Intelai
AI in Regulated Work Is an Evidence Problem
Healthcare and defense have the same blocker. It is not model quality. It is whether you can show a control was satisfied when the system acted.
A hospital compliance officer and a defense program manager will describe their AI problem in completely different vocabulary. Strip the vocabulary and it is the same problem. Neither one can approve a system that cannot show what it did.
That is not a technology gap. It is an evidence gap, and most AI vendors are still answering a question about capability when the buyer is asking a question about proof.
Regulation is a specification for evidence
HIPAA tells a covered entity how to protect patient information and when it may be used. The Security Rule asks for administrative, physical, and technical safeguards. The Privacy Rule includes a minimum necessary standard for most uses and disclosures.
CMMC Level 2, for defense contractors handling controlled unclassified information, maps to 110 requirements drawn from NIST SP 800-171. FedRAMP does the same kind of thing for cloud services, at a different altitude, with a different assessor.
None of those frameworks say anything about model architecture. All of them say something about access, logging, retention, and the ability to reconstruct who did what.
That is the specification an AI system has to satisfy in a regulated environment. Not "is it accurate." Can you show, after the fact, that only authorized people could reach the data, that the use was permitted, and that the record is intact.
Translate the control into a behavior
The useful exercise is one column of controls and one column of system behaviors, and a line between each pair. If a line has no counterpart, the control is being satisfied by a policy document rather than by the system.
Access control becomes: who can read the prompt, the retrieved context, and the output, and is that enforced at the data layer or the interface.
Minimum necessary becomes: what the retrieval step is allowed to pull, and whether a query can reach records outside the purpose of the task.
Audit becomes: what is recorded per action, where the record lives, how long it is retained, and whether it can be altered.
Data retention becomes: where the vendor stores the prompt and for how long, including in logs and in any training pipeline, with a clause that binds it.
Incident response becomes: what the system does when it produces a disclosure it should not have, and who is notified.
Doing this once, for one system, takes a week. It also converts a vague compliance conversation into a list of engineering tasks, which is the only form in which that conversation produces anything.
A control that lives only in a policy document is a hope with a signature block.
Why the culture advice does not survive contact
The standard consulting answer to regulated AI is a culture problem: run hackathons, celebrate small wins, empower teams, build cross functional collaboration.
I have never seen a hackathon produce an authorization to operate.
The reason is structural. In a regulated organization, the person who can say yes is not the person who is enthusiastic about the demo. The approver is a compliance officer, a security assessor, or an authorizing official, and their job is to find the gap. Enthusiasm does not close a gap. Evidence does.
That does not mean culture is irrelevant. It means the culture that matters is the one where an engineer can tell a compliance officer that a control is not yet satisfied without being treated as an obstacle. Programs where bad news travels are the ones that get approved, because they fix things before the assessment rather than during it.
Three gates, in order
Every regulated AI program I have seen pass through the same three gates. The order is not optional.
The data gate. Where the data comes from, what basis you have for using it, and where it is allowed to live. Most programs fail here first and spend months discovering that the data they planned to use cannot be used for the purpose they planned.
The decision gate. What the system is allowed to decide, what a human decides, and what happens when the two disagree. This is where a compliance officer can actually help you, because they have seen what happens when an organization cannot explain a decision.
The evidence gate. What record exists after an action, who can read it, and whether it would survive an assessment. This is the gate that gets deferred, and it is the gate that determines whether you ship.
Programs that sequence the gates backwards, buying technology and then looking for a use, spend their budget on a pilot that cannot be approved.
What to ask a vendor in a regulated market
Do not ask whether they are compliant. Every vendor says yes, and the word means nothing without a scope.
Ask which controls they satisfy, which ones the customer has to satisfy, and where the boundary sits. Ask for the shared responsibility model in writing. Ask what the audit record contains and whether you can export it. Ask what happens to your data when the contract ends.
A vendor that can answer those questions has done this before. A vendor that answers with a certification logo has not, or is hoping you will not notice the difference.
Regulated industries are slow for a reason that is easy to mock and hard to argue with. The cost of a mistake is borne by someone who did not consent to the experiment.
AI does not change that arithmetic. It changes the speed at which a mistake can propagate, which makes the evidence more important, not less.
The organizations that get value from AI in these sectors will be the ones that treated compliance as a design constraint from the first week, and built the record before they built the model. Everyone else will build something impressive that nobody can approve.