State AI law after the 99-to-1 vote: build for notice, explanation and human review

The week

On July 1, 2025, the Senate voted 99 to 1 to strip a single sentence from a giant budget bill. For several months that sentence had threatened to reshape how American organizations govern artificial intelligence. In the early hours of the morning, during an overnight series of amendment votes, it was gone.

The amendment that removed it came from Senator Marsha Blackburn of Tennessee. Its stated purpose was to strike the section relating to support for artificial intelligence. The Senate roll call recorded the vote at just after 4 a.m. Senator Thom Tillis of North Carolina cast the only vote against.

The sentence was a moratorium: a pause, written into federal law, on states enforcing their own AI regulations. The House-passed version of the budget reconciliation bill, approved by a single vote, 215 to 214, on May 22, would have barred states from enforcing AI regulations for ten years. Its reach went beyond what most people call AI. It covered state laws regulating AI models, AI systems and automated decision systems, a phrase broad enough to catch older scoring models and rules engines. Budget reconciliation is a special fast-track process Congress uses for tax and spending legislation; it needs only a simple majority in the Senate, which is why large and varied provisions tend to ride along with it.

In the Senate, drafters had tied a version of the provision to federal broadband funding. In effect, states that wanted certain broadband money would have had to hold off on enforcing their AI laws. A late compromise to shorten the pause to five years, with carve-outs, collapsed. Opposition had come from a broad coalition: a bipartisan group of state attorneys general, governors and legislators.

Their objections were practical as well as constitutional. States had spent several years writing rules on deepfakes, children’s online safety, automated hiring tools and insurers’ use of algorithms. A blanket pause threatened to suspend those rules along with the broad AI statutes the moratorium’s backers had in mind. Supporters argued the opposite risk: that dozens of different state regimes would burden developers and slow adoption. Both arguments survived the vote, and both shaped what came next.

The margin of that vote left little doubt about where the Senate stood. It passed the bill without the moratorium later that day, by 51 to 50, with Vice President JD Vance breaking the tie. The House agreed to the Senate version on July 3, by 218 to 214, and the President signed it on July 4.

By then the state laws at stake were not hypothetical. Colorado had enacted its AI Act in May 2024, with its main duties due to start on February 1, 2026. Utah had passed disclosure rules for generative AI the same year. Texas had enacted its Responsible AI Governance Act in June 2025, days before the vote. New York City had been enforcing its rules on automated employment decision tools since July 2023. A ten-year pause would have frozen enforcement of all of them, along with the bills still moving through legislatures.

What did that mean in practice? Nothing new was required of anyone. But nothing was taken away either. States remained free to enforce existing AI laws and pass new ones. For compliance teams at banks, health systems and government agencies, many of whom had been hoping a single federal rule would replace a patchwork of state rules, the vote settled the near-term question. The patchwork was here to stay, and it was going to grow.

The vote touched each sector differently. Lenders and insurers still answered to federal fair lending law and to state insurance regulators, whatever happened to the moratorium. Health systems had a new federal rule on clinical decision support tools that had taken effect on May 1. Federal agencies followed OMB guidance on AI issued in April. State and local agencies answered to their own legislatures, which kept writing rules for government use of AI.

For a chief executive or general counsel, the practical question shifted from “Will we have to deal with state AI law?” to “How do we build a program that can keep up with it?”

For much of 2025, many compliance teams hoped federal preemption would spare them a growing body of state AI law. Preemption is the legal principle that a federal law can override state laws on the same subject; a ten-year moratorium would have worked much like it. The Senate vote ended that hope for the near term. It created no new obligations. It left existing ones in place and signaled more would follow.

The practical question for banks, health systems and agencies was how to build an AI compliance program on moving ground. Our answer starts by sorting state AI laws into three types: laws on consequential decisions, laws on specific uses or sectors, and laws on transparency and disclosure. Within each type, some obligations recur in almost every version and others change from bill to bill. The program should build the recurring core first.

A builder on uncertain ground works the same way: without knowing exactly what the upper floors will look like, you pour a foundation that will carry any reasonable design.

The sections that follow first pin down the software these laws actually cover, because scope decides everything else. They then take each of the three types of law in turn. The last three sections turn to operations. They cover how the program runs day to day, how the pieces fit into one backbone, and a short set of tests for the stable core.

The software these laws actually cover

State AI laws are not mainly about futuristic robots or chatbots that write poetry. Most of them are about something more ordinary and more consequential: software that helps decide what happens to people.

In practice, that means systems such as a credit model that scores loan applicants, a tool that ranks job candidates, a program that flags insurance claims for review, or a system that predicts which patients need follow-up care. Some of these use modern machine learning, in which software learns patterns from large amounts of past data rather than following rules a person wrote by hand. Others are older statistical models. Many laws define AI broadly enough to catch both.

Generative AI, the kind that produces text, images or audio on request, shows up in state law too, but mostly where it faces the public: a chatbot answering customer questions, a letter drafted by software, or synthetic video. That distinction between AI that decides and AI that talks maps neatly onto the three types of law below.

Definitions decide scope

Each statute draws its own boundary, and the boundary is usually set by the decision, not the technology. Colorado’s law turns on whether a system makes, or is a substantial factor in making, a consequential decision. It also excludes a list of ordinary tools that do not make such decisions. Other laws use phrases such as “automated decision tool” or “automated employment decision tool,” each defined differently. A compliance team that starts from a list of “AI projects” will miss the scoring model built a decade ago, and will over-count the chatbot pilot that decides nothing.

The same vendor product can therefore fall in scope in one use and out of it in another. A document-reading tool that summarizes loan files for an underwriter, who then decides, may or may not be a substantial factor depending on how heavily the underwriter relies on it. The same tool used to route mail is plainly out of scope. Scope belongs to each use, recorded in the inventory, rather than to each product.

How scope differs by type of organization

The mix of in-scope uses differs sharply by sector. A credit union’s list is dominated by underwriting, pricing, fraud holds and collections. A broker-dealer or wealth manager will find fewer consequential decisions about consumers, but may have hiring tools and suitability support. A health system has patient-facing triage, scheduling and outreach models, plus clinical decision support. A state benefits agency has eligibility screening, fraud detection and caseworker prioritization. Each starts the same exercise from a different place, and each should expect the hiring and human resources tools to appear on its list whatever its sector.

Laws on consequential decisions

The first type regulates AI used to make, or substantially influence, important decisions about people: lending, employment, housing, insurance, education and health care. These are the decisions that change the course of someone’s life, which is why legislators care most about them.

Colorado’s AI Act, enacted in 2024, was the leading example in mid-2025. It placed duties on developers and deployers of high-risk systems. A developer builds the AI system; a deployer is the organization that uses it to make decisions, such as a bank using a vendor’s credit model. The duties included risk management programs, impact assessments, consumer notices, a route to appeal adverse decisions and a duty of reasonable care to avoid algorithmic discrimination. Algorithmic discrimination is when an automated system produces unfair results for people based on characteristics such as race, sex or age, often because the historical data it learned from carried those patterns.

Two layers: one stable, one shifting

Those duties fall into two layers. The first is procedural and faces the consumer: tell people when AI is involved in a decision about them, explain adverse outcomes, let them correct their data and offer human review.

That layer echoes long-standing rules. Lenders already give adverse action notices under the Equal Credit Opportunity Act and the Fair Credit Reporting Act. An adverse action notice is the letter a lender must send when it turns someone down, explaining the main reasons. Health systems already run appeals and clinical review. The most important consumer-facing AI duties are close cousins of things regulated organizations have done for decades. That is good news: much of the machinery already exists.

The second layer is programmatic: risk management frameworks, annual impact assessments and attorney general notification. An impact assessment is a structured written review of what a system does, who it could affect and how risks are managed. This layer is where legislatures and industry argue, and where drafts change most.

What the Colorado text asked of deployers

The Colorado statute is worth reading closely because so many later bills borrowed from it. Deployers of high-risk systems owe a duty of reasonable care to protect consumers from known or reasonably foreseeable risks of algorithmic discrimination. They earn a rebuttable presumption that they met that duty by doing a list of things. They adopt a risk management policy and program and complete impact assessments. They review each deployed system annually. They notify consumers when a system makes or substantially factors into a consequential decision about them.

The list continues. Deployers let consumers correct inaccurate personal data and appeal adverse decisions, with human review where technically feasible. They publish a statement summarizing the high-risk systems they use. They tell the attorney general within 90 days if they discover algorithmic discrimination. Developers carry parallel duties to document their systems for deployers and to disclose known risks. The attorney general has exclusive enforcement authority, and a violation counts as a deceptive trade practice under Colorado’s consumer protection law. An affirmative defense is available to organizations that follow a recognized AI risk management framework and find and fix violations.

Read that way, the statute splits cleanly into the two layers. Notice, correction, appeal and human review are the procedural core. The risk program, impact assessments, annual reviews, public statements and attorney general reporting are the programmatic layer. The presumption and the defense reward the second layer, which explains why vendors and consultants pushed it hardest. The first layer is what a complaining consumer actually experiences.

The carve-outs for federally supervised firms

Colorado also wrote in carve-outs that matter for regulated industries. Banks, credit unions and their affiliates are treated as compliant if a state or federal prudential regulator examines them under guidance or regulations on AI that meet the statute’s conditions. Insurers are treated as compliant if they follow Colorado’s insurance rules on external data, algorithms and models.

These carve-outs are narrower than they look. Whether existing supervisory guidance qualifies is a legal question, and the answer may differ for a national bank, a state-chartered credit union and a nonbank lender. A fintech without a prudential regulator gets no carve-out at all. A health system has no equivalent sector exemption in the Colorado text. Organizations should check with counsel before relying on any carve-out. Even where one applies, the procedural core is still expected by the federal regulator, so the carve-out saves paperwork rather than process.

How to build the stable layer

A bank or health system can build the procedural layer once and reuse it across states. That means four things:

• A register of decision points where AI materially influences an outcome. Not a list of every algorithm, but a list of the moments where a person gets a yes or a no and software played a meaningful part.
• Standard notice and explanation templates, so that every adverse decision comes with a clear, consistent account of what happened and why.
• A human review queue with service levels, meaning a defined process for a person to look again at a decision, with a target time for doing it.
• Records retained long enough to answer a complaint, so the organization can show what the system did, what the person was told and what the reviewer decided.

Programmatic duties can then be added state by state as they settle. Organizations should check with counsel on exemptions for entities already supervised under federal law, because they vary by statute.

One credit line, start to finish

Take a regional bank that uses a vendor’s model to help approve small-business credit lines. Under the procedural approach, the bank’s compliance team first lists that credit decision on its register. It then checks that its existing adverse action letters explain the main reasons in language applicants can follow, adding a sentence that automated analysis contributed. It routes any applicant who asks for reconsideration to a named team of credit officers, who aim to respond within a set number of business days. It keeps the model’s inputs, the letter and the reviewer’s notes together. Nothing here waits on any particular state’s final rules; all of it would satisfy the consumer-facing core of most of them.

Where the existing machinery falls short

The strongest objection to this approach is that adverse action processes were built for simpler models. Regulation B, which implements the Equal Credit Opportunity Act, requires a statement of specific reasons for denial. With a scorecard of a dozen variables, the reasons are easy to find. With a machine learning model using hundreds of inputs, producing accurate principal reasons takes real technical work. The reasons generated must reflect what actually drove the decision, not a convenient list of standard phrases.

Human review has its own weakness. A reviewer who sees the model’s answer first, has little time and lacks authority to change it will tend to confirm it. That pattern is widely described as automation bias. A review queue that rubber-stamps decisions satisfies no statute in spirit and will not survive an examiner’s sampling. Reviewers need the underlying information, enough time, a clear standard and the power to reverse. Their reversals should be recorded and fed back to the model owner.

Neither objection defeats the argument. Both show where the existing machinery needs upgrading for AI. The upgrade is still cheaper than building a parallel AI process, and it improves the treatment of customers whether or not any AI statute ever takes effect.

Health systems already had a federal version

Health systems had a federal analog in place before the moratorium debate. The 2024 rule implementing Section 1557 of the Affordable Care Act added a provision on patient care decision support tools. Covered entities must make reasonable efforts to identify uses of such tools that rely on race, color, national origin, sex, age or disability. They must then make reasonable efforts to mitigate the risk of discrimination. That provision took effect on May 1, 2025. For a hospital, the state laws on consequential decisions extend an existing obligation rather than creating a new one, and the inventory built for Section 1557 is the natural starting point.

Measuring it

The procedural layer also produces evidence that regulators of every kind understand. Useful measures include the share of adverse decisions with a completed explanation, the time taken to complete a human review, and the rate at which reviews overturn the original outcome.

That last number carries more weight than it seems. A high overturn rate is an early sign that a model needs attention, well before any statute requires a formal assessment. If people keep reversing the machine’s calls, something about the machine is off. Model risk and fair lending teams can feed those measures into existing reporting instead of building a separate AI dashboard. Model risk management is the discipline, long established at banks, of validating and monitoring the models a firm relies on; it is a natural home for this work.

Banks know this discipline from the Federal Reserve’s SR 11-7 and the OCC’s matching bulletin on model risk management, both issued in 2011. Those documents already expect inventories, validation, ongoing monitoring and documentation for models that drive business decisions. A bank that classifies AI-influenced credit decisions as models under that framework gets most of the programmatic layer for free. A health system or agency has no such inherited discipline, and must decide which team plays the model risk role.

Laws on specific uses or sectors

The second type targets particular applications rather than decisions in general. Examples include New York City’s audit and notice rules for automated employment decision tools. Others are Colorado’s 2021 law on insurers’ use of external consumer data and algorithms. Illinois has rules on AI analysis of video job interviews, and California has a requirement that health care providers disclose generative AI use in certain patient communications. Many state insurance regulators have also adopted guidance based on the NAIC’s 2023 model bulletin on insurers’ use of AI. The NAIC, the National Association of Insurance Commissioners, is the body through which state insurance regulators coordinate; its model bulletins are templates that individual states can adopt.

These laws are narrow but concrete. They require specific notices, bias audits or human review at defined points. A bias audit is an independent check of whether a tool’s results differ unfairly across groups of people.

What the narrow laws demand, in detail

New York City’s rule shows how concrete these laws get. An employer or employment agency using an automated employment decision tool for candidates in the city needs a bias audit by an independent auditor within the year before use. It must publish a summary of the results and give candidates notice at least ten business days before the tool is used. Enforcement began in July 2023. A national employer with a single New York City office still has to meet all of it for those candidates.

Illinois has two layers in employment. Its Artificial Intelligence Video Interview Act, in force since 2020, requires notice, an explanation of how the AI works and consent before AI analyzes recorded interviews. A 2024 amendment to the Illinois Human Rights Act, effective January 1, 2026, bars employers from using AI in ways that discriminate and requires notice when they use it for employment decisions. California’s AB 3030, effective in January 2025, requires health care providers that use generative AI for patient clinical communications to include a disclaimer and instructions for reaching a human. Communications reviewed by a licensed provider before sending are exempt.

Each law fits a different team. The employment rules land on human resources and talent acquisition, often through a recruiting platform the organization does not control. The California rule lands on the patient portal and the clinicians who use drafting tools. The insurance rules land on actuarial and underwriting. A central AI committee cannot carry out these controls itself. It can only make sure each owner knows the rule exists and has evidence that the control runs.

The inventory is the control that scales

The control that scales is a use case inventory tagged by function, sector and jurisdiction. In effect, it is a well-labeled catalog of every place the organization uses AI: what it does (screening job applicants, drafting patient messages), which business it sits in, and which states or cities its decisions touch.

When a new law targets AI in hiring or in patient messaging, the compliance team can query the inventory and find every affected use within a day. Without it, each new law starts with a survey of business units, which takes weeks and misses shadow deployments bought on a corporate card. “Shadow” deployments are tools that a team adopted on its own, often a low-cost online service, without going through IT or procurement. They are common, and they are exactly the uses most likely to be missed.

Keeping it current

The inventory has a cost. Someone must keep it current as vendors add AI features to existing products, which happens through routine software updates. A human resources system that had no AI last year may gain a candidate-ranking feature in a quarterly release, with nothing more than a line in the release notes.

Procurement is the natural control point. A standard contract clause requiring vendors to disclose new AI functionality keeps the inventory honest without a separate discovery exercise. Rather than hunting for AI after the fact, make vendors say when they add it.

The contract clauses that do the work

A vendor disclosure clause works best with specifics. It should require notice before a release that adds or materially changes AI functionality, with a description of what the feature does and what data it uses. It should require documentation the deployer needs for an impact assessment or bias audit, which Colorado’s law already obliges developers to provide. It should give the customer the ability to switch a feature off without losing the product. And it should require the vendor to cooperate with audits and complaints.

The trade-off is negotiating effort. Large platform vendors rarely accept bespoke clauses from a single customer, so smaller banks, credit unions and county agencies may get only the vendor’s standard terms. Cooperative purchasing vehicles and state contracts can help, because they negotiate once for many buyers. Where the clause cannot be won, the fallback is a procurement step: reading release notes for AI features each quarter and updating the inventory by hand.

Laws on transparency and disclosure

The third type requires organizations to tell people when they are dealing with AI, to label synthetic content or, for developers of large models, to publish information about safety practices. Synthetic content means images, audio, video or text generated by AI rather than captured or written by a person. Utah’s disclosure rules for generative AI, which apply with particular force to regulated occupations, and state laws on deepfakes in political advertising fall here. A deepfake is a convincing fake video or audio clip of a real person.

For most banks, health systems and agencies, these laws mainly touch customer-facing chatbots, virtual assistants and generated letters. The controls sit at product level: standard disclosure language, logs of AI-generated content and a pre-launch review against current state requirements.

The logs serve a second purpose. When a customer disputes what an assistant told them, the organization needs the transcript. Suppose a customer says the bank’s chat assistant promised to waive a fee. Without a record of the conversation, the bank is left arguing from memory. Building logging once, for all customer-facing AI, costs less than retrofitting it for each new law.

How disclosure duties differ

The disclosure laws vary in trigger and timing. Colorado’s act requires anyone deploying an AI system intended to interact with consumers to tell them they are interacting with AI, unless that would be obvious. Utah’s 2024 law drew a line by profession: people in state-regulated occupations had to disclose generative AI use prominently, while other businesses had to disclose when a consumer asked. Some laws focus on synthetic media and political advertising and barely touch a bank or hospital.

The design choice for a product team is where disclosure lives. A banner on a chat window is simple. A voice assistant on a phone line needs a spoken statement at the start. A letter drafted by AI and reviewed by a staff member before sending raises a harder question about whether disclosure applies at all, and the answer differs by statute. A single disclosure standard set at the strictest common level, applied everywhere, avoids state-by-state variants. The cost is some disclosure in places that do not strictly require it, which customers rarely object to.

Logging carries its own trade-off. Transcripts help resolve disputes, but they also hold personal and sometimes sensitive information. Retention should follow the organization’s existing records schedule, with access controls that match the data. Health systems must treat transcripts that contain protected health information under HIPAA. Banks should consider whether transcripts form part of a consumer complaint record.

The extra layer for public agencies

Public agencies have one more layer. Several states have passed laws requiring their own agencies to inventory AI systems and assess them before use. An agency’s program should fold those duties into the same inventory, so that one record supports both public reporting and internal governance. A county human services department, for instance, would list its benefits-eligibility screening tool once, with its owner, purpose, assessment and review process, and draw both its state report and its internal oversight from that single record.

Public records law adds a twist that private firms do not face. Chatbot transcripts, inventory entries and impact assessments held by a public agency may be subject to disclosure on request, depending on the state’s statute. Agencies should write them on the assumption that the public may read them, and should agree retention periods with their records officer before launch rather than after.

Federal agencies had their own regime. OMB memorandum M-25-21, issued in April 2025, required them to keep AI use case inventories and apply minimum risk management practices to what it called high-impact AI. State and local agencies administering federal programs may meet those expectations through grant terms or partner agencies. A state Medicaid or unemployment agency should expect federal partners to ask for the same inventory data the state law requires.

How the program runs day to day

Someone has to do this work, and AI compliance tends to fall between chairs. Technology teams assume it is a legal matter; lawyers assume it is a technical one; business units assume someone else has it in hand.

A workable division of labor looks something like this. The business owner of each use case, such as the head of consumer lending or the director of patient access, is accountable for the decision and for the customer’s experience of it. Compliance and legal track which laws apply where, maintain the notice and explanation templates, and decide when a new statute requires a change. Model risk or data science teams validate how systems perform and watch measures such as the review overturn rate. Procurement enforces the vendor disclosure clause and flags new AI features before contracts renew. And an operations team staffs the human review queue and keeps its service levels.

The inventory is the shared tool that ties these roles together. Each entry names the owner, the decision involved, the jurisdictions affected, the notice used, the review route and where records are kept. When something changes, whether a new law, a new vendor feature or a spike in complaints, everyone works from the same record.

The rhythm matters as much as the structure. A short quarterly review, in which each business owner confirms their entries are accurate and compliance reports on new or changed state laws, keeps the program alive without turning it into a bureaucracy. Most of the time the review takes an hour. Occasionally it surfaces a tool nobody knew about, which is exactly what it is for.

For a board, the useful report is short: how many AI-influenced decision points the organization has, how many adverse decisions received a completed explanation, how long human reviews took, how often they overturned the system, and which state requirements changed this quarter. Those five figures say more about real risk than a long policy document.

Sequencing the first ninety days

A program starting from nothing can be sequenced. In the first month, compliance and the business owners build the register of decision points, starting with lending, hiring, claims, eligibility and patient outreach. In the second month, they compare existing notices and appeal routes against the procedural core and fix the gaps, beginning with the highest-volume decisions. In the third month, procurement adds the disclosure clause to new contracts and renewals, and model risk sets up the overturn-rate measure.

The programmatic layer can follow on a separate track, led by counsel. That track watches each state’s effective dates and rulemaking, drafts impact assessment templates when the requirements settle, and decides whether to adopt a recognized framework such as the NIST AI Risk Management Framework. Keeping the two tracks apart stops a delay in one state’s law from stalling work that helps customers now.

The evidence an examiner would ask for

An examiner or state attorney general investigating a complaint will usually ask for the same records. They want the inventory entry for the system, the notice the consumer received and the reasons given. They want the reviewer’s file and the time taken, along with evidence that the system was tested for disparate outcomes and the result. An organization that can produce those within a day is in a strong position, whichever statute the inquiry rests on. One that has policies but no records is not.

One inventory, one backbone

Pulling the three types together, a sound AI compliance program on moving ground has a recognizable shape. It has one inventory, not several. It has one procedural backbone for notice, explanation, correction, human review and records, reused across every state. It treats impact assessments and program duties as a volatile layer, tracked by counsel and added as each state settles. And it ties vendor contracts to the inventory, so new AI features are disclosed rather than discovered.

The mistakes are equally recognizable:

• Building state by state. Creating a separate Colorado program, then a separate program for the next state, produces duplicated effort and inconsistent treatment of customers.
• Starting with the volatile layer. Spending the first year drafting detailed impact assessment templates for a law whose timing and shape may change, while customers still receive unclear letters.
• Creating parallel processes. Standing up a new AI appeals process next to an existing credit or clinical appeals process, which confuses staff and customers alike.
• Trusting the survey. Relying on business units to self-report AI use, which reliably misses tools bought on a card and features added in a software update.
• Forgetting the records. Making good decisions but being unable to show, months later, what was decided and why.

The business case is straightforward. A stable procedural core reduces the cost of each new law, because most of the work is already done. It lowers complaint and litigation risk, because customers get explanations and a real route to review. It gives boards and regulators evidence rather than assurances. And it treats customers better, which is what the laws are trying to achieve in the first place.

The case for waiting, and why it fails

The best argument against this program is that federal preemption could still arrive, making state-specific work wasted. That argument had force before July 1, 2025, and it returned in a different form later that year. It fails for the procedural core because that core does not depend on state law. ECOA and the Fair Credit Reporting Act already require reasons for adverse credit decisions. Section 1557 already reaches clinical decision tools. Unfair and deceptive practices law already reaches misleading chatbots. A federal statute that preempted every state AI law would leave all of that in place.

A second argument is that small institutions cannot afford this. A community bank or rural hospital may have a single compliance officer. For them the inventory may be a spreadsheet, the human review queue an existing complaints process, and the vendor clause a standard addendum. The structure scales down because it reuses what already exists. The programmatic layer is where small organizations should lean hardest on exemptions, frameworks and vendor documentation, after checking with counsel which apply.

Where the argument is weakest

The approach is weakest for frontier model developers and large technology vendors, whose state obligations concentrate in the programmatic and transparency layers. It is also weaker for insurers, whose state regulators set detailed algorithm rules that are themselves the core. For banks, health systems and agencies that buy rather than build, the procedural core remains the right first investment.

Tests for the stable core

1. Where in our business does AI materially influence a decision about a customer, patient, employee or resident, and do we have that list written down?
2. When one of those decisions goes against someone, what exactly do we tell them, how do they ask for human review, and how long does it usually take?
3. How often do human reviewers overturn the original outcome, who looks at that number, and are we reusing our existing adverse action, appeals and clinical review processes or running something parallel?
4. Is our AI use case inventory tagged by function, sector and jurisdiction, and do our vendor contracts require disclosure of new AI features?
5. Which of our obligations are stable and which are still changing, who is tracking the changes state by state, and have we checked with counsel which federal-supervision exemptions apply to us?

A board or audit committee does not need to answer these questions itself. It should expect management to answer them in writing once a year, with evidence attached. The first answers will expose gaps. The value lies in the second year, when the same questions show whether those gaps closed.

Internal audit can use the same tests as the basis of a review. Each one produces evidence an auditor can sample: the register, a set of adverse decision files, the review queue’s timing data, the inventory tags and a sample of vendor contracts. A finding against any of them is specific enough to fix within a quarter.

With hindsight

States kept legislating. Texas enacted its Responsible AI Governance Act in June 2025, effective January 2026, and California enacted a frontier AI transparency law in September 2025. “Frontier” refers to the largest and most capable AI models, built by a small number of developers. The patchwork that the July vote preserved did not stand still.

Texas’s law, which took effect on January 1, 2026, put much of its weight on government use. It required state agencies to disclose AI use to people interacting with them and barred certain intentional harmful uses, with enforcement by the attorney general. California’s frontier law, SB 53, was signed on September 29, 2025, and aimed at the largest model developers rather than the banks, hospitals and agencies that buy from them. Neither law changed the procedural core described here. Both added to the transparency and government-use layers.

Federal pressure returned from a different direction. In December 2025, an executive order directed the Justice Department to form a task force to challenge state AI laws it considered burdensome, and asked Commerce to evaluate them. Where Congress had declined to preempt state law, the executive branch moved to contest it through other means.

The order, Executive Order 14365, was signed on December 11, 2025. It gave the Attorney General 30 days to set up an AI Litigation Task Force. It gave Commerce 90 days to publish an evaluation identifying onerous state AI laws. It moved to make states with such laws ineligible for non-deployment BEAD broadband funds, as far as federal law allowed, which echoed the Senate’s tie to broadband money that summer. It directed the FCC and FTC to take related steps and asked White House advisers to draft legislation for a uniform federal standard.

The order’s own carve-outs are telling for public agencies. The legislative recommendation was not to preempt otherwise lawful state laws on child safety, on AI compute and data center infrastructure, or on state government procurement and use of AI. A state’s rules for its own agencies were therefore the part of the patchwork the order explicitly left alone. An executive order cannot itself repeal a state statute. Any preemption would have to come from Congress or the courts, a process measured in years.

Colorado shows why the procedural core was the safer investment. In August 2025, lawmakers delayed the AI Act to June 30, 2026. That delay left the timing and final shape of the assessment and risk program duties uncertain well into 2026. A team that had poured its first year into those duties found its work in limbo.

The delay came through SB25B-004, passed in an extraordinary session and signed on August 28, 2025. It moved only the effective date. The substance of the 2024 act, including the procedural duties, stayed in the statute while lawmakers and industry continued to argue about the programmatic layer.

The consumer-facing duties, by contrast, echo adverse action and appeal processes that lenders and health systems already run. Organizations that built notice, explanation, data correction, human review and records first had work that would carry over whatever form the law finally took. Those processes help customers today, satisfy regulators who already expect fair treatment, and would remain useful even if every state AI statute were challenged or rewritten.

The same principle applies to any regulation still being written. When the rules are moving, invest first in the parts that every version of the rules will need, and treat the rest as an option you exercise later. Preemption litigation and federal legislation remain open questions to follow with counsel, and a program built on the stable core is well placed to absorb whatever answer arrives.