Three connected practices produce a single maintained record — not three reports that age on a shared drive. The work is designed so that what it leaves behind can be opened, questioned and reviewed by someone who was not in the room.
It arrives through procurement, through vendor features switched on by default, and through people who found a tool that helped and used it. By the time a board, a client, an insurer or a regulator asks what governs it, the honest answer is usually that nobody has looked at all of it at once.
That is not a technology problem. It is an accountability problem, and it is answerable. The work is to establish what is running, what it could affect, who owns it, and what evidence exists that anyone reviewed it — in that order, because a policy written before anyone knows what is running governs a system nobody has described.
Three measures are reported separately and never blended into a score. Exposure describes what a system could affect. Readiness describes whether anyone is managing it. Evidence describes whether any of that can be shown to someone who asks. A single figure hides all three, and hiding them is the opposite of the job.
What rules and responsibilities govern our AI?
Every AI system in use, including embedded vendor features and tools staff adopted themselves. Unapproved use is recorded rather than treated as an embarrassment — an inventory that omits it describes a different organization.
A named person accountable for each material system. A specific person, not a committee and not a department, because a committee cannot be asked a question on a Tuesday.
Who may approve a new system, who must be consulted, who can stop one. Written down at a level of detail that survives the person who wrote it leaving.
A defined route a new AI tool travels before it is used, with a decision recorded at the end of it. Informal approval is the most common finding and the cheapest to fix.
Rules that reference the inventory rather than the abstraction. A policy that cannot be tested against what is running is a statement of intent.
Atlas is a capability, not a product with a scope and a price attached to it. What it covers in a given engagement depends on what the inventory turns out to contain.
What could go wrong — and what proves our safeguards work?
What each system could affect, classified on its own terms. Exposure floors do not move: a system that materially influences decisions about people, or acts without a person approving each step, sits at the top of the scale regardless of how strong governance is elsewhere.
Risks stated specifically enough to be wrong. "Bias" is not a risk; a named system producing a named effect on a named group, through a described mechanism, is.
Each risk mapped to the safeguard that addresses it, with a named owner, recorded exceptions, and residual risk stated rather than implied.
What would have to exist for each control to be demonstrable — before anyone goes looking for it. This is where most organizations discover the gap between what they believe they can show and what they can produce.
What the supplier is contractually responsible for, what they actually publish, and what remains yours regardless of what the contract says.
How you would find out that a system acted outside its intended scope, who is told, and what stops it. Detection that depends on the system reporting itself is not detection.
Strata does not provide independent assurance. Where independence is required, the engagement is routed to a qualified independent provider — that is a boundary, not a preference.
Do our people understand and apply those rules?
What each role may and may not do, written for the role. A single all-staff email is not guidance; it is a record that something was sent.
Who a person tells when a system produces something wrong, and what happens next. Most staff can describe the tool and not the route.
Confirmation that the guidance was understood, not that it was delivered. The difference is the whole point.
Completion recorded per role, so the record can answer who was prepared for what, and when — rather than that training happened.
Nexus prepares people to operate and oversee systems. It does not certify competence, and no completion record here should be presented as one.
Each practice writes into the same record rather than producing its own document. Six sections, maintained, with review dates — so the question "what governs our AI?" has a single answer that can be opened rather than three that have to be reconciled.
The record is a records mark, not an approval mark. It states what exists and what has been reviewed. It does not certify anything, and the seal that appears on it says maintained · reviewable for that reason.
The record separates what an organization has said, what it can point to, and what someone accountable has actually examined. Almost every governance failure sits in the distance between the first and the third. Take one claim through all three.
“We have an AI policy, and it requires a risk assessment before any system goes live.”
That the organization has taken a position and can say what it is. This is not nothing — an organization that cannot state its own rule has no rule.
That the policy matches what is actually running, that anyone owns it, that it has ever been applied, or that a single assessment exists. A declaration is a claim about the world, not a fact about it.
Only the third is evidence in the sense a regulator, an insurer or a board means it. The other two are worth having — you cannot reach the third without them — but a record that does not distinguish between them will describe an organization as prepared when it is merely articulate.
What you are running, who owns it, what evidence exists. Short, and specific to your organization. Scope and fee are confirmed after this — never before, and never from a form.
Drawn from the practices in whatever combination the situation requires. Atlas, Strata and Nexus are capabilities, not packages: an organization with a good inventory and no evidence needs a different engagement from one with neither.
The Governance Record, with review dates, in your hands. The test is whether it still answers questions six months later without us in the room.
Zevana provides governance, risk-readiness and workforce-readiness services. It does not provide legal advice, independent assurance, certification or statutory audit, and nothing produced in an engagement should be presented as any of those. Where independence is required — and for some questions it is — the engagement is routed to a qualified independent provider.
Stating this is part of the method rather than a disclaimer attached to it. A governance firm that is vague about the limits of its own work has already demonstrated the problem it claims to solve.
Twelve questions, about five minutes, nothing stored — or a conversation, if you would rather begin there.