Inventory your AI. Control it. Prove it.
A customer security review, your board, your insurer, legal, or whoever finds the tool nobody approved. Assay makes sure the answer already exists. We find the AI you are actually running, put controls on the systems that create exposure, and build the evidence behind them, so answering takes an afternoon instead of three weeks.
The gap
88%
of organizations report using AI in at least one business function.
McKinsey
1 in 3
have reached governance maturity level three or higher out of four. Roughly two thirds are running AI on governance that is still early.
McKinsey
That gap is not a failure of anyone's judgment. Adoption was encouraged, it was fast, and it happened team by team. Governance is the slower, duller work that nobody owned while it was happening.
Where it becomes expensive
The questionnaire you used to clear in two days now has a section nobody owns, and the deal waits while engineering and legal argue about who answers it.
You shipped an AI feature into a system already covered by controls you attest to. Nobody updated the control language, because nobody thought of it as a compliance change. The next person to look will find it before you do.
A prospect asks whether their data trains anything, and the honest internal answer is that it depends who you ask. That is a governance gap wearing a technical costume.
Teams adopted tools without review because adoption was encouraged and governance was not. You need an inventory before you need a policy.
The obligation with a date on it
Article 4 of the EU AI Act requires every organization that provides or deploys an AI system to ensure its staff have a sufficient level of AI literacy for the role they are in. The duty itself has applied since 2 February 2025. What changed on 2 August 2026 is that national market surveillance authorities and their penalty regimes came online, so there is now somebody whose job it is to ask.
It reaches further than most US companies assume. You are in scope if you have staff in the EU, if you place an AI system on the EU market, or if the output of your AI system is used in the EU. Risk classification does not get you out of it. This one applies to any AI system, including the chat assistant somebody turned on last quarter.
Article 4 carries no dedicated fine tier of its own. That is not comfort. It means a literacy failure gets characterised by the authority inside whatever broader investigation is already open, and the tier it lands in is their call, not yours.
We build the training and the record that shows it happened: role appropriate sessions for the people who actually touch these systems, a written literacy plan, and attendance and comprehension evidence that survives being asked for. The evidence is the half organizations forget, and it is the only half an authority can read.
Assay is not a law firm and this is not legal advice. Confirm applicability with your own counsel. What we do is build the program and the proof once you know you are in scope.
What we do
Discovery tooling does the mechanical part, and we stand it up so it keeps running after we leave rather than going stale the day we do. Then the part tooling cannot do: which systems touch customer data, which make decisions about people, and which six of the forty actually create exposure.
Human review where it matters, access and data boundaries, change and model update handling, incident paths. Mapped to the NIST AI Risk Management Framework and ISO/IEC 42001 so the crosswalk already exists when someone asks for it.
Evidence that accumulates from work people already do, rather than a scramble every time a review lands. This is the part policy-only exercises skip, and it is the part that decides whether any of it survives contact with an auditor.
The order matters. A policy written before the inventory describes a company you do not have.
When the inventory turns up something that needs rebuilding rather than documenting, that is Strake, our sister practice.
Being straight with you
How we see it
Most organizations can produce an AI policy in a week. Templates are everywhere and most of them are fine. What almost nobody can produce is evidence: which systems exist, what data moved through them, who reviewed what, and when.
That is what a customer's security review actually asks for, and it is the one thing a policy cannot answer. Governance that lives only in a document fails the first time somebody checks it.
Questions we get
Possibly, and it catches people out. Article 4, the AI literacy duty, reaches you if you have staff in the EU, if you place an AI system on the EU market, or if the output of your AI system is used in the EU. It applies to any AI system regardless of risk classification, so an off the shelf assistant counts. The duty has applied since 2 February 2025 and national enforcement authorities came online on 2 August 2026. We are not a law firm, so confirm scope with counsel. Once you know you are in scope, we build the training and the evidence that it happened.
No. Assay is not an audit firm and issues no certifications or opinions. We build the inventory, controls, documentation and evidence so that a customer review, an auditor, or your own committee gets a straight answer quickly.
Primarily the NIST AI Risk Management Framework and ISO/IEC 42001. In practice the framework matters less than what your customers are actually asking you, and that is usually the real driver.
Mostly not, and that is good news. If you already have controls that work, this is a scoping, documentation and monitoring extension of them rather than a second program built from nothing. Starting from something is far cheaper than starting cold.
Often yes, and that case is usually harder, because the tools were adopted without review and nobody has an inventory. Using someone else's model does not move accountability off you in your customer's eyes.
For discovery, largely yes, and you should. Tools in this category run a few dollars per employee per month, they run continuously, and they will find more than a one-time manual sweep. We will tell you which one fits and help you set it up. What they give you is a list. What they cannot give you is which items on it create exposure, what you are obligated to say about them, and the evidence trail behind the answer. That judgment is the engagement.
A gap assessment is fixed price and scoped on a short call, so you know the number before you commit. Larger readiness engagements are also quoted fixed before work starts. The first call is free.
Control and Function LLC, a Denver based consulting practice.
Start here
Thirty minutes, no charge. Bring the questionnaire that stalled, or just the question you could not answer cleanly. You will leave knowing whether you have an inventory problem, a control problem, an evidence problem, or none of the above.