Diagnostic

Can you account for what your AI agents did?

Four modules — readiness scoring, obligation mapping, evidence gaps and modeled exposure. It starts at eight questions and goes deeper only where your answers reveal a gap.

Discuss an audit deadline Or run the assessment below ↓

No email required. Nothing you enter identifies you or your organization. An anonymous copy of the four scores is retained to build a sector benchmark — no organization name, no identifier, no IP address, no cookies.

This site tracks nothing that identifies you — no cookies, no session recording, no person profiles. The same standard it assesses you against, applied to itself: here is the evidence.

0 answered · 0% complete

Context

what the agents touch, and who supervises you

Which regime governs you most directly?

Where does the obligation land?

Pick the jurisdiction whose supervisor would examine you first.

What is the most consequential thing an agent can do today without a human approving it?

Roughly how many agent-initiated actions occur per month?

Entitlements

Not assessed—

When an agent takes an action, can you name the human authority it derives from?

Not which credential it used — which person or governing body authorized that class of action.

Policy enforcement

Not assessed—

Where is policy evaluated for an agent action?

The question is when the check runs, not whether a policy document exists.

Attestation & audit

Not assessed—

Could you reconstruct, six months from now, exactly why a specific agent action was permitted?

The reconstruction test is the single most predictive question in this diagnostic.

Model risk governance

Not assessed—

How do your agents appear in model risk inventory?

Answer five questions to unlock the full report. The panel beside these questions is already live, and updates with every answer.

Live score0% complete

—0 of 4 layers assessed

An overall score is published once all four layers have been asked about. Until then it would average real answers against layers nobody put a question to.

Entitlementsnot assessed — no question covered this layer

Policy enforcementnot assessed — no question covered this layer

Attestation & auditnot assessed — no question covered this layer

Model risk governancenot assessed — no question covered this layer

A layer with no questions behind it is reported as not assessed, never as a zero. A score for something nobody asked about is the failure this instrument exists to find.

Method, in full

What the diagnostic tests

The interactive assessment above is backed by a deterministic scoring model. The same answers always produce the same score. These are the questions and decision rules in prose so a reviewer—or a crawler that does not execute JavaScript—can examine the method.

Entitlements

Core test: When an agent takes an action, can you name the human authority it derives from?

Most agent stacks inherit a service account holding the union of everyone's permissions. Authority has to be granted, scoped, time-boxed and recorded — per agent, per action class, traceable to the human who delegated it.

Follow-up depth: What do your agents' credentials actually permit? · Does an agent's authority expire or get re-examined? · Are business limits (value ceilings, rate caps, jurisdiction) enforced on the action, or assumed in the prompt?

Policy enforcement

Core test: Where is policy evaluated for an agent action?

A control that runs at login and never again is not a control. Enforcement belongs at the moment of the action, between the agent's intent and the system of record, where it can still refuse.

Follow-up depth: Can the enforcement layer actually refuse an action in flight? · When an agent makes a customer-facing commitment, is it checked against an authoritative source? · For third-party agent platforms, can you enforce your own policy inside their execution path?

Attestation & audit

Core test: Could you reconstruct, six months from now, exactly why a specific agent action was permitted?

Chat transcripts are not traceability. Evidence means the decision, the authority it rested on, the policy evaluated, and the inputs — retained and reconstructible on a reviewer's timeline, not yours.

Follow-up depth: Do your records capture which model and prompt versions were in effect at the moment of the action? · Is evidence retained for as long as your regime requires? · Where a human is 'in the loop', can you evidence what they actually saw and how long they had?

Model risk governance

Core test: How do your agents appear in model risk inventory?

Agent behavior changes the day a provider ships a new model version, with no code change on your side. Inventory, validation, effective challenge and monitoring all have to account for a system that acts rather than predicts.

Follow-up depth: When a provider ships a new model version, what happens? · Can an independent reviewer mount effective challenge to an agent's behavior? · If an agent starts acting wrongly at scale, what stops it?

Band	Score	Interpretation
Critical gap	0–29	Authority or evidence is absent; a review is likely to stop here.
Weak	30–54	Controls exist informally but cannot be reconstructed or enforced consistently.
Developing	55–79	Core controls exist; gaps remain in coverage, evidence or independent challenge.
Defensible	80–100	The design is likely defensible, subject to testing against a real historical action.

Context that changes the obligation map

Which regime governs you most directly?
Where does the obligation land?
What is the most consequential thing an agent can do today without a human approving it?
Roughly how many agent-initiated actions occur per month?

Entitlements and attestation receive additional weight because enforcement and model-risk controls depend on knowing who authorized an action and being able to reconstruct it later. The exposure range is directional prioritization—not a penalty forecast.