AI policy template for US businesses: NIST and state laws

This AI policy template is the US layer, not a second policy. Take the 15 clauses and four annexes from our free AI policy template, which downloads in Word and PDF with no email, and use this page to change the lines that American law reaches.

No federal statute tells a private business what its AI policy must say. What shapes the US version is a voluntary framework from NIST that Texas now rewards, and a handful of state laws that decide particular clauses.

The short version: Texas makes the NIST generative AI profile part of a defense, Colorado’s rewritten law names an acceptable use policy in its text, Connecticut wants a sentence about training large language models in your privacy notice, and California asks for a risk assessment before sensitive data goes into a new system. Health and financial firms already owe written programs under HIPAA and GLBA. This guide takes each in turn, then maps them onto the template.

AI policy template for US businesses: what this layer changes

The root template is written for any jurisdiction. Its clauses say who is covered, which tools are approved, what never goes in, what may go in once reduced, when a human reviews output, and who owns the policy. Every one of those still applies in Seattle or Dallas. What changes is the evidence a US regulator or court will look for, and a few deadlines.

Template clause	What the US layer adds	Where it comes from
3. Approved tools	An approved provider list and vendor due diligence	NIST AI 600-1, GV-6.1-007 and GV-6.1-009
4 and Annex A. Never goes in	US identifiers: SSN, driver’s license, MRN, Medicare number	HIPAA and state breach laws
9. Human review	Review by someone with authority to change the decision	Colorado SB 26-189, CPPA ADMT rules
11. Use case register	A record that answers the Texas attorney general’s questions	TRAIGA, section 552.103
14. Incidents	State breach clocks, and the FTC notice for financial firms	State law, 16 CFR 314.4(j)
15. Review	January 1, 2027, and NIST’s revision of its framework	SB 26-189, CPPA, NIST

Our mapping of the root template to the sources below, read October 13, 2026.

Two things the US layer does not change. The policy is still a control on people, not on text: it gives you a basis to act, and it does not stop anyone pasting a client file. And the template is still not legal advice.

If a state law applies to you in a way this page does not cover, our map of state privacy and AI laws is the place to check next, and the root template’s own section on what it does not cover still applies.

No federal AI law requires the policy, but several laws read it

There is no federal privacy statute of general application and no federal AI statute that tells an ordinary business to adopt an AI acceptable use policy. That is the honest starting point, and it is why a page promising a “compliant” US template is selling something. What exists instead is a set of laws that look for a policy when something goes wrong.

Texas is the clearest. Under section 552.104 of TRAIGA, a business that gets a notice of violation can avoid a lawsuit if, within 60 days, it cures the violation and sends a written statement that, among other things, it has made any necessary changes to internal policies to reasonably prevent further violations.

60 days

A company with no policy has nothing to change, and writes its first one under deadline.

The sector rules already ask for most of it

Health and financial businesses are further along than they think. The HIPAA Security Rule requires a risk analysis, a security awareness and training program for all workforce, and assurances from any business associate handling electronic PHI. The GLBA Safeguards Rule requires a written information security program run by a Qualified Individual. Both are covered further down, and our HIPAA compliant AI guide works through the vendor side.

For everyone else, the national picture is in our guide to which AI tools comply with US privacy law, which explains why compliance follows the use rather than the brand.

NIST AI RMF 1.0: the voluntary framework a US policy can follow

NIST released the AI Risk Management Framework, NIST AI 100-1, on January 26, 2023. Its own page says it is intended for voluntary use, and the document describes itself as rights preserving, not sector specific and use case agnostic, for organizations of all sizes. It is the closest thing the US has to a common vocabulary for AI governance, and Texas law now points to its generative AI companion by name.

4functions: Govern, Map, Measure, Manage

January 26, 2023AI RMF 1.0 released

July 26, 2024generative AI profile, NIST AI 600-1

The core has four functions. Govern is cross cutting: it is meant to inform the other three. Map is about understanding the context a system is used in, Measure about assessing its risks, and Manage about acting on them. NIST says plainly that the actions “do not constitute a checklist” and need not be done in order.

What the Govern function asks of a small office

For a firm that buys AI rather than builds it, almost everything a policy can do sits under Govern. Three subcategories line up with the template. GOVERN 1.1 asks that legal and regulatory requirements involving AI be understood, managed and documented. GOVERN 1.6 asks for mechanisms to inventory AI systems. GOVERN 2.2 asks that personnel and partners receive AI risk management training consistent with policies.

GOVERN 1.5 adds ongoing monitoring and periodic review, with the frequency of review decided in advance. That is clause 15 of the root template, and it is why the template dates every version. A firm that wants to say it follows the AI RMF can point to those four lines in its own document. Our guide to AI in HR shows the same inventory idea applied to hiring tools.

Still voluntary, and being revised

NIST’s framework page states that AI RMF 1.0 is being revised as part of the White House AI Action Plan. On October 13, 2026 the page still presents version 1.0 as the framework, so a policy should cite the version it follows and its date, and clause 15 should send someone to check NIST when the revision lands. For lawyers, the court filings guide shows how far courts have gone without waiting for any framework.

The NIST Generative AI Profile: twelve risks and the actions a policy can borrow

NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, was released on July 26, 2024. It names twelve risks that are unique to or made worse by generative AI, and then lists suggested actions under the same Govern, Map, Measure and Manage headings. For an office using ChatGPT, Claude or Copilot, four of the twelve do most of the work.

Data Privacy is defined as impacts due to leakage and unauthorized use, disclosure or de-anonymization of biometric, health, location or other personally identifiable information. Confabulation is confidently stated but false content. Information Security covers attacks that use the model and attacks on it. Value Chain and Component Integration is the risk of tools and data from outside vendors that nobody vetted.

The suggested actions that become clauses

The profile’s actions are suggestions, not requirements. Several read like policy clauses already, which makes them easy to cite.

NIST AI 600-1 action	What it suggests	Root template clause
GV-1.4-002	Transparent acceptable use policies addressing illegal use	The policy itself
GV-1.6-001	Add generative AI systems to the AI inventory	11, the register
GV-1.6-002	Define inventory exemptions for AI embedded in other software	2, definitions
GV-6.1-007	Inventory outside entities with access to content; approved provider lists	3, approved tools
GV-6.1-009	Vendor due diligence covering data privacy and security	3 and 12
GV-6.1-010	Update the policy for contractors and consultants	1, scope
MP-4.1-009	Detect PII or sensitive data in generated output	9, human review

NIST AI 600-1, July 2024, tables for GOVERN 1.4, 1.6 and 6.1 and MAP 4.1.

The root template already covers contractors and embedded features, because its clause 1 reaches anyone working for the organization on any device, and clause 2 counts features inside products you already use. What the profile adds is the vocabulary. If a client’s security questionnaire asks whether you follow NIST, the honest answer is “we follow these actions from AI 600-1”, with the list attached. The cyber insurance questionnaire guide shows the questions where that list helps.

What the profile says about personal data in prompts

The profile’s privacy section notes that generative models can leak or infer personal information, and one of its Measure actions, MS-2.2-002, written about content provenance data, suggests removing personally identifiable information to prevent harm or misuse. It does not say how. Clause 5 of the template does: cut input to what the task needs, and remove or replace direct identifiers first. What to remove before AI lists those identifiers for US paperwork.

Texas TRAIGA: where following NIST becomes a defense

The Texas Responsible Artificial Intelligence Governance Act, HB 149, took effect on January 1, 2026. It is the first statute on this page that turns the NIST profile into something with legal weight. Section 552.105(e) says a defendant may not be found liable if, among other routes, it discovers a violation through an internal review process while it substantially complies with the most recent version of the NIST generative AI profile, or another nationally or internationally recognized AI risk framework.

552.105(e)

The defense needs both halves: the framework and an internal review that caught the problem. A policy is where you write down that the review exists, who runs it, and how often. That is clause 15 of the template with one addition, a line saying which framework the business follows.

The register that answers the attorney general

Section 552.103 lets the attorney general, after a complaint, issue a civil investigative demand asking for a description of the AI system’s purpose and use, the categories of data processed as inputs, its outputs, metrics, known limitations, and the post deployment monitoring and safeguards the business uses. For a firm using a general chatbot, the inputs question is the one that bites.

Clause 11 of the template keeps a use case register with the tool, the purpose and the categories of input. In Texas, that register is the first draft of your answer. Our Texas privacy and AI law guide covers the rest of TRAIGA, including the health care disclosure in section 552.051(f), which a medical practice should copy into clause 10.

TRAIGA gives the attorney general exclusive enforcement authority, apart from licensing agencies under section 552.106, and section 552.101(b) rules out private lawsuits; how other states enforce is in the national guide.

Colorado SB 26-189: the acceptable use policy the law names

Colorado’s 2024 AI Act was repealed and reenacted by SB 26-189, signed on May 14, 2026. The new text takes effect on January 1, 2027 and applies to consequential decisions made from that date. It regulates what the statute calls automated decision-making technology, or ADMT, when it materially influences decisions about education, employment, housing, financial or lending services, insurance, health care, and government services. The definition of consumer includes employees and job applicants.

January 1, 2027

The new law does not mention NIST at all. What it does mention is an acceptable use policy, and that is the line a US template should be built around.

A clause that keeps chatbot output out of consequential decisions

Section 6-1-1701(2)(b) takes three things out of ADMT. One is a tool used by an individual solely to summarize, organize, translate, draft, route or present information for human review of administrative processing. Another is technology that talks to consumers in natural language, provided it is not contracted, marketed, configured or intended for consequential decisions, and it is subject to an acceptable use policy that prohibits generated content from being used in one.

The statute does not say whose acceptable use policy it means, and the vendor’s may not be written in those terms. A clause in your own policy closes part of that gap. Something like this, added to clause 9:

AI tools may be used to draft, summarize, translate, organize or route information. Output from an AI tool is not used to score, rank, recommend or decide on an applicant, employee, tenant, borrower, policyholder or patient in any decision about employment, housing, credit, insurance, health care, education or government services, unless the use has been approved under clause 3 and recorded in the register under clause 11.

That does not make a covered use legal by itself. It records that the business drew the line, and it gives staff a rule they can follow in the window. Our Colorado AI Act guide goes through where the line usually falls in an office.

If the business does use ADMT for decisions

A deployer that does use covered ADMT owes five things from January 1, 2027: a clear and conspicuous notice at the point of interaction, a plain language explanation within 30 days of an adverse outcome, a way to correct inaccurate personal data, meaningful human review, and records kept for three years. The attorney general enforces, with a 60 day cure period, where a cure is possible, until January 1, 2030, and the act creates no new private right of action.

Two things are open as of October 13, 2026. The attorney general must adopt rules by January 1, 2027, and they are still a proposal.

And a federal court order of April 27, 2026, in xAI’s lawsuit against the 2024 law, pauses enforcement of that law, or of any legislation replacing or amending it enacted that session, for violations up to 14 days after the court rules on xAI’s motion for a preliminary injunction. Neither changes what to write in the policy. The HR guide covers the hiring side.

Connecticut PA 25-113: the privacy notice line about training models

Connecticut amended its Data Privacy Act with Public Act 25-113, mostly from July 1, 2026, and again with Public Act 26-64 from October 1. Three changes from the first round reach an AI policy. The law now applies to a business that controlled or processed the personal data of at least 35,000 consumers in the prior year, or that processes sensitive data at all, or that offers personal data for sale. The main threshold used to be 100,000.

35,000consumers, down from 100,000

July 1, 2026amendments in force

August 1, 2026profiling impact assessments apply

The privacy notice must now include a statement disclosing whether the controller collects, uses or sells personal data for the purpose of training large language models. That is a sentence your privacy team writes, but the policy decides what it can truthfully say.

The act does not spell out how a vendor’s own training on your input fits that statement. A policy whose clause 12 keeps work input on business plans where the vendor does not train on it makes the sentence easy to write; staff on consumer accounts make it guesswork.

Profiling impact assessments, and the right to question a result

A controller that uses profiling to make a decision with a legal or similarly significant effect must now carry out an impact assessment, for processing activities created on or after August 1, 2026. It covers the purpose and context, the risks and mitigations, the categories of input and output, performance metrics and known limitations, and the safeguards in use. Processing sensitive data still needs a data protection assessment.

Consumers subject to such a decision can, if feasible, question the result, be told the reason and review the personal data used; where the decision concerned housing, they can also correct the data and have it reevaluated. A policy that keeps chatbots out of those decisions, as in the Colorado clause above, also keeps this machinery out of most offices. Our Connecticut guide has the rest, and you can watermark a Connecticut driver’s license before a copy leaves the office.

California: risk assessments now, ADMT rules from 2027

The California Privacy Protection Agency’s regulations on risk assessments, cybersecurity audits and automated decisionmaking technology took effect on January 1, 2026. Two parts belong in a US policy. Risk assessments are required from January 1, 2026, with an attestation and summary due to the agency by April 1, 2028. And businesses that use ADMT for significant decisions must comply with the ADMT rules from January 1, 2027.

Sensitive data in a new system needs an assessment first

Section 7150(b) of the regulations lists activities that present significant risk, and a business must assess them before it starts. Processing sensitive personal information is one.

January 1, 2026risk assessments required

January 1, 2027ADMT rules for significant decisions

April 1, 2028attestation and summary due to the agency

There is a narrow exception for employee data used only for payroll, work authorization, benefits, accommodation or wage reporting; any other processing of sensitive personal information is in. Because California employees are consumers under the CCPA, a manager pasting an employee’s diagnosis into a chatbot to help draft a performance review is outside the exception.

That is why clause 4 of the template, the list of data that never goes in, matters more in California than anywhere else on this page. A policy that keeps sensitive personal information out of AI tools avoids a risk assessment it would otherwise owe. Our California guide lists the categories.

What the ADMT rules mean by human involvement

The regulations define ADMT as technology that processes personal information and uses computation to replace or substantially replace human decisionmaking. Substantially replace means using the output without human involvement, and human involvement requires a reviewer who knows how to interpret the output, reviews and analyzes it with other relevant information, and has the authority to make or change the decision.

Those three conditions are a better human review clause than most templates carry. Clause 9 of the root template says output is a draft until a competent person checks it; in California, add that the reviewer must have the authority to change the outcome. Significant decisions cover financial or lending services, housing, education, employment or contracting, and health care.

Employment: the Civil Rights Council rules

Separately, the California Civil Rights Council’s regulations on automated decision systems in employment took effect on October 1, 2025.

The council says they make clear that using such a system may violate California law if it harms applicants or employees based on protected characteristics, and that employers must keep employment records, including automated decision data, for at least four years. Clause 11 should keep hiring records for four years in California, longer than the three Colorado asks. The PHI and PII guide explains which HR records are sensitive in the first place.

Hiring and HR: where the state rules cluster

Of everything a US business does with AI, hiring draws the most state and local law. The rules differ in what they ask, but each of them reaches an employer that uses software to sort people for jobs.

Where	What reaches AI in hiring	From
Colorado	Applicants and employees count as consumers under SB 26-189	January 1, 2027
California	Civil Rights Council rules on automated decision systems, records for four years	October 1, 2025
California	CPPA rules on ADMT for employment decisions	January 1, 2027
Connecticut	Employment opportunities are a significant effect; profiling impact assessments	August 1, 2026
Connecticut	CART Act, PA 26-15: using the tool is no defense to a discrimination claim	October 1, 2026
New York City	Local Law 144: a bias audit within the past year, a public summary and notices to candidates	Enforced since July 5, 2023
Illinois	Human Rights Act: notice of AI use, no discriminatory effect, no ZIP code proxies	January 1, 2026

The sources listed at the end of this page; New York City from the Department of Consumer and Worker Protection.

Connecticut’s CART Act adds a notice duty later. For automated employment decision technology deployed on or after October 1, 2027, the employer must tell employees and applicants when they are interacting with it and, where its output is a substantial factor in an employment decision, give them written notice before that decision. Clause 8 of the template, on notices, is where that goes for Connecticut staff.

The practical line for a policy is the same in all of them: who may use an AI tool on candidate or employee data, for what, and with what human in the loop. Our guide to AI for HR and employee data puts the rules side by side, and the Illinois AI hiring law guide covers that state.

HIPAA and GLBA: the clauses a regulated business already owes

If a business is a HIPAA covered entity or business associate, the Security Rule at 45 CFR 164.308 already requires three things an AI policy should point to. An accurate and thorough risk analysis of risks to electronic PHI. A security awareness and training program for all members of the workforce, including management. And satisfactory assurances, in a business associate agreement, before a vendor creates, receives, maintains or transmits electronic PHI on its behalf.

For an AI policy, that means clause 3 approves no tool for PHI without a signed business associate agreement on the plan in use, and the risk analysis names the AI tools in scope. Which vendors sign one, and on which plans, is in our HIPAA compliant AI guide, and our page for healthcare practices shows the workflow.

The Safeguards Rule for financial firms

Financial institutions under the FTC’s Safeguards Rule have their own list in 16 CFR 314.4. The FTC’s guidance says coverage turns on activities, not labels, and its examples include tax preparation firms, mortgage brokers, collection agencies and investment advisors that are not required to register with the SEC. They designate a Qualified Individual, train staff in security awareness, and oversee service providers by selecting capable ones, requiring safeguards by contract and assessing them periodically.

Section 314.2 defines a service provider as anyone that receives, processes or is permitted access to customer information by providing services directly to the institution, which describes an AI vendor fed client files.

Safeguards Rule duty	16 CFR 314.4	In the AI policy
Qualified Individual	(a)	Owner named in clause 15
Security awareness training	(e)(1)	Annex C acknowledgment
Oversee service providers	(f)	Clause 3 approval and contract check
Report to the board at least yearly	(i)	Clause 15 review feeds the report
Notify the FTC within 30 days, 500 or more consumers	(j)	Clause 14 incidents

16 CFR 314.4 and 314.6, read on eCFR October 13, 2026. Institutions with fewer than 5,000 consumers are exempt from (b)(1), (d)(2), (h) and (i), not from (j).

A firm that prepares tax returns should assume it is in. Our page for CPA firms covers what accounting staff typically paste into an AI tool.

Putting the US layer into the template, clause by clause

Here is the order to do it in, starting from the root template’s numbering. Each step changes a clause you already have.

1. Clause 1, scope. Name contractors and consultants explicitly, as NIST’s GV-6.1-010 suggests, and say the policy covers AI features inside other software.
2. Clause 3, approved tools. Keep an approved provider list with the plan and account type. For PHI, require a business associate agreement; for customer financial data, a Safeguards Rule contract.
3. Clause 4 and Annex A. Add the US column below. In California, keep sensitive personal information out unless a risk assessment has been done.
4. Clause 9, human review. Bar chatbot output from scoring or ranking people in consequential decisions, and require a reviewer with authority to change the outcome.
5. Clause 11, register. Record purpose, input categories, outputs and safeguards for each use case, in the shape of a Texas civil investigative demand.
6. Clause 15, review. Name the framework you follow, with its version, and diary January 1, 2027 for Colorado and California’s ADMT rules.

Annex A: the US column

Add these rows to the root template’s Annex A. They are the identifiers that turn up most in American office paperwork, and most of them appear on HIPAA’s de-identification list or in state breach laws.

Category	United States
Government identity number	Social Security number, ITIN, passport number
State identifiers	Driver’s license or state ID number
Health identifiers	Medical record number, Medicare Beneficiary Identifier, health plan ID
Financial	Bank account and routing numbers, card numbers
Tax	EIN of a sole proprietor, tax return data
Health information	Diagnoses, treatment, claims, leave notes

For the documents themselves, our guide to redacting a Social Security number before AI and the free US document tools handle the common cases.

Nonimo alongside a US AI policy

A policy says what may go into an AI tool. Nonimo is the app that does the reducing clause 5 asks for, on the computer, before the text reaches ChatGPT, Claude or Copilot. Whoever is drafting highlights the passage and presses one key, and the names, Social Security numbers and other identifiers it recognizes become placeholders on that machine. When the answer comes back, the real details go back in on the same machine.

In the terms of this page, that means less personal data in the vendor’s systems, a shorter answer to the Texas question about input data, and less sensitive information for a California risk assessment to describe. It does not replace the vendor contract, the business associate agreement or the human review.

The conversation is saved on the computer, encrypted, so the names can be put back, and what leaves the machine is a daily count with no text, as set out on the security page. Start from Nonimo for US teams, or see the page for law firms.

Sources

Read October 13, 2026.

• NIST, AI Risk Management Framework: released January 26, 2023; intended for voluntary use; AI 600-1 released July 26, 2024; “The AI RMF 1.0 is being revised as part of the White House AI Action Plan.”
• NIST, AI 100-1, AI RMF 1.0: voluntary, rights preserving, non sector specific; four functions; actions “do not constitute a checklist”; GOVERN 1.1, 1.5, 1.6 and 2.2.
• NIST, AI 600-1, Generative Artificial Intelligence Profile: the twelve risks and the definitions of Data Privacy and Confabulation; actions GV-1.4-002, GV-1.6-001, GV-1.6-002, GV-6.1-007, GV-6.1-009, GV-6.1-010, MP-4.1-009 and MS-2.2-002.
• Texas Legislature Online, HB 149, 89th Legislature, enrolled: effective January 1, 2026; sections 552.051(f), 552.101, 552.103(b), 552.104 (60 days, changes to internal policies) and 552.105(e).
• Colorado General Assembly, SB 26-189 and its signed act of May 14, 2026: effective January 1, 2027 for decisions on or after that date; section 6-1-1701(2)(b) exclusions, including the acceptable use policy condition; consumer includes employees and applicants; deployer duties; 60 day cure until January 1, 2030; no new private right of action; rules due by January 1, 2027. The act does not mention NIST.
• Connecticut General Assembly, Public Act 25-113: sections 6 (35,000 consumers, sensitive data, sale), 8 (profiling rights), 9 (privacy notice statement on training large language models) and 11 (impact assessments from August 1, 2026); effective July 1, 2026. Public Act 26-64: further amendments, effective October 1, 2026. Public Act 26-15, the CART Act: sections 7 to 10 (interaction disclosure and written notice for automated employment decision technology deployed on or after October 1, 2027) and 13 and 14 (use of the technology is no defense to a discriminatory practice complaint, from October 1, 2026).
• California Privacy Protection Agency, CCPA updates, risk assessments and ADMT, the announcement of September 23, 2025 and the approved text: effective January 1, 2026; risk assessments from January 1, 2026, submissions by April 1, 2028; ADMT from January 1, 2027; definitions of ADMT and significant decision; section 7150(b).
• California Civil Rights Department, Civil Rights Council regulations on automated decision systems: approved June 27, 2025, in effect October 1, 2025; records for at least four years.
• NYC Department of Consumer and Worker Protection, Automated Employment Decision Tools: Local Law 144 of 2021, bias audit within one year, public summary, notices; enforcement from July 5, 2023.
• eCFR, 45 CFR 164.308: risk analysis, workforce security training, business associate assurances.
• FTC, Safeguards Rule: what your business needs to know: coverage depends on activities; examples include tax preparation firms, mortgage brokers, collection agencies and investment advisors not required to register with the SEC.
• eCFR, 16 CFR 314.2: definition of service provider.
• eCFR, 16 CFR 314.4 and 314.6: Qualified Individual, training, service providers, annual report, FTC notice within 30 days for 500 or more consumers, and the exception below 5,000 consumers.