Do state AI procurement rules apply if AI is only a feature of my product?

Usually yes. Most state AI purchasing standards are written around whether the system uses automated decision-making or generative output in the course of the agency's work, not around whether the product is marketed as an AI product. A scheduling tool with an AI ranking feature, a case-management platform with a summarisation feature, or an analytics product with a predictive score all tend to fall inside the definition. Vendors are frequently caught out here, because they answer 'no' on the AI-disclosure question and then have to amend the response once the agency's technical reviewer identifies the feature.

What is an AI inventory requirement and why does it reach vendors?

Many states require agencies to maintain a public or internal inventory of the automated systems they use, including purpose, data categories, and whether the system informs a decision affecting a person's rights or benefits. Agencies cannot populate that inventory without vendor-supplied facts, so the obligation is pushed into the solicitation and then into the contract as an ongoing disclosure duty. The practical effect for a vendor is that model changes, new data sources, and new inference types become reportable events during the contract term, not just at award.

Will an agency accept our SOC 2 report as AI compliance evidence?

It covers a different question. SOC 2 speaks to security and operational controls; AI procurement standards ask about intended use, training data provenance, testing for disparate outcomes, human review, and failure behaviour. A security attestation is necessary and usually requested separately, but it does not answer the AI-specific items. Vendors that have an ISO/IEC 42001 certificate or a documented NIST AI RMF mapping are in a materially stronger position, because those frameworks produce artefacts that line up with what the AI questions are asking for.

Can a public-sector contract force us to disclose training data?

It can require disclosure of categories, provenance, and licensing basis, and increasingly it does. It rarely requires the dataset itself. The more consequential term is usually the other direction: restrictions on the vendor using agency data or resident data to train or improve models, which is now a common default clause. Vendors whose product roadmap assumes customer-data training should read that clause before pricing the bid, because agreeing to it and then breaching it is a contract-termination and debarment issue, not a renegotiation.

Is public records law a risk for the documents we submit?

Yes. Materials submitted to a public agency are presumptively disclosable under state public records law unless a trade-secret or proprietary-information exemption applies, and the exemption generally has to be claimed properly at submission rather than after a request lands. Model documentation, evaluation results, and architecture descriptions are exactly the kind of material a competitor or journalist will request. Mark and segregate proprietary material at the time of filing and follow the state's specific procedure for it.

Selling AI Software to State Government 2026: The Procurement Questions That Disqualify You

Why the Federal Playbook Does Not Transfer

Vendors selling into federal agencies have a well-worn path: get a cloud authorisation, satisfy the security controls, land on a schedule. State and local procurement has no equivalent single gate for AI. Instead, states have been attaching AI-specific requirements to ordinary solicitations — a disclosure section in the RFP, a set of mandatory terms in the resulting contract, and a reporting duty that runs afterwards.

That structure has a consequence worth planning around. There is nothing to certify once and reuse everywhere. Each jurisdiction asks its own version of the same handful of questions, and the differences are in wording rather than substance. A vendor that builds one well-evidenced answer set can respond to most of them; a vendor that treats each RFP as a fresh writing exercise burns weeks per bid and produces inconsistent answers that a diligent evaluator will notice.

The Five Questions Behind Every Variation

1. Is there AI in it, and where exactly?

A feature-level answer, not a product-level one. Name the component, the task it performs, and whether its output reaches a person or only an internal reviewer.

2. What decision does it touch?

Eligibility, benefits, enforcement, employment, and housing are the categories that trigger the heavier requirements. Say plainly whether the system informs one of them.

3. What have you tested, and what did the test show?

Accuracy figures alone rarely suffice. Expect to describe evaluation across subgroups, the dataset used, when it was run, and who ran it.

4. What happens with agency and resident data?

Whether it trains models, whether it leaves the jurisdiction, which subprocessors touch it, and how it is deleted at contract end.

5. What does a human do?

Where review sits in the workflow, what the reviewer sees, and whether they can override. 'A human is in the loop' with no described mechanism reads as a non-answer.

The Bid-Ready Artefact Set

Prepare Before the Next Solicitation

☐A one-page intended-use statement per AI feature, written for a non-technical evaluator
☐A data-flow description naming every subprocessor, storage region, and retention period
☐Evaluation results with dates, dataset description, and subgroup breakdowns where the use case affects individuals
☐A written human-review procedure describing what the reviewer sees and what they can change
☐A model-change notification process you can actually operate at contract scale

Contract Terms to Price In, Not Argue With

☐No training on agency or resident data — assume this is non-negotiable and check your architecture against it
☐Audit and inspection rights covering AI documentation, not just security posture
☐Notice obligations when the underlying model or a foundation-model dependency changes
☐Deletion and return of data at termination, including derived features and embeddings
☐Indemnity for third-party IP claims arising from generated output

Filing Hygiene

☐Claim trade-secret protection at submission using the jurisdiction's stated procedure — not afterwards
☐Segregate proprietary material into clearly marked exhibits so an agency can release the rest
☐Keep answers consistent across jurisdictions; evaluators do compare against public award files
☐Answer the AI-disclosure question 'yes' whenever a feature could plausibly qualify

The Disqualification Is Usually Procedural

Public procurement runs on responsiveness. A bid that omits a required disclosure, or answers it in a way that does not match the form, can be set aside without reaching the merits — and the vendor often never learns that was the reason. This is a different failure mode from private enterprise sales, where an incomplete security questionnaire triggers a follow-up email rather than an elimination.

It is also why the artefact set above is worth building once, in advance. The response window on a public solicitation is short, the questions are known in substance, and the evidence they ask for — subgroup evaluation, data-flow mapping, documented human review — takes longer to produce than the window allows if you start from nothing.

Frequently Asked Questions

Do these rules apply to city and county buyers too?

Increasingly. Large municipalities have adopted their own AI purchasing policies, and smaller ones frequently borrow the state's language wholesale or buy from a state master contract that carries the terms through. Assume the requirements travel down, and read the flow-through clauses in any cooperative-purchasing vehicle you list on.

We use a third-party foundation model. Whose disclosures are these?

Yours, as the contracting party. The agency's counterparty is you, and the disclosure and notice duties land on the entity holding the contract. That makes your upstream provider's terms a compliance dependency: if you cannot get notice of model changes from your provider, you cannot reliably give notice to the agency. Negotiate that visibility upstream before you promise it downstream.

Is a bias audit the same thing as the testing evidence they want?

Related but not identical. A bias audit under an employment-specific law like NYC Local Law 144 is a narrowly scoped, formatted exercise for a defined tool category. Procurement testing evidence is broader and less standardised: it asks what you evaluated, on what data, when, and what the results were, for whatever the system actually does. An audit report is useful supporting material, but it does not by itself answer the general question.

How long do these obligations last after award?

For the contract term and often past it. Reporting duties run continuously, records-retention clauses commonly outlive the engagement, and data-deletion obligations attach at termination. Treat the award as the beginning of the compliance work rather than the end of it, and staff the notification process accordingly.

The buyer-side version of the same questions, and how to answer them consistently.

FedRAMP AI SaaS Authorization 2026

What the federal cloud authorisation path does and does not cover for AI features.

ISO 42001 AI Management System Certification 2026

The certification that produces most of the artefacts a public buyer asks for.

NIST AI RMF Implementation Checklist 2026

A framework mapping that gives structure to testing and documentation answers.

Build the Answer Set Once

Every jurisdiction asks the same five questions in different words. The vendors that win public-sector AI work are not the ones with the best answers — they are the ones who had answers ready, with dated evidence behind them, on the day the solicitation dropped.

Public-sector buyers also carry digital accessibility obligations that flow through to the software they procure, so the accessibility posture of your interface is scored in the same bid. Fix that before it costs you a responsive submission.

The AI Questions in a Government Bid Are Scored Before Anyone Looks at Your Demo