Group the AI estate by EU AI Act tier, read the written reason behind every classification, and declare the profile that produces a real tier.

Risk classification groups the AI estate by EU AI Act tier and prints the classifier's own reason beside every asset. Nothing on this page shows a tier without showing why: a badge whose reason is one hover away is a badge that gets screenshotted without it.

What do I need?

Licence: Any
Role: admin_userviewer

Prerequisites

Discovery has run at least once, so there are assets to classify.
You are signed in as admin_user or superadmin to declare or change a risk profile.
You know what the asset actually does — the domain it is used in, whether a person reviews its output, and whether it affects decisions about people.

admin_user, superadmin, auditor and viewer may read the page. Declaring a profile is a write and a viewer is refused with READ_ONLY_ROLE.

Screenshot pendingRisk classification grouped by tier, each asset card listing the triggers behind its classification/assets/screenshots/governance-risk@0.4.png

How a tier is decided

The page restates the classifier's rule table on screen, always visible, for the person who has to defend a tier in an audit:

Tier	Decided by
Unacceptable	A declared capability that AI Act Art. 5 prohibits — social scoring, untargeted face scraping, emotion inference at work, and the rest of that list
High	A declared Annex III domain and a declared effect on decisions about people. Both halves are required: drafting text in HR is not deciding who gets hired
Limited	Users not informed; an Annex III domain without a decision effect; a decision effect without human oversight — or the observed signals, an unmanaged destination or sensitive data seen in traffic
Minimal	A profile was declared and nothing in it escalates
Unclassified	No profile has been declared

Three properties are load-bearing, and each of them is a decision somebody will eventually challenge:

No profile is Unclassified, never Minimal. An absence of information must not become a compliance claim. An empty profile and no profile behave identically, and an absent boolean is an absence, not a "No".
Observation never goes above Limited. The page carries this as its own callout, The escalation ceiling: High and Unacceptable are legal classifications with duties attached, and a hostname in a log cannot make that claim.
Every tier ships its reasons. A tier with no trigger listed is treated as a defect by the page, not a pass — it prints No reason recorded — a tier without a stated reason is a defect, not a pass.

Note

When observed traffic suggests more than the declaration does, the result is a Declared vs observed mismatch finding for a human to review — never a silent promotion to a higher tier.

Read the estate by tier

Open the page

Select Governance → Risk Classification.

Assets are grouped into sections, riskiest first: Unacceptable, High, Limited, Unclassified, Minimal. A tier with nothing in it says Nothing in this tier rather than disappearing.

Narrow to one tier

Select a tier in the segmented control, or All tiers.

The choice lands in the address bar, so a link to the unacceptable tier is shareable. The count beside the control is the number of matching assets.

Read why an asset is where it is

Read the trigger list under each asset card.

Each trigger states its code, whether it came from the declaration or from observation, and the detail behind it. That pair — declared or observed — is what tells you whether the tier rests on something a person said or on something the traffic showed.

Open the asset

Select the asset's key.

The drawer opens with the declared profile, the evidence, the findings against it and who has been using it.

Warning

This page requests one page of up to 200 assets and groups them in the browser. Above that it shows Showing the riskiest 200 and points at the inventory. Page through the whole estate on Governance inventory with the Risk filter set.

Classify an asset

Start the declaration

Select Register on an unclassified asset, or Update registration on one that already has an owner.

The same dialog is used for both; on an already-owned asset the current profile is the starting point.

State the domain

Set Domain, or leave Not stated.

Annex III domains are marked in the list — employment, education, credit scoring, insurance, law enforcement, migration and borders, justice, critical infrastructure, essential services, biometrics, medical. They are the only route to High, and only alongside a declared decision effect.

State the capabilities

Select everything under Capabilities that the asset actually does.

A single prohibited capability classifies the asset Unacceptable on its own, and the dialog warns before you submit.

Answer the three questions

Set Human oversight, Affects decisions and Users informed to Yes, No or Not stated.

Three states, not a checkbox, because an unanswered question and a declared "No" reach different conclusions. Not stated does not escalate; No on Users informed does.

Submit

Select Register.

The toast names the resulting tier and how many findings closed. If the declaration and the observation disagree it adds Declaration and observation disagree — a mismatch finding was opened for review.

Fields reference

Tier filter

Value	Shows
All tiers	Every asset, grouped by tier
Unacceptable / High / Limited / Minimal / Unclassified	Only that group

Risk profile

Field	Values	Effect on the tier
Domain	The domain vocabulary, or Not stated	An Annex III domain reaches High with a decision effect, Limited without one
Capabilities	Multi-select over the capability vocabulary	A prohibited capability alone reaches Unacceptable
Human oversight	Yes / No / Not stated	No alongside a decision effect reaches Limited
Affects decisions	Yes / No / Not stated	Yes is the second half of High, and reaches Limited without oversight
Users informed	Yes / No / Not stated	No reaches Limited on its own

Triggers

Source	Meaning
Declared	The reason comes from the profile a person filled in
Observed	The reason comes from traffic — an unmanaged destination, or sensitive data seen. Never reaches above Limited

Verify

The asset you classified appears under the tier the toast named, and its card lists at least one trigger.
Filter to that tier. The asset is in the group; filter to Unclassified and it is not.
Open Event logs. The declaration wrote GOVERNANCE_ASSET_REGISTERED, carrying the previous status, the resulting tier, the triggers and the findings it closed — audit category, 365-day retention.
Declare a profile whose observed signals contradict it. A Declared vs observed mismatch finding opens and the tier does not move: the count on Governance overview rises by one.

If it fails

Access denied — your role is not admitted to the governance pages.
A 403 with FEATURE_NOT_LICENSED — the licence does not carry the governance capability.
This is a read-only demo account (viewer role) — actions and changes are disabled. — the declaration was refused with READ_ONLY_ROLE.
No reason recorded — a tier without a stated reason is a defect, not a pass. — the asset carries a tier with no triggers. Re-register it; if it recurs, it is a gateway defect worth reporting.
Nothing in this tier in every group — discovery has not run, or every asset is filtered out.
An asset stays Unclassified after you registered it — you gave it an owner and left the profile empty. Reopen Update registration and declare the profile.
The estate is bigger than the page — Showing the riskiest 200 is displayed. Use the inventory's Risk filter and paging.
An asset you expected in High landed in Limited — one of the two halves is missing. High needs an Annex III domain and Affects decisions set to Yes.

Next steps

Governance inventory — page through the whole estate and retire what is no longer in use.
Ownership — the assets that have no owner, which are almost always the unclassified ones too.
Monitor regulatory compliance — what the controls actually caught, the other half of the picture.

Classify AI assets by risk tier