Orchestrate

Report an issue

Something wrong, out of date or unclear? Tell us. Corrections are prioritized over new topics. You don't need to sign in.

0 / 2000

Don't include personal or confidential company details. Reports are reviewed with AI assistance.

US state AI laws for an SAP deployment

State and city laws now regulate AI decisions about people. See which reach an SAP HR or credit agent, from when, and what to document.

Updated Nov 11, 2026Foundational 9 minDeep 36 min
Foundational layer · 9 min read

What changed, and when

There is still no federal US statute written for AI decisions about people. Existing anti-discrimination and credit laws still apply, but the AI-specific rules come from states and one city, and in 2026 several of them firmed up:

Date What changed
1 January 2026 Texas's Responsible AI Governance Act (HB 149) took effect, as did Illinois's amendment on AI in employment decisions.
14 May 2026 Colorado signed SB 26-189. It repealed the 2024 Colorado AI Act before that law ever applied and replaced it with a narrower, notice-based law on automated decision-making technology (ADMT).
30 September 2026 California signed SB 947, the "No Robo Bosses Act", which requires a person to corroborate AI-driven discipline and termination from 1 July 2027 (SHRM).
1 January 2027 Colorado's law applies to decisions made from this date. California's privacy rules on ADMT for "significant decisions" apply. New York's RAISE Act reporting starts for large frontier AI developers.

Two headline laws, California's SB 53 and New York's RAISE Act, bind companies that build frontier models. For an SAP customer they matter when choosing a model vendor, not when designing an agent.

At federal level, a December 2025 executive order told the Justice Department to challenge state AI laws and asked Congress for a preempting law (White House). An executive order does not cancel state law by itself, and as of November 2026 we found no preempting statute. Plan to comply with the state laws, and watch the courts.

Why it matters to the business

The state laws regulate decisions about individuals: applicants, employees and consumers. They don't care whether the AI runs in SAP SuccessFactors, on SAP BTP or in a spreadsheet macro. What they care about is whether its output shapes a decision about a person's job, credit, housing, insurance, health care or education.

Take the fictional Calder Pumpworks, the course's running company. Calder hires assemblers and test technicians at its plant in Rockford, Illinois (CP11), sales staff in New York City, and field service technicians who live in Colorado, California and Texas. It is building two agents:

  • An applicant screening agent that ranks applicants for open requisitions and rejects the lowest scores. In 2027 that hiring decision is covered in Colorado, California, Illinois and New York City at once, each with a different notice.
  • A credit agent that proposes credit limits for distributors. Distributors are companies, so the consumer-focused laws largely don't reach it. But Calder also sells spare parts on account to sole-trader contractors, and a sole trader is a person. Score them, and the agent becomes a consequential decision about an individual.

The costs are mostly enforcement and trust, not giant fines. Colorado treats a violation as a deceptive trade practice with civil penalties up to USD 20,000 per violation (fiscal note). Texas penalties for violations that can't be cured start at USD 80,000. Illinois lets applicants file discrimination charges and then sue.

How SAP does it

As of November 2026, SAP's Responsible AI page describes a Global AI Ethics Policy, an AI Ethics impact assessment for every SAP AI use case, an ISO/IEC 42001 certificate and alignment work for the EU AI Act. It cites NIST frameworks. It does not name any US state AI law.

SAP's announced governance tooling is also EU-first. SAP says its AI Governance Assistant in SAP AI Agent Hub supports EU AI Act assessments and produces drafts "for legal and compliance review" (SAP News, 30 October 2026). No US state module is announced.

What SAP does give you is control over where AI acts. In SuccessFactors Recruiting, an administrator switches AI-assisted skills matching on per feature, and a business rule decides which requisition templates use it (SAP Learning).

SAP builds the feature; you use it to decide about people in your states. The notices, human review, corrections and records are yours. For an agent your team builds on BTP, you are both developer and deployer.

The laws on one page

Law Who it binds Trigger Core duties From
Colorado SB 26-189 Developers and deployers ADMT that materially influences a consequential decision about a Colorado consumer, such as hiring or lending Notice before use; within 30 days of an adverse outcome, explain the decision and the tool's role; correction of wrong data and human review on request; records for 3 years 1 Jan 2027
California CCPA ADMT rules Businesses under the CCPA ADMT that replaces or substantially replaces human decision-making in a significant decision Pre-use notice; opt-out unless an exception applies (such as a human appeal); access to the logic and outcome; risk assessment 1 Jan 2027
California SB 947 Employers Primary reliance on an automated decision system for discipline or termination A person corroborates the output first; written notice; description of the data on request 1 Jul 2027
Illinois Human Rights Act (HB 3773) Employers AI used in hiring, promotion, discipline or discharge Notice; no discriminatory effect 1 Jan 2026
NYC Local Law 144 Employers and agencies hiring in NYC Automated employment decision tool Yearly independent bias audit, published summary, notice 10 business days before use 5 Jul 2023
Texas TRAIGA Developers and deployers; disclosure only for government and health care Intent to discriminate or manipulate Don't build or deploy with that intent; keep testing records 1 Jan 2026
California SB 53, New York RAISE Act Frontier model developers Training very large models Safety frameworks, reports, incident reporting 2026 and 2027

Questions to ask

  1. Which of our AI features or agents shape decisions about applicants, employees or individual customers, and in which states do those people live or work?
  2. For each one, does the AI decide, or does a person who understands the output and has authority to change it make the decision?
  3. Who sends the Colorado adverse-outcome disclosure within 30 days, and from which system?
  4. Can an applicant or customer ask for a human review and a correction of wrong data? Who handles it, and how fast?
  5. Do we keep each AI decision's version, inputs, output and reviewer for at least three years?
  6. Who watches the court cases and Colorado's rules, and updates the register when they change?

Common misconceptions

  • "The federal executive order wiped out state AI laws." It directs challenges and funding pressure. Until a court strikes a law down or Congress passes a preempting statute, the state law applies.
  • "Colorado repealed its AI Act, so Colorado is off the table." It replaced it. The new law is narrower but still requires notices, explanations, corrections, human review and records from 1 January 2027.
  • "A human clicks approve, so no AI law applies." California's carve-out needs a reviewer who understands and weighs the output and can change the decision. Colorado covers tools that materially influence a decision, not only fully automated ones.
  • "B2B credit is never covered." Company credit is generally outside the consumer laws. A sole trader is a person, so check before scoring them.
  • "Frontier AI laws apply to us because we use a frontier model." They bind the model's developer. Your duties come from what you do with the model.

Key terms

All course terms, A to Z, in the glossary

  • Automated decision-making technology (ADMT): technology that uses personal data and computation to produce outputs (scores, rankings, recommendations) that make or assist a decision about a person. Colorado and California each define it in their own way.
  • Consequential decision (Colorado): a decision about a consumer's access, eligibility, selection or compensation in a covered domain, such as employment or lending.
  • Significant decision (California): a decision about financial or lending services, housing, education, employment or health care, among others.
  • Adverse outcome: a decision that denies or materially reduces access to an opportunity or service, or gives materially worse terms.
  • Developer and deployer: who builds an AI system, and who uses it to decide.
  • Frontier developer: under SB 53, a company that trains a model with more than 10^26 computing operations.
  • Preemption: federal law overriding state law. It needs a statute or a court ruling, not only an executive order.

Check yourself

Pick one answer for each question. The explanation appears after you choose.

1A federal executive order of December 2025 targets state AI laws. What should a US SAP program assume for 2027 planning?

Answer: A. An executive order directs agencies; it does not by itself cancel state law. As of November 2026 no preempting statute had passed, so the state laws stand unless a court rules otherwise.

2Colorado repealed its 2024 AI Act in May 2026. What does that mean for a Calder screening agent that rejects Colorado applicants?

Answer: A. SB 26-189 repealed and replaced the 2024 law. It drops impact assessments but keeps notice before use, a 30-day explanation after an adverse outcome, correction and human review on request, and three-year records.

3Why does Calder's credit-limit agent for distributors mostly fall outside the consumer AI laws?

Answer: C. The laws regulate decisions about consumers, applicants and employees. A company is not a person, but a sole trader is, so scoring sole-trader contractors brings the agent back in scope.

4A vendor says "a recruiter clicks approve on every AI ranking, so California's ADMT rules don't apply." What do you ask?

Answer: A. California carves out decisions with real human involvement: a reviewer who understands the output, weighs it with other information and has authority to change the result. A rubber-stamp click doesn't meet that.

5Which law is mainly a question for your model vendor rather than your HR agent design?

Answer: A. SB 53 binds developers of frontier models. Ask your vendor for its published framework; your own duties come from the decisions you use the model for.

6Calder plans an agent that drafts written warnings from attendance data for California staff. What does SB 947 add from July 2027?

Answer: C. SB 947 bars relying solely on an automated decision system for discipline or termination. A person must examine and corroborate the output, and the worker gets written notice. The yearly bias audit is New York City's rule for hiring tools.

7What is the most useful first step for a program with AI in HR and credit across several states?

Answer: C. Every law here is triggered by the kind of decision and where the person is. A register with those facts lets counsel see which laws apply and keeps notices, reviews and records owned.
Deep layer · 36 min read

Mental model: a decision about a person, in a place

The EU AI Act asks "which tier is this use case?" The US state laws ask something simpler and messier: is an automated output shaping a decision about a person, and where is that person?

Three facts about each use case decide almost everything:

  1. The decision. Hiring, promotion, discipline, credit for an individual, housing, insurance. Back-office automation (invoice extraction, order status) rarely triggers these laws.
  2. The AI's role. Does it decide, is it relied on primarily, or does it assist a person who weighs it? The laws draw their lines at different points on that scale.
  3. The states. Where the applicant, employee or customer lives or works. One agent can fall under four regimes at once.

Whether you are the developer or the deployer then decides which duties are yours, just as in the EU AI Act topic.

flowchart LR
  U[AI use case] --> D{Decision<br/>about a person?}
  D -- no --> L[Low exposure:<br/>keep in register]
  D -- yes --> R{AI role?}
  R --> S{Which states?}
  S --> O[Duties per law,<br/>with dates]

How it works

Colorado: notice, explanation, correction, records

SB 26-189 was introduced on 1 May 2026 and signed on 14 May 2026. It repeals and re-enacts the Colorado Consumer Protection Act sections created by the 2024 Colorado AI Act (SB 24-205), which never took effect.

Scope. ADMT is technology that processes personal data and uses computation to produce outputs such as predictions, recommendations, classifications, rankings or scores used to make or assist a decision about an individual. The fiscal note lists exclusions, including some cybersecurity and web tools, spreadsheets that need human analysis, and tools that communicate with consumers or summarize information for later human review. A consequential decision relates to access, eligibility, selection or compensation in a covered domain: education, employment, residential real estate, financial and lending services, insurance and benefits, health care, and essential government services.

Deployer duties (the company using the ADMT):

  • A clear and conspicuous notice before ADMT is used to materially influence a consequential decision.
  • After an adverse outcome (a denial, termination, material reduction or materially worse terms), within 30 days: a plain-language description of the decision and the ADMT's role, how to ask for more information, and the person's rights.
  • On request, after an adverse outcome: access to and correction of factually incorrect personal data (not opinions, predictions or scores), and meaningful human review where commercially reasonable. Venable reads the reviewer requirements as authority to approve, modify or override, training for the review, and not simply accepting the output (Venable).
  • Records reasonably necessary to show compliance, kept for at least three years.

Developer duties: technical documentation for deployers covering intended and known harmful uses, categories of training data, known limits and risks, instructions including human review, information the deployer needs for its notices, and notice of material changes. The 2024 law's impact assessments and filings with the Attorney General are gone (McDermott).

Enforcement. The Attorney General enforces it as a deceptive trade practice, with civil penalties up to USD 20,000 per violation under the Consumer Protection Act. Before suing, the Attorney General must give notice and 60 days to cure, except for knowing or repeated violations; the cure right ends on 1 January 2030. There is no private right of action, but existing anti-discrimination claims are unaffected. Deployers already under comparable federal notice rules for credit or education decisions are treated as meeting the disclosure duties.

Status in November 2026. The Attorney General must adopt rules before 1 January 2027. The proposed rules were filed on 11 August 2026, an interim draft followed on 6 October, and the hearing and comment deadline were on 26 October 2026; the rulemaking page showed no adoption date when read on 11 November. Separately, xAI sued over the 2024 law on 9 April 2026, the Justice Department filed its own complaint on 24 April, and a stipulated order of 27 April paused enforcement of the 2024 law pending rulemaking and a ruling. McDermott expects both to challenge the new law too. Design to the statute; adjust the wording of notices when the final rules appear.

California: ADMT in privacy law, and humans before firing

CCPA regulations on ADMT. The California Privacy Protection Agency's regulations took effect on 1 January 2026. Businesses that use ADMT to make significant decisions must comply with the ADMT requirements from 1 January 2027 (CPPA). The CPPA's FAQ describes the consumer side as notice, an opt-out "where applicable", and access to meaningful information about how the ADMT worked and affected the person.

Per Manatt's summary of the final text:

  • ADMT is technology that processes personal information and uses computation to replace or substantially replace human decision-making. Decisions with real human involvement are carved out: the reviewer must understand and analyze the output, weigh it with other information, and have authority to make or change the decision.
  • A pre-use notice covers the purpose, the categories of personal information, the intended outputs, how humans take part and the person's rights.
  • The opt-out has exceptions: a clearly described appeal to a human who can overturn the decision, and decisions made solely for admission, hiring, allocation of work or compensation, provided the ADMT works for its purpose and doesn't discriminate unlawfully.
  • Using ADMT for a significant decision needs a documented risk assessment. Processing that started before 2026 has until 31 December 2027; attestations are due from 1 April 2028.

The CCPA covers California residents, including applicants and employees, when the company meets the CCPA's business thresholds. Check those with counsel.

SB 947, the No Robo Bosses Act. Signed on 30 September 2026 and operative on 1 July 2027 (SHRM). An employer may not rely solely on an automated decision system for discipline or termination. When it primarily relies on one, a person must examine and corroborate the output with other information, such as personnel records or interviews; output that can't be corroborated, or that is inaccurate, can't be the basis. The worker gets written notice and a contact, and can ask for a meaningful description of the data used. SB 947 also bars using such systems to infer protected status.

Illinois and New York City: employment rules already in force

Illinois. HB 3773 amended the Illinois Human Rights Act from 1 January 2026. Employers may not use AI in a way that discriminates against protected classes in recruitment, hiring, promotion, renewal, discharge, discipline or selection for training, and must give notice when AI is used to influence or facilitate those decisions. As presented at the Illinois Human Rights Commission in April 2026, the Department of Human Rights' notice rules were still proposed. Applicants and employees can file charges and then sue.

New York City Local Law 144. Since enforcement began on 5 July 2023, an employer or agency may use an automated employment decision tool only if it had a bias audit within one year of use, a summary of the results is published, and candidates got notice at least 10 business days before use (DCWP).

Texas: intent, not impact

Texas's chapter 552 of the Business and Commerce Code, added by HB 149, took effect on 1 January 2026 (Texas statutes). For a private manufacturer it works differently from Colorado:

  • The consumer disclosure duty applies to governmental agencies and to health care providers, not to private companies in general.
  • The prohibitions turn on intent: AI developed or deployed to manipulate people into harm, to impair constitutional rights, or "with the intent to unlawfully discriminate". The statute says "a disparate impact is not sufficient by itself" to show intent.
  • Only the Attorney General enforces it. After a complaint, it can demand a description of the system, data, metrics, limits and safeguards. There is a 60-day cure period; penalties run from USD 10,000 to 12,000 for curable violations, USD 80,000 to 200,000 for uncurable ones, and up to USD 40,000 a day for continuing ones.
  • There is no liability where a violation is found through testing or red-teaming, or with substantial compliance with the NIST Generative AI Profile or another recognized framework.

The practical consequence: your testing records are your defense. Keep them.

Frontier laws, and the federal picture

California SB 53, signed on 29 September 2025, binds frontier developers: those training models with more than 10^26 operations. Large frontier developers (above USD 500 million in revenue) carry more duties. They publish frameworks and transparency reports and report critical safety incidents within 15 days; penalties reach USD 1 million per violation (FPF). New York's RAISE Act targets large frontier developers too: safety protocols, 72-hour incident reports and an oversight office in the Department of Financial Services, with compliance and reporting from January 2027 (Governor of New York, 21 September 2026). Neither binds a company that only calls a model through an API.

Federal. Executive Order 14365 of 11 December 2025 set up an AI Litigation Task Force, asked Commerce to evaluate state AI laws and called for preempting legislation. The order does not preempt anything on its own; preemption needs a statute or a court ruling. As of a September 2026 review we read, no preempting statute had passed. Treat a court ruling against a specific state law as the event that changes your register.

Build it yourself: check your AI register against the state laws

You will build radar/us_state_ai_check.py. It reads a list of AI use cases with the states they touch, flags which laws need counsel's look, lists the duties with start dates, shows the EU AI Act analogue for each use case, and drafts a Colorado adverse-outcome notice for counsel to edit.

It is a triage tool. It turns the rules above into explicit logic so nothing is missed and every flag has a written reason. It does not decide whether a law applies; counsel does.

Before you start: complete Set up your computer for this course. It creates your orchestrate-course folder with its .venv. This lab uses only Python's built-in modules, so there is nothing new to install. If you did the EU AI Act topic, keep its radar folder; this script sits next to ai_act_register.py.

flowchart LR
  C[us_ai_use_cases.csv] --> K[Rules per law]
  K --> F[Flags:<br/>APPLIES / CHECK]
  K --> T[Duties<br/>with dates]
  F --> R[Register:<br/>CSV + Markdown]
  T --> R
  K --> N[Colorado notice<br/>draft]

What you need

  • Your course folder orchestrate-course with its .venv.
  • About 35 minutes.
  • Cost: free. No account, no key, no network call.

Step 1: Open your course folder and turn on the virtual environment

  1. Open VS Code, choose File > Open Folder, and open orchestrate-course.

  2. Open a terminal: Terminal > New Terminal.

  3. If the prompt doesn't start with (.venv), turn it on:

    • Windows (PowerShell):

      .venv\Scripts\Activate.ps1
    • macOS / Linux:

      source .venv/bin/activate
  4. Check Python:

    python --version

    You should see Python 3.10 or newer. On macOS or Linux without .venv, use python3 in every command.

Run every command in this topic from the course folder, not from inside radar.

Step 2: Create the script

  1. Create the radar folder if it doesn't exist yet:

    • Windows (PowerShell):

      New-Item -ItemType Directory -Force radar
    • macOS / Linux:

      mkdir -p radar
  2. In VS Code, create a new file radar/us_state_ai_check.py, paste the whole script below, and save.

"""US state AI law check for an AI use-case register.

A teaching tool, not legal advice. It reads AI use cases with the states
they touch, flags which of a handful of state and city AI laws need a
lawyer's look, and lists what to document, with start dates.

Commands:
  python radar/us_state_ai_check.py sample             write radar/us_ai_use_cases.csv
  python radar/us_state_ai_check.py check              check every use case
  python radar/us_state_ai_check.py check --as-of 2027-07-01
  python radar/us_state_ai_check.py explain US-02      the reasoning for one use case
  python radar/us_state_ai_check.py notice US-02       draft a Colorado adverse-outcome notice
  python radar/us_state_ai_check.py dates              the timeline of the laws covered
"""
import argparse
import csv
import sys
from datetime import date
from pathlib import Path

HERE = Path(__file__).resolve().parent
CASES = HERE / "us_ai_use_cases.csv"
OUT_CSV = HERE / "us_state_ai_register.csv"
OUT_MD = HERE / "us_state_ai_register.md"

# Laws covered, as read on official or official-hosted pages in November 2026.
# Each entry: short name, where it applies, start date of the duties used here.
LAWS = {
    "CO-ADMT": ("Colorado SB 26-189 (ADMT in consequential decisions)", "CO", date(2027, 1, 1)),
    "CA-ADMT": ("California CCPA regulations on ADMT", "CA", date(2027, 1, 1)),
    "CA-SB947": ("California SB 947 (No Robo Bosses Act)", "CA", date(2027, 7, 1)),
    "IL-HRA": ("Illinois Human Rights Act, AI in employment (HB 3773)", "IL", date(2026, 1, 1)),
    "NYC-LL144": ("New York City Local Law 144 (AEDT)", "NYC", date(2023, 7, 5)),
    "TX-TRAIGA": ("Texas HB 149 (TRAIGA), Bus. & Com. Code ch. 552", "TX", date(2026, 1, 1)),
    "CA-SB53": ("California SB 53 (frontier AI developers)", "CA", date(2026, 1, 1)),
    "NY-RAISE": ("New York RAISE Act (large frontier developers)", "NY", date(2027, 1, 1)),
}

TIMELINE = [
    ("2023-07-05", "NYC Local Law 144: DCWP enforcement of bias audits and notices begins"),
    ("2026-01-01", "Texas TRAIGA in effect; Illinois AI-in-employment amendment in effect"),
    ("2026-01-01", "California CCPA regulations in effect (risk assessments for new processing)"),
    ("2026-05-14", "Colorado SB 26-189 signed; repeals and re-enacts the 2024 Colorado AI Act"),
    ("2027-01-01", "Colorado SB 26-189 applies to consequential decisions from this date"),
    ("2027-01-01", "California ADMT requirements for significant decisions apply"),
    ("2027-01-01", "New York RAISE Act compliance and reporting for large frontier developers"),
    ("2027-07-01", "California SB 947: human corroboration for ADS-based discipline and termination"),
    ("2027-12-31", "California: risk assessments due for processing that started before 2026"),
]

EMPLOYMENT = {"hiring", "promotion_pay", "discipline_termination"}
FIELDS = ["id", "name", "process", "sap_area", "built_by", "affects", "decision",
          "ai_role", "states", "talks_to_people", "trains_frontier_model"]

SAMPLE = [
    ["US-01", "Skills match shown to recruiters", "hire-to-retire", "SuccessFactors Recruiting",
     "sap", "applicants", "hiring", "assists", "IL;CO;CA", "no", "no"],
    ["US-02", "Applicant screening agent that ranks and rejects", "hire-to-retire", "BTP + Joule Studio",
     "us", "applicants", "hiring", "decides", "IL;CO;CA;NYC", "yes", "no"],
    ["US-03", "Attendance agent proposing written warnings", "hire-to-retire", "BTP + Employee Central data",
     "us", "employees", "discipline_termination", "primary", "IL;CA", "no", "no"],
    ["US-04", "Credit limit proposal for distributors", "order-to-cash", "S/4HANA Credit Management + BTP",
     "us", "business_customers", "credit_business", "assists", "CO;CA;TX", "no", "no"],
    ["US-05", "Open-account check for sole-trader contractors", "order-to-cash", "BTP + S/4HANA",
     "us", "individuals", "credit_individual", "decides", "CO;CA;TX", "no", "no"],
    ["US-06", "Supplier invoice field extraction", "procure-to-pay", "Document AI",
     "sap", "none", "none", "none", "IL", "no", "no"],
    ["US-07", "Order-status chatbot for distributors", "order-to-cash", "BTP + Joule Studio",
     "us", "business_contacts", "none", "none", "IL;CO;CA;TX", "yes", "no"],
]

ALLOWED = {
    "built_by": {"sap", "partner", "us"},
    "affects": {"applicants", "employees", "individuals", "business_customers", "business_contacts", "none"},
    "decision": EMPLOYMENT | {"credit_individual", "credit_business", "none"},
    "ai_role": {"decides", "primary", "assists", "none"},
    "talks_to_people": {"yes", "no"},
    "trains_frontier_model": {"yes", "no"},
}
STATES = {"CO", "CA", "IL", "TX", "NY", "NYC", "OTHER"}

# The EU AI Act analogue, so the register lines up with the EU topic's register.
EU_ANALOGUE = {
    "hiring": "EU: Annex III 4(a) recruitment, high-risk from 2027-12-02",
    "promotion_pay": "EU: Annex III 4(b) work terms and promotion, high-risk from 2027-12-02",
    "discipline_termination": "EU: Annex III 4(b) termination and monitoring, high-risk from 2027-12-02",
    "credit_individual": "EU: Annex III 5(b) natural persons, high-risk from 2027-12-02",
    "credit_business": "EU: not in Annex III (5(b) names natural persons)",
    "none": "EU: no high-risk area; check Article 50 if it talks to people",
}


def load(path):
    if not path.exists():
        sys.exit(f"No file at {path}. Run: python radar/us_state_ai_check.py sample")
    with path.open(newline="", encoding="utf-8") as f:
        rows = list(csv.DictReader(f))
    for n, row in enumerate(rows, start=2):
        missing = [k for k in FIELDS if k not in row]
        if missing:
            sys.exit(f"{path.name} is missing columns: {', '.join(missing)}")
        for key, allowed in ALLOWED.items():
            value = row[key].strip().lower()
            if value not in allowed:
                sys.exit(f"Line {n} ({row['id']}): {key}='{row[key]}' is not one of {sorted(allowed)}")
            row[key] = value
        states = {s.strip().upper() for s in row["states"].split(";") if s.strip()}
        unknown = states - STATES
        if unknown:
            sys.exit(f"Line {n} ({row['id']}): unknown state code(s) {sorted(unknown)}; use {sorted(STATES)}")
        row["states"] = states
    return rows


def when(law, as_of):
    start = LAWS[law][2]
    return "in force" if as_of >= start else f"from {start.isoformat()}"


def check(uc, as_of):
    """Return a list of (law, verdict, reasons, duties). Verdicts: APPLIES, CHECK, NOT LIKELY."""
    out = []
    st, dec, role = uc["states"], uc["decision"], uc["ai_role"]
    about_people = uc["affects"] in {"applicants", "employees", "individuals"}
    builder = uc["built_by"] == "us"

    # Colorado: ADMT that materially influences a consequential decision about a Colorado consumer.
    if "CO" in st:
        domain = dec in EMPLOYMENT or dec == "credit_individual"
        if about_people and domain and role in {"decides", "primary"}:
            verdict, why = "APPLIES", "ADMT output drives a consequential decision in a covered domain"
        elif about_people and domain and role == "assists":
            verdict, why = "CHECK", "AI output is shown to the decision maker; is that 'material influence'?"
        else:
            verdict, why = "NOT LIKELY", "no consequential decision about an individual in a covered domain"
        duties = []
        if verdict != "NOT LIKELY":
            w = when("CO-ADMT", as_of)
            duties = [f"Clear and conspicuous notice before ADMT is used in the decision [{w}]",
                      f"Within 30 days of an adverse outcome: plain-language description of the decision and "
                      f"the ADMT's role, how to ask for more information, the person's rights [{w}]",
                      f"On request: correct factually inaccurate personal data; meaningful human review "
                      f"where commercially reasonable [{w}]",
                      f"Keep compliance records at least 3 years after the decision [{w}]"]
            if builder:
                duties.append(f"As developer: technical documentation for the deployer (intended and known "
                              f"harmful uses, training data categories, limits, human review instructions) [{w}]")
        out.append(("CO-ADMT", verdict, [why], duties))

    # California CCPA ADMT rules: technology that replaces or substantially replaces human decisionmaking.
    if "CA" in st:
        significant = dec in EMPLOYMENT or dec == "credit_individual"
        if about_people and significant and role in {"decides", "primary"}:
            verdict, why = "APPLIES", "ADMT substantially replaces the human in a significant decision"
        elif about_people and significant and role == "assists":
            verdict, why = "CHECK", "outside ADMT only if a reviewer understands, weighs and can change the output"
        else:
            verdict, why = "NOT LIKELY", "no significant decision about a California consumer"
        duties = []
        if verdict != "NOT LIKELY":
            w = when("CA-ADMT", as_of)
            duties = [f"Pre-use notice: purpose, personal information used, outputs, human role, rights [{w}]",
                      f"Opt-out, unless an exception applies (for example a human appeal that can "
                      f"overturn the decision) [{w}]",
                      f"Access: purpose, logic, outcome and how the output affected the decision [{w}]",
                      "Documented risk assessment for ADMT in significant decisions "
                      "[new processing since 2026-01-01; older processing by 2027-12-31]"]
        out.append(("CA-ADMT", verdict, [why], duties))

        # California SB 947: discipline and termination.
        if dec == "discipline_termination" and uc["affects"] == "employees":
            if role in {"decides", "primary"}:
                verdict, why = "APPLIES", "an ADS is primarily relied on for discipline or termination"
            else:
                verdict, why = "CHECK", "does the employer 'primarily rely' on the ADS output?"
            w = when("CA-SB947", as_of)
            out.append(("CA-SB947", verdict, [why],
                        [f"A person examines and corroborates the output with other information before "
                         f"acting; uncorroborated output can't be the basis [{w}]",
                         f"Written notice to the worker that an ADS was relied on and a human corroborated "
                         f"it, with a contact for more information [{w}]",
                         f"On request: a meaningful description of the worker data the ADS used [{w}]"]))

    # Illinois: AI that influences or facilitates employment decisions.
    if "IL" in st and dec in EMPLOYMENT and role != "none":
        w = when("IL-HRA", as_of)
        out.append(("IL-HRA", "APPLIES", ["AI influences or facilitates an employment decision"],
                    [f"Notice to applicants and employees that AI is used [{w}; IDHR notice rules were "
                     f"proposed, check their final form]",
                     f"No discriminatory effect on protected classes; test for it and keep the results [{w}]"]))

    # New York City: automated employment decision tools in hiring or promotion.
    if "NYC" in st and dec in {"hiring", "promotion_pay"} and role != "none":
        verdict = "APPLIES" if role in {"decides", "primary"} else "CHECK"
        w = when("NYC-LL144", as_of)
        out.append(("NYC-LL144", verdict, ["tool used to screen candidates for jobs in New York City"],
                    [f"Independent bias audit within one year before use; publish a summary [{w}]",
                     f"Notice to candidates at least 10 business days before use [{w}]"]))

    # Texas: intent-based prohibitions for everyone; disclosure only for agencies and health care.
    if "TX" in st:
        reasons = ["private company, so the consumer disclosure duty (agencies, health care) does not apply",
                   "prohibitions turn on intent; disparate impact alone does not show intent"]
        w = when("TX-TRAIGA", as_of)
        out.append(("TX-TRAIGA", "CHECK" if about_people else "NOT LIKELY", reasons,
                    [f"Never develop or deploy with intent to unlawfully discriminate [{w}]",
                     f"Keep testing and framework records (for example a NIST AI RMF profile): they support "
                     f"the statutory defenses if the Attorney General asks [{w}]"] if about_people else []))

    # Frontier laws bind model developers, not the companies that call their models.
    for law, state in (("CA-SB53", "CA"), ("NY-RAISE", "NY")):
        if state in st or (state == "NY" and "NYC" in st):
            if uc["trains_frontier_model"] == "yes":
                out.append((law, "CHECK", ["you train a model that may be a frontier model"],
                            [f"Frontier framework, transparency reports, incident reporting [{when(law, as_of)}]"]))
            else:
                out.append((law, "NOT LIKELY", ["binds frontier model developers; ask your model vendor "
                                                 "for its published framework"], []))
    return out


def summary_line(results):
    order = {"APPLIES": 0, "CHECK": 1, "NOT LIKELY": 2}
    return ", ".join(f"{law} {v}" for law, v, _, _ in sorted(results, key=lambda r: order[r[1]])
                     if v != "NOT LIKELY") or "none flagged"


def cmd_sample(_args):
    with CASES.open("w", newline="", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerow(FIELDS)
        writer.writerows(SAMPLE)
    print(f"Wrote {len(SAMPLE)} sample use cases to {CASES.relative_to(HERE.parent)}")


def cmd_check(args):
    rows = load(Path(args.file))
    print(f"US state AI law check as of {args.as_of.isoformat()} ({len(rows)} use cases). Not legal advice.\n")
    print(f"{'ID':<6} Flags (APPLIES first, then CHECK)")
    results = []
    for uc in rows:
        res = check(uc, args.as_of)
        results.append((uc, res))
        print(f"{uc['id']:<6} {summary_line(res)}")

    applies = sum(1 for _, res in results for r in res if r[1] == "APPLIES")
    checks = sum(1 for _, res in results for r in res if r[1] == "CHECK")
    print(f"\nTotals: {applies} APPLIES, {checks} CHECK across {len(rows)} use cases")

    with OUT_CSV.open("w", newline="", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerow(["id", "name", "law", "verdict", "duties", "reasons", "eu_analogue"])
        for uc, res in results:
            for law, verdict, reasons, duties in res:
                writer.writerow([uc["id"], uc["name"], law, verdict, " | ".join(duties),
                                 "; ".join(reasons), EU_ANALOGUE[uc["decision"]]])
    lines = [f"# US state AI law check ({args.as_of.isoformat()})", "",
             "Not legal advice. Every APPLIES and CHECK row goes to counsel.", ""]
    for uc, res in results:
        lines += [f"## {uc['id']}: {uc['name']}", "",
                  f"{uc['process']}, {uc['sap_area']}. States: {', '.join(sorted(uc['states']))}. "
                  f"{EU_ANALOGUE[uc['decision']]}.", ""]
        for law, verdict, reasons, duties in res:
            why = "; ".join(reasons)
            lines.append(f"- **{law}: {verdict}.** {why}{'' if why.endswith('?') else '.'}")
            lines += [f"  - {d}" for d in duties]
        lines.append("")
    OUT_MD.write_text("\n".join(lines), encoding="utf-8")
    print(f"Wrote {OUT_CSV.relative_to(HERE.parent)} and {OUT_MD.relative_to(HERE.parent)}")


def find(rows, wanted):
    match = [uc for uc in rows if uc["id"].lower() == wanted.lower()]
    if not match:
        sys.exit(f"No use case {wanted}. IDs: {', '.join(uc['id'] for uc in rows)}")
    return match[0]


def cmd_explain(args):
    uc = find(load(Path(args.file)), args.id)
    print(f"{uc['id']}: {uc['name']} ({uc['process']}, {uc['sap_area']})")
    print(f"States: {', '.join(sorted(uc['states']))}. AI role: {uc['ai_role']}. {EU_ANALOGUE[uc['decision']]}\n")
    for law, verdict, reasons, duties in check(uc, args.as_of):
        print(f"{law} ({LAWS[law][0]}): {verdict}")
        for r in reasons:
            print(f"  why: {r}")
        for d in duties:
            print(f"  - {d}")


def cmd_notice(args):
    uc = find(load(Path(args.file)), args.id)
    flags = {law: v for law, v, _, _ in check(uc, args.as_of)}
    if flags.get("CO-ADMT") not in {"APPLIES", "CHECK"}:
        sys.exit(f"{uc['id']} is not flagged for Colorado, so no Colorado notice is drafted.")
    print(f"DRAFT for counsel review. Colorado post-adverse-outcome disclosure for {uc['id']}.")
    print("Send within 30 days of the decision. Fill every <...> from the decision record.\n")
    print(f"Subject: About our decision on your {'application' if uc['affects'] == 'applicants' else 'request'}\n")
    print("1. The decision: <what was decided, on which date, for which position or account>.")
    print(f"2. The technology's role: we used an automated tool ({uc['name'].lower()}) that "
          f"{'scored and ranked' if uc['ai_role'] in {'decides', 'primary'} else 'produced a result shown to the person deciding on'} "
          "<inputs, for example the skills in your application> against <criteria>. "
          "<Say whether a person reviewed the result before the decision.>")
    print("3. More information: reply to <contact> or call <number> to ask how the tool works "
          "and which of your data it used.")
    print("4. Your rights: you can ask us to correct personal data that is factually wrong, "
          "and ask for a person to review the decision. <Describe how, and by when.>")
    print("\nCheck the wording against the Colorado Attorney General's final rules before first use.")


def cmd_dates(args):
    print(f"Timeline of the laws in this check, against {args.as_of.isoformat()}:\n")
    for day, text in TIMELINE:
        mark = "done " if date.fromisoformat(day) <= args.as_of else "ahead"
        print(f"  [{mark}] {day}  {text}")


def main():
    parser = argparse.ArgumentParser(description="US state AI law check for an AI use-case register.")
    sub = parser.add_subparsers(dest="command", required=True)
    common = argparse.ArgumentParser(add_help=False)
    common.add_argument("--file", default=str(CASES), help="CSV of use cases (default radar/us_ai_use_cases.csv)")
    common.add_argument("--as-of", type=date.fromisoformat, default=date.today(),
                        help="date to check duties against, YYYY-MM-DD (default today)")
    sub.add_parser("sample", help="write the sample use cases")
    sub.add_parser("check", parents=[common], help="check every use case and write the register")
    for name, text in (("explain", "show the reasoning for one use case"),
                       ("notice", "draft a Colorado adverse-outcome notice")):
        p = sub.add_parser(name, parents=[common], help=text)
        p.add_argument("id")
    sub.add_parser("dates", parents=[common], help="print the timeline")
    args = parser.parse_args()
    {"sample": cmd_sample, "check": cmd_check, "explain": cmd_explain,
     "notice": cmd_notice, "dates": cmd_dates}[args.command](args)


if __name__ == "__main__":
    main()

Step 3: Write the sample use cases

  1. Run:

    python radar/us_state_ai_check.py sample
  2. You should see:

    Wrote 7 sample use cases to radar/us_ai_use_cases.csv
  3. Open radar/us_ai_use_cases.csv. Each row is one of Calder's (fictional) AI use cases. The columns are facts, not conclusions: who built it (built_by), who it affects (affects), which decision it shapes (decision), how strongly (ai_role: decides, primary, assists or none), the states it touches, separated by semicolons (NYC for New York City), whether it talks to people, and whether you train a frontier model.

Step 4: Check every use case

  1. Run the check against a fixed date so your output matches this page:

    python radar/us_state_ai_check.py check --as-of 2026-11-11
  2. You should see:

    US state AI law check as of 2026-11-11 (7 use cases). Not legal advice.
    
    ID     Flags (APPLIES first, then CHECK)
    US-01  IL-HRA APPLIES, CO-ADMT CHECK, CA-ADMT CHECK
    US-02  CO-ADMT APPLIES, CA-ADMT APPLIES, IL-HRA APPLIES, NYC-LL144 APPLIES
    US-03  CA-ADMT APPLIES, CA-SB947 APPLIES, IL-HRA APPLIES
    US-04  none flagged
    US-05  CO-ADMT APPLIES, CA-ADMT APPLIES, TX-TRAIGA CHECK
    US-06  none flagged
    US-07  none flagged
    
    Totals: 10 APPLIES, 3 CHECK across 7 use cases
    Wrote radar/us_state_ai_register.csv and radar/us_state_ai_register.md
  3. Read the pattern. The screening agent (US-02) hits four regimes at once. The distributor credit agent (US-04) hits none, but the sole-trader check (US-05) hits two, because a sole trader is a person. The SuccessFactors skills match (US-01) is CHECK in Colorado and California: whether it counts depends on how recruiters use it.

  4. Open radar/us_state_ai_register.md. It has one section per use case with every flag, its reason, the duties with their status, and the EU AI Act analogue.

Leave out --as-of to check against today's date.

Step 5: Read the reasoning for one use case

  1. Ask why the screening agent is flagged:

    python radar/us_state_ai_check.py explain US-02 --as-of 2026-11-11
  2. You should see (first lines):

    US-02: Applicant screening agent that ranks and rejects (hire-to-retire, BTP + Joule Studio)
    States: CA, CO, IL, NYC. AI role: decides. EU: Annex III 4(a) recruitment, high-risk from 2027-12-02
    
    CO-ADMT (Colorado SB 26-189 (ADMT in consequential decisions)): APPLIES
      why: ADMT output drives a consequential decision in a covered domain
      - Clear and conspicuous notice before ADMT is used in the decision [from 2027-01-01]
      - Within 30 days of an adverse outcome: plain-language description of the decision and the ADMT's role, how to ask for more information, the person's rights [from 2027-01-01]
      - On request: correct factually inaccurate personal data; meaningful human review where commercially reasonable [from 2027-01-01]
      - Keep compliance records at least 3 years after the decision [from 2027-01-01]
      - As developer: technical documentation for the deployer (intended and known harmful uses, training data categories, limits, human review instructions) [from 2027-01-01]

    The list continues with California, Illinois and New York City, then two NOT LIKELY lines for the frontier laws with the reason "binds frontier model developers; ask your model vendor for its published framework".

  3. Now run explain US-04. Every law comes out NOT LIKELY, with the reason that there is no decision about an individual. That reason line is what you show when someone asks why the credit agent has no notice.

Step 6: Move the clock forward

  1. Check the attendance agent as it will stand on 1 July 2027:

    python radar/us_state_ai_check.py explain US-03 --as-of 2027-07-01
  2. The SB 947 duties now say [in force]: a person corroborates the output before any warning, and the worker gets written notice. The flags don't change with time; only the status of each duty does.

  3. Print the timeline:

    python radar/us_state_ai_check.py dates --as-of 2026-11-11
    Timeline of the laws in this check, against 2026-11-11:
    
      [done ] 2023-07-05  NYC Local Law 144: DCWP enforcement of bias audits and notices begins
      [done ] 2026-01-01  Texas TRAIGA in effect; Illinois AI-in-employment amendment in effect
      [done ] 2026-01-01  California CCPA regulations in effect (risk assessments for new processing)
      [done ] 2026-05-14  Colorado SB 26-189 signed; repeals and re-enacts the 2024 Colorado AI Act
      [ahead] 2027-01-01  Colorado SB 26-189 applies to consequential decisions from this date
      [ahead] 2027-01-01  California ADMT requirements for significant decisions apply
      [ahead] 2027-01-01  New York RAISE Act compliance and reporting for large frontier developers
      [ahead] 2027-07-01  California SB 947: human corroboration for ADS-based discipline and termination
      [ahead] 2027-12-31  California: risk assessments due for processing that started before 2026

Step 7: Draft the Colorado notice

  1. Draft the adverse-outcome disclosure for a rejected Colorado applicant:

    python radar/us_state_ai_check.py notice US-02 --as-of 2026-11-11
  2. You should see:

    DRAFT for counsel review. Colorado post-adverse-outcome disclosure for US-02.
    Send within 30 days of the decision. Fill every <...> from the decision record.
    
    Subject: About our decision on your application
    
    1. The decision: <what was decided, on which date, for which position or account>.
    2. The technology's role: we used an automated tool (applicant screening agent that ranks and rejects) that scored and ranked <inputs, for example the skills in your application> against <criteria>. <Say whether a person reviewed the result before the decision.>
    3. More information: reply to <contact> or call <number> to ask how the tool works and which of your data it used.
    4. Your rights: you can ask us to correct personal data that is factually wrong, and ask for a person to review the decision. <Describe how, and by when.>
    
    Check the wording against the Colorado Attorney General's final rules before first use.
  3. Every <...> must be filled from the decision record. If your agent doesn't keep the inputs, criteria and reviewer for each decision, you can't fill them, and that is the finding.

  4. Try notice US-06. You get US-06 is not flagged for Colorado, so no Colorado notice is drafted. That is the expected "empty" result.

Step 8: Save your work

git add radar/us_state_ai_check.py radar/us_ai_use_cases.csv
git commit -m "Radar: US state AI law check for the AI register"

The generated register files are outputs. Regenerate them whenever the CSV changes.

What each part of the script does

Part What it does
LAWS, TIMELINE The laws covered and their start dates, as read in November 2026. Change them here when a law or rule moves
FIELDS, ALLOWED, STATES, load Reads the CSV and refuses unknown values or state codes, so a typo can't silently become "not flagged"
check One block per law: Colorado and California test the decision and the AI's role, SB 947 tests discipline, Illinois and NYC test employment, Texas notes intent, the frontier laws point to the vendor
when Adds [in force] or [from date] to each duty, compared with --as-of
EU_ANALOGUE The matching EU AI Act area, so this register lines up with the EU one
cmd_check Prints the summary and writes us_state_ai_register.csv and .md
cmd_explain, cmd_notice, cmd_dates The reasoning for one use case, the Colorado notice draft, the timeline

If something goes wrong

What you see What it means What to do
python: command not found or 'python' is not recognized Python isn't on your path, or .venv isn't active Turn on .venv (Step 1). On macOS/Linux without .venv, use python3.
ModuleNotFoundError Unusual: the script needs only built-in modules Check that you saved the whole file and are running radar/us_state_ai_check.py.
No file at ... us_ai_use_cases.csv The sample file doesn't exist yet Run Step 3 first.
Line 3 (US-02): ai_role='decide' is not one of [...] A value in your CSV isn't in the allowed list Use one of the listed values, in lowercase.
unknown state code(s) ['NJ'] The script only knows the states it covers Use OTHER for other states and check them with counsel.
missing columns A column was renamed or deleted in a spreadsheet Run sample again into a fresh file and copy your rows across.
invalid isoformat string The date isn't in YYYY-MM-DD form Write --as-of 2027-07-01.
Network or proxy errors Not expected: the script makes no network calls If you see one, you are running a different script.

The SAP way

As of November 2026, SAP offers no US state AI law module. What it offers are the controls and records you need to meet these laws, and the obligation to use them sits with you.

  • Feature-level switches. In SuccessFactors Recruiting, AI features are switched on individually in AI Services Administration, need an AI Units license and provisioning options, and a business rule picks the requisition templates that use skills matching (SAP Learning). That is how you keep AI out of requisitions in a state until the notice process is ready.
  • Candidate-facing text. The same course describes an optional data privacy consent statement before resume upload. It is not an AI notice by itself; your AI notice text has to be written and placed deliberately.
  • SAP's own governance. SAP runs an AI Ethics impact assessment for every SAP AI use case and holds an ISO/IEC 42001 certificate (SAP Responsible AI). Ask SAP for the documentation of each AI feature you use for decisions about people: intended use, data, limits, and how human review is meant to work. Colorado's developer duties describe exactly that package.
  • AI Governance Assistant. SAP describes it as EU AI Act support in SAP AI Agent Hub, producing drafts for legal review (SAP News, 30 October 2026). Use its inventory and fact sheets as inputs to your US register; don't expect US classifications from it.
  • Your own agents on BTP. For a Joule Studio or BTP agent, your company is developer and deployer. The notice, the decision record, the human review path and the three-year retention are part of your design, as the next sections show.

Build vs. SAP

Need Use SAP Build yourself Notes
Turn AI on or off by process and requisition Feature switches and business rules in the SAP application Feature flag per state in your agent Do this first: it buys time
AI inventory SAP AI Agent Hub (agents); the register you keep A register CSV or table, like this topic's SAP's inventory covers what runs in SAP
Notices before use Application text where the product allows it Notice text in your career site or portal Wording is legal work, not configuration
Adverse-outcome disclosure (Colorado) Not offered as a feature A notice step fed by the decision record Must go out within 30 days
Human review and correction Workflow and roles in the application A review queue with reviewer, reason and override Record who reviewed, and the result
Bias audit (NYC) Not offered Independent auditor, with your data export Plan the data extract early
Records for three years Application logs, change logs Decision log with version, inputs, output, reviewer Check retention against HR data rules

Architecture view

Where it sits. Calder's (fictional) screening agent runs on BTP. It reads requisitions and applications from SuccessFactors Recruiting, writes a ranking back, and a recruiter decides. A decision log and a notice service sit next to it; the HR system stays the system of record for the application status.

flowchart LR
  SF[SuccessFactors<br/>Recruiting] --> AG[Screening agent<br/>on BTP]
  AG --> SF
  AG --> DL[Decision log<br/>3+ years]
  SF --> NS[Notice service]
  DL --> NS
  NS --> C[Applicant]
  C --> RQ[Review and<br/>correction queue]
  RQ --> SF

Non-functional requirements that decide the design:

Requirement Typical target What breaks if missed
Decision record completeness 100% of decisions with version, inputs, output, reviewer You can't explain an adverse outcome or defend it
Retention At least 3 years after the decision (Colorado) Records gone before a complaint arrives
Notice latency Adverse-outcome disclosure well inside 30 days A statutory deadline missed for every rejected Colorado applicant
Location accuracy State of residence or work known per applicant Wrong notice, or none
Data minimization Only the fields the ranking needs, no protected attributes Discrimination risk; California's ban on inferring protected status for workers

Integration and identity. The agent calls SuccessFactors APIs under a technical user with read access to the requisitions it screens and write access only to the ranking field. The recruiter's own identity records the decision in SuccessFactors. The notice service reads the decision record, not the model, so the explanation matches what happened. Applicant data in the log is pseudonymized where possible and kept in the region your privacy policy states.

The decision to record (ADR-US-01, course example):

  • Context: the screening agent would reject the lowest-ranked applicants automatically, across Colorado, California, Illinois and New York City.
  • Options: (A) automatic rejection with notices in every state; (B) ranking only, a recruiter decides with real authority, notices where the ranking materially influences the decision; (C) no AI for these requisitions.
  • Decision: B, with a feature flag per state and a review queue.
  • Consequences: recruiters spend time on review; California's "human involvement" carve-out becomes arguable but still needs counsel; Colorado notices and records are still needed, because the ranking materially influences the decision; New York City needs a bias audit before use.

What a design review will ask:

  1. Which states does each requisition touch, and how does the agent know?
  2. Can a reviewer override the ranking, and where is that recorded?
  3. What exactly goes into the adverse-outcome disclosure, and from which data?
  4. Which protected attributes or proxies (such as ZIP codes) can reach the ranking?
  5. Who changes the register when Colorado's final rules or a court ruling arrive?

Production concerns

  • Authorizations. The agent's technical user must not read compensation, health or protected-attribute fields. Reviewers need an SAP role that lets them change the status and requires a reason. See agent permissions and SAP authorizations.
  • Evaluation. Test ranking outcomes by group before go-live and after every model or prompt change. Keep the results; under Texas law they support a defense, and in Illinois and New York City they are what an auditor or complainant will ask for.
  • Versioning. Record the model, prompt and rule version with every decision. A three-year record that can't say which version decided is not much of a record.
  • Cost. Notices and reviews cost more than tokens. Count reviewer minutes per decision and the volume of adverse outcomes when you estimate the agent's business case.
  • Operations. Monitor review queue age and notice deadlines like any other SLA. A queue that backs up turns into missed statutory dates.
  • Clean core. Keep the agent side-by-side on BTP and write back through released APIs. Don't add custom fields to store AI explanations in the core if the log can hold them.

Pitfalls

  • Treating "a human approves" as an exemption. Rubber-stamp review doesn't meet California's carve-out and doesn't take you out of Colorado's "materially influence".
  • Not knowing the person's state. Requisitions for remote roles touch every state. Capture residence or work location before the AI runs.
  • Explaining from the model, not the record. An explanation regenerated by the model later may not match the decision. Generate it from the stored inputs and output.
  • Forgetting sole traders. "We only score companies" stops being true the day an individual opens an account.
  • Waiting for federal preemption. It may come, through the courts or Congress, but no one can plan a 2027 go-live around it.
  • Copying the EU work and stopping. The EU register is a good base, but the US triggers (states, adverse outcomes, bias audits) need their own columns.

Exercise

Add your own use case and its Colorado notice.

  1. Open radar/us_ai_use_cases.csv and add a row for a promotion recommendation agent: id US-08, name Promotion shortlist for team leads, process hire-to-retire, area BTP + Employee Central data, built by us, affects employees, decision promotion_pay, AI role primary, states CO;IL, talks to people no, trains frontier model no.
  2. Run python radar/us_state_ai_check.py check --as-of 2027-02-01 and find US-08 in the summary.
  3. Run python radar/us_state_ai_check.py explain US-08 --as-of 2027-02-01 and copy the duties into a short note for your HR lead.
  4. Run python radar/us_state_ai_check.py notice US-08 --as-of 2027-02-01. The subject line says "request"; change it in your note to fit a promotion decision, and fill every <...> with made-up but realistic values.
  5. Write one paragraph under the note: which record fields your agent must store so the notice can be filled from data.

Done when: US-08 shows CO-ADMT APPLIES, IL-HRA APPLIES in the check, and your note lists the duties, a filled notice and the record fields. Keep the note; the governance work in the portfolio projects reuses it.

Going further

For readers who want the primary texts and the open legal questions.

  • Colorado SB 26-189 fiscal note (leg.colorado.gov): the most complete official summary of the duties, exclusions, the cure period and the litigation context. Read it with the Attorney General's rulemaking page and the final rules once adopted; the rules will settle how detailed the notices must be.
  • Texas chapter 552 (Texas statutes): short enough to read in full. Note how the defenses reward testing and recognized frameworks. That makes an internal AI risk framework a legal asset, not only a governance one.
  • The CPPA regulations (CPPA): the human-involvement carve-out and the opt-out exceptions are where most design choices land. Read the final text, not a summary, before deciding a reviewer step takes a tool out of scope.
  • Executive Order 14365 (White House fact sheet): the open question is whether litigation narrows state laws such as Colorado's. The Justice Department's complaint against the 2024 Colorado law argued that its disclosure rules compel speech and its anti-discrimination provisions impose race- and sex-conscious obligations. If a court accepts parts of that, notice duties and anti-discrimination testing may be affected differently, so keep them as separate rows in your register.

Two harder edges are worth tracking. First, "materially influence" (Colorado) and "substantially replace" (California) draw the line at different points; a tool can be in one and out of the other. Second, the same ranking can be an "assist" for the EU's Article 6(3) test and still a material influence in Colorado. Record the reasoning per law, not one verdict per tool.

Check yourself

Pick one answer for each question. The explanation appears after you choose.

1What three facts about a use case decide most of the US state law exposure?

Answer: B. The laws are triggered by decisions about people (hiring, discipline, credit), by how strongly the AI shapes them, and by where the person is. The technology inside matters much less.

2Under Colorado SB 26-189, what must a deployer send after an adverse outcome, and when?

Answer: B. The 30-day post-adverse-outcome disclosure is the core new duty. Impact assessments were part of the repealed 2024 law, and published bias audits are New York City's rule.

3Your screening agent's explanation is generated by the model when an applicant asks, weeks after the decision. What is the risk?

Answer: D. The disclosure must describe the decision and the tool's role. Only a record of the version, inputs, output and reviewer at the time shows that. A later regeneration can drift.

4Why does the script flag the SuccessFactors skills match (US-01) as CHECK in California rather than APPLIES?

Answer: C. California's ADMT definition covers technology that replaces or substantially replaces human decision-making. Real human involvement is carved out, so the answer depends on how recruiters use the match.

5What does Texas TRAIGA mean for Calder, a private manufacturer using a credit agent on sole traders?

Answer: A. The disclosure duty covers governmental agencies and health care providers. The prohibitions turn on intent, and the statute shields findings from testing and substantial compliance with recognized frameworks.

6A manager wants the attendance agent in California to send warnings automatically from July 2027. What would you do?

Answer: A. SB 947 bars relying solely on an automated decision system for discipline and requires a person to examine and corroborate the output, then notify the worker. Data center location doesn't change who the worker is.

7In the ADR, why does option B (ranking only, a recruiter decides) still need Colorado notices and records?

Answer: B. Colorado's law covers ADMT that materially influences a consequential decision, not only fully automated ones. A ranking the recruiter works from still shapes the outcome.

8Which artifact is most useful when Colorado's final rules or a court ruling change the picture?

Answer: A. Rules and rulings change one law at a time. If the reasoning is recorded per law, you can update the affected rows instead of re-assessing every tool.

Sources

Sign in to track your progress

We'll email you a one-time sign-in link. No password needed.

or

Tell us a little about you

Optional, every field. It helps us pitch answers to your questions at the right level and decide which topics to write next. It is never shown publicly, and you can change or clear it anytime from the account menu.

SAP areas you work in