolio and career path

the instruction message that sets the rules for the whole call.

From Prompt and context engineering

the one system where a piece of data is created and changed; others receive copies.

From Hire-to-retire: where people data meets logistics and finance

a timed practical exam in SAP systems, graded from the work you produce.

From Preparing for SAP's AI certifications

where SAP physically stores data, with technical names such as VBAK.

From The SAP data model for AI engineers: tables, CDS views and OData

a tabular foundation model from Prior Labs, an SAP company since July 2026.

From Tabular AI and SAP's predictive models

a model pretrained on many tables that predicts a column of a new table without being trained on it.

From Tabular AI and SAP's predictive models

the slowest few requests, the ones above p95 or p99.

From Latency budgets and performance

the promised value or change, with a date.

From Defining success metrics with the business

one test: an input, plus what success means.

From Evaluating agents: tasks, trajectories and tool calls

the status that says the work is done.

From Maintain to settle: plant maintenance and assets

how it is built: technology, mapping, security, retries, monitoring.

From The integration lead's toolkit: inventory, specs, testing and cutover

explaining one solution to several audiences, each with the picture, numbers, risks and ask that fit their decision.

From Technical storytelling: one architecture, five audiences

a system account an application uses to call SAP. SAP checks the account, not the person behind it.

From Grounding on SAP data without breaking authorizations: live queries, access labels and the late check

a setting that sharpens (low) or flattens (high) the probabilities before sampling.

From How LLMs generate text, and what that means for enterprise use

extraction rules tied to one exact layout.

From Document AI, built and bought: from supplier invoice PDF to checked SAP data

a text window where you type commands.

From Set up your computer for this course

a small program that checks another program and reports pass or fail.

From Testing AI applications: unit and integration tests

any stand-in for a real part during tests. Fakes, stubs and mocks are kinds of test doubles.

From Testing AI applications: unit and integration tests

examples held back to check the model on cases it hasn't seen.

From Machine learning in plain terms

choosing one label from a fixed list for a piece of text.

From Classifying text: LLMs, embeddings or a trained model?

a model translating a question into an SQL query.

From Asking SAP data questions in plain language: text-to-SQL and its limits

the original request, in the requester's own words.

From Problem framing: from 'we need AI' to a measurable problem

the list of who might attack, what they want, and how they could get in.

From Red-teaming your AI system

a low, likely and high value for one task.

From Scoping and estimating an AI engagement

comparing purchase order, goods receipt and invoice before paying.

From Procure to pay end to end

the cut-off score at which a case is flagged.

From Classification and the metrics that matter

money spent on something that must be rebuilt at the migration.

From AI for SAP ECC customers before they migrate

paying for the days worked at agreed rates, often with a cap.

From Scoping and estimating an AI engagement

values measured at regular intervals, such as weekly goods issues for one material.

From Time-series forecasting for SAP planning

how long until the model produces the first piece of its answer.

From Latency budgets and performance

the longest time a stored item may be reused.

From Caching for LLM applications

train on older records, test on newer ones.

From Predicting late deliveries with SAP data

a fixed period of work whose length doesn't move; the scope does.

From Rapid prototyping: one day, one week, two weeks

how long the flow waits for an answer before giving up.

From Calling AI from SAP Integration Suite

a piece of text, often part of a word, that the model processes as one unit.

From Attention, explained

the most tokens you allow one call's context to use.

From Prompt and context engineering

swapping a token for a new, narrower one that names both the agent and the user.

From MCP security in 2026: lessons from the first exploited vulnerabilities

a server forwarding the caller's token to another system unchanged. The MCP guidance forbids it.

From MCP security in 2026: lessons from the first exploited vulnerabilities

how fast a model writes. It decides whether an answer feels instant or slow.

From Set up for Unit 12: local models

how far an invoice may differ from the order or receipt before it is blocked.

From Procure to pay end to end

one action your application offers the model, described by a name, a description and an input schema.

From Structured outputs and function calling

MCP's labels for how a tool behaves, such as read-only or destructive. Hints, not enforcement.

From Tool design for agents

the name, description and input schema the model reads. It is sent with every model call.

From Tool design for agents

the cycle of model asks, application runs, model reads the result, until the model answers.

From Structured outputs and function calling

hiding instructions to the model inside a tool's description or results.

From The Model Context Protocol (MCP)

a list of requests with the right tool for each, used to measure whether the model picks correctly.

From Tool design for agents

the first k results; k is often the number of passages sent to the model.

From Evaluating retrieval

sample only from the most likely tokens that together reach a probability p, such as 0.9.

From How LLMs generate text, and what that means for enterprise use

build cost plus run cost over the life of the solution.

From Buy vs. build for SAP AI

the share of cases with no manual step after creation.

From Process mining to find AI opportunities

the record of every step: what was decided, what ran, what came back.

From Agents from first principles

the report Python prints when it stops: which lines were running, ending with the error type and message.

From Errors, logging and debugging

the company on the other side of an intercompany posting; matching depends on it.

From Record to report, end to end

the step where the model learns from past examples.

From Machine learning in plain terms

one input with the ideal output, written or approved by a person.

From When to fine-tune (and when not to)

one round of "predict, measure the error, adjust the weights".

From Set up for Unit 4: PyTorch and an optional free GPU

the full record of a trial: every tool call, its result and the final answer.

From Evaluating agents: tasks, trajectories and tool calls

SAP's code for a kind of posting, such as BSX for stock or WRX for GR/IR clearing.

From Document flow and how logistics becomes finance

the price one group company charges another; set by policy, checked by tax authorities.

From Intercompany and multi-company scenarios

the model design built from attention layers; the next topic in this unit.

From Attention, explained

a failure that may go away on its own, such as a timeout, a 429 "too many requests" or a 503 "service unavailable".

From Errors, logging and debugging

a time-limited SAP subscription that keeps ECC-based systems running after 2030, under conditions.

From AI for SAP ECC customers before they migrate

converting a company's figures into group currency.

From Intercompany and multi-company scenarios

a recorded package of changes moved from one system to the next.

From Landscape and deployment: editions, systems and the migration approach

SAP's container for moving a change from one system to the next.

From Set up for Unit 13: ABAP development tools and a coding agent

one run of the agent on one task. Run several per task.

From Evaluating agents: tasks, trajectories and tool calls

the time left before a trial account or service is suspended or deleted.

From Set up for Unit 14: capstone environment and a public portfolio

a line in a diagram where data passes from one owner's control to another's, such as from the SAP system to a model provider.

From Technical storytelling: one architecture, five audiences

a bot that runs with no person present, usually on a server.

From Agents that work SAP screens: computer use, RPA and when APIs are missing

OWASP's name for uncontrolled model use that causes outages or runaway cost.

From AI gateways: governing model and agent traffic

multiplies inputs by weights, adds a bias, then applies an activation function.

From Neural networks from scratch

a test of one small piece, such as one rule, with nothing external involved.

From Testing AI applications: unit and integration tests

the single S/4HANA table (ACDOCA) for finance and controlling line items.

From Document flow and how logistics becomes finance

content without fixed fields, such as a PDF purchase order or an email.

From Where AI creates value in S/4HANA processes

learning without labeled examples of the problem.

From Anomaly detection in finance and procurement

the accept or reject decision that closes an inspection lot and posts its stock.

From Quality, projects and service: the processes between the big four

the business person accountable for one use case's result.

From An AI roadmap and operating model for an SAP customer

the list of every AI use, with owner, tier, status and vendor details.

From AI governance for SAP programs

passing the business user's identity to SAP, so the call runs as that person.

From An agent that calls a RAP service

a fast tool that manages Python projects, environments and lock files in one place.

From Python projects done right: structure, virtual environments and packages

a check that runs before saving and can reject the change with a message.

From ABAP Cloud and RAP for AI builders

a material setting that groups materials for account determination.

From Document flow and how logistics becomes finance

SAP AppHaus's chart for comparing use cases, usually business value against complexity.

From AI opportunity assessment: AI, automation or neither?

the period in which an item has value, for example only until a migration replaces the process.

From An AI roadmap and operating model for an SAP customer

a set of cases that follow the same sequence of steps.

From Process mining to find AI opportunities

an ordered list of numbers. An embedding is a vector.

From Embeddings and semantic similarity

the SAP HANA Cloud features that store and compare embeddings.

From SAP HANA Cloud vector engine: embeddings next to business data, searched with SQL

a store that keeps embeddings and searches them quickly.

From Embeddings and semantic similarity

ranking by closeness of embeddings, which represent meaning.

From Hybrid search and metadata filtering: keywords, vectors and the filters that keep answers in scope

a database that stores embeddings and finds the closest ones to a question.

From RAG fundamentals: the architecture end to end

an SAP HANA Cloud function that turns text into an embedding inside the database.

From SAP HANA Cloud vector engine: embeddings next to business data, searched with SQL

anything that checks the agent's work automatically: a test, a linter, a syntax check.

From Coding agents for SAP developers

one audience's picture of the solution, showing some parts and leaving out others.

From Technical storytelling: one architecture, five audiences

S/4HANA's layered set of CDS views.

From The SAP data model for AI engineers: tables, CDS views and OData

a private folder with its own Python and packages for one project.

From Python projects done right: structure, virtual environments and packages

a tool that runs many known attacks automatically and reports which worked.

From Set up for Unit 11: security testing tools

the time between one step and the next in a case.

From Process mining to find AI opportunities

a package of warehouse tasks for one worker or one trip.

From Warehouse and transportation execution

the warehouse's copy of a delivery, called an outbound delivery order in EWM.

From Warehouse and transportation execution

one move of goods from a source bin to a destination, confirmed when done.

From Warehouse and transportation execution

a RAP service binding meant for programs rather than a user interface.

From An agent that calls a RAP service

one of the learned numbers inside a model. Billions of them make up the model file.

From Quantization and local models

the numbers inside a model that training adjusts.

From Gradient descent, by hand

the numbers training adjusts.

From Neural networks from scratch

a test where a person secretly produces the system's outputs, to learn whether users want them before building.

From Rapid prototyping: one day, one week, two weeks

the hierarchy of a project; each WBS element can collect costs.

From Quality, projects and service: the processes between the big four

the costs on orders that aren't finished at period end.

From Plan to produce end to end

a sequence of steps fixed in code, with the model called at set points.

From Agents from first principles

the ranked list of open deliveries a planner checks.

From Predicting late deliveries with SAP data

a tool that changes something, such as creating a request. It needs the strongest checks.

From Structured outputs and function calling

asking how to make a chosen fix work instead of describing the underlying problem.

From Problem framing: from 'we need AI' to a measurable problem

giving another platform access to data without making a second copy.

From SAP Knowledge Graph and Business Data Cloud: giving AI a map of the business and data it can trust

asking an LLM with no examples / a few examples in the prompt.

From Classifying text: LLMs, embeddings or a trained model?