EU AI Act
Register the AI systems your organisation builds or uses, record each one's risk level, and work through the obligations that follow — documentation, conformity, transparency and registration.
The EU AI Act (Regulation 2024/1689, a European law that regulates AI systems according to how much risk they pose) sets different duties for different systems. A spam filter is treated lightly; a system that screens job applicants or scores creditworthiness is high-risk and carries documentation, oversight and registration duties. This module holds one inventory of those systems and the flows that produce the paperwork each risk level demands. It does not classify a system for you, and it never makes a system lawful on its own.
The EU AI Act is a pay-per-framework module. Without it in your licence the
module home shows a short "not available on your current plan" message
instead of the dashboard, and every sub-page — the inventory, a system page,
the wizards, the prohibited practices screening and the two signpost pages —
shows an EU AI Act Not Available notice in place, asking you to
contact your Aegis administrator. Nothing redirects you elsewhere, so you
can see why the page is empty.
A licensed module can also be switched off in
Settings → Frameworks. Every page then reads
EU AI Act is switched off, with the reassurance that its data
is kept and returns when an administrator switches it back on.
Administrators see an Open Settings → Frameworks link on that
notice; other roles do not.
Who uses it
Everyone who reaches the Compliance area can read the register. Changing it needs a higher role, and controls you may not use are hidden rather than shown and refused.
-
Viewer reads the module home,
the inventory, each system page and the prohibited practices screening. No
Register AI Systembutton appears, theAI Inventory Readinessbutton is hidden (a Viewer cannot commission AI actions), the edit, classify and delete controls on a system page stay hidden, and so do the links to the Annex IV wizard and the conformity attestation. - Contributor adds registering, editing and risk-classifying a system, runs the readiness assessment, and opens the Annex IV wizard and the conformity attestation. A Contributor cannot delete a system: removing an entry from a regulatory register is kept for Managers and above.
- Manager adds deleting a system and the right to complete a FRIA (Fundamental Rights Impact Assessment — a structured review of how a system might affect people's rights). A Manager records a conformity assessment but does not issue the Declaration of Conformity.
- Admin has full access across the module, including exporting the Declaration of Conformity.
What's on this screen
Open Compliance in the left sidebar and choose
EU AI Act. You land on the module home at
/compliance/eu-ai-act, headed
EU AI Act Compliance with the line "Manage compliance with EU AI
Act (Regulation 2024/1689) requirements for AI system risk classification." To
its right sits the one action button on this screen,
AI Inventory Readiness. Below the header the page stacks four
bands:
-
Five count cards.
Total Systems,High Risk,Conformity Passed,FRIA RequiredandReassessment Required. On the captured screen they read 13, 4, 1, 5 and 1. All but the first colour themselves once they rise above zero — amber forHigh Risk, green forConformity Passed, red forFRIA Requiredand amber forReassessment Required;Total Systemsstays neutral whatever it counts.Reassessment Requiredcounts systems whose classification was invalidated by a material change after it was recorded (see "Work through a system's obligations" below). Such a system keeps its recorded risk level, so it is never folded intoHigh Risk. -
Four navigation cards.
AI Systems("Register and manage AI systems subject to EU AI Act requirements"),Technical Documentation("Annex IV technical documentation guidance — files are completed per AI system"),Transparency Obligations("Article 50 transparency obligations — notices are created per AI system") andProhibited Practices Check.Technical DocumentationandTransparency Obligationsopen with aBack to EU AI Actlink;Prohibited Practices Checkcarries a plainBacklink to the same place. TheAI Systemsinventory has no back link — use the sidebar or your browser's back control to return here. - Saved AI insights. Keeps readiness narratives you chose to store. Empty on the captured screen: "No AI insights saved yet. Run an AI action and choose "Save as record" to keep it here."
- Action items. The follow-up list, likewise empty there: "No action items yet. Use "Create action item" inside an AI action's result to track one here."
There is no filter bar and no table here — this screen is a hub, and the working
lists live one level down. Its three live areas are numbered on the figure
below: the readiness button
1, the saved-insights panel 2 and the action-items
panel 3.
If Aegis cannot read the counts, it does not show zeros. An amber banner reads
"Could not load your EU AI Act overview counts. This is a temporary problem —
your registered systems are safe", each card reads Unknown, and a
Try again button reloads the counts without leaving the page.
The AI assist
AI Inventory Readiness reads your registered systems and narrates
where you stand. The scoring is done by Aegis, not the model: the counts and the
severity-ranked gap list — each gap tied to an article, such as Article 6 for
classification or Article 43 for conformity — are worked out first. The model
writes them up and proposes three to five next steps. It never recomputes a
score, reorders the gaps, or invents a system or an article. You decide what to
do with it.
-
Select
AI Inventory Readinessat the top right. A dialog titledEU AI Act Inventory Readinessopens explaining what the assessment covers, with aRun Inventory Readinessbutton. Choose it and the answer streams in while four progress lines tick past — reading the inventory, checking classification, conformity, FRIA and registration, scoring the gaps, then drafting the read. When it finishes, a confidence indicator and aVerified figurespanel appear beneath the text, alongsideCopy,Export DOCXandExport PDF, and the start button becomesRun again. -
Select
Save as recordto keep the narrative. Close the dialog and theSaved AI insightspanel lists it with the date it was saved, replacing the empty-state line. -
Select
Create action itemto turn a recommendation into tracked work. Give it a title and description and confirm; it then appears inAction itemsat the foot of this page and in the main action items list.
If Total Systems reads 0, there is nothing to
score and the read will say so. Register your systems first, then come back.
Each run consumes one AI credit from your tenant's monthly pool.
Register an AI system
The AI Systems card opens
/compliance/eu-ai-act/systems. Its table has columns for
Name, Provider, Risk Level,
Status, Domain, Conformity and
Actions; the only row action is a Details link,
because editing happens on the system page. A system whose classification no
longer holds carries an amber Reassessment required badge beside
its risk level; hover over it to see which fields changed. The table shows 25
systems per page, with a total count and Back/Next
paging controls below once you exceed that. Before anything is registered you
see an empty state — "No AI systems registered yet. Register your first AI
system to begin compliance tracking" — with a
Register AI System button beneath it.
-
Select
Register AI System— top right of the inventory, or in the middle of the empty state. The form opens at/compliance/eu-ai-act/systems/new, with aBack to AI systemslink above its title. -
Type a
Name. It is the only required field, and the form will not submit without it ("Name is required"). A recognisable name such as "Resume Screening Assistant" is what colleagues search for later. -
Set
Operator role— your organisation's relationship to the system:Provider,Deployer,Importer,DistributororAuthorized Representative. Duties differ by role, so this drives what you owe. -
Choose a
Risk level:Not Classified,Unacceptable,High Risk,Limited RiskorMinimal Risk. Leave it atNot Classifieduntil the decision is taken, then classify it from the system page, where the reasoning is recorded with it. -
Set the operational
Status(Development,Testing,Production,SuspendedorDecommissioned) and fillProvider,Intended purposeandDescriptionif you know them. -
Select
Register system. The button reads "Registering…" while it saves, then Aegis opens the new system's detail page, where the rest of the record and the obligation flows live. If the server rejects a field, the problem is shown beside that field rather than as a general error.Cancelreturns you to the inventory without saving.
Work through a system's obligations
The detail page at /compliance/eu-ai-act/systems/[id] is the
workbench for one system. An AI Systems link at the top returns you
to the inventory. The header shows the name with coloured risk-level and status
badges, and the provider and version beneath it. Under it is an action bar —
Edit system, Classify risk and
Delete system — each opening a dialog and each hidden from roles
lacking the matching right. Classify risk asks for a risk level, a
risk category (such as Employment & Worker Management) and a
written justification, so the decision carries its reasoning. For
High Risk or Unacceptable both the category and the
justification are required: the dialog says so, and its
Classify button stays disabled until both are filled.
Below that, an overview grid lays out the purpose and description, the operator
role, conformity status including
CE marking (the mark declaring a product
meets EU rules) and the EU database identifier, deployment details, and a side
column with the classification, FRIA status and key dates. A system marked as
general-purpose AI adds a panel for the model name and systemic risk; one
flagged for emotion recognition, biometric categorisation or deepfakes shows
those flags in an amber Special Capabilities
panel. If the system is high-risk and unregistered, an amber banner says the EU
database filing is outstanding.
The classification card in the side column carries three signals worth knowing:
-
Classification not substantiated— an amber badge beside a High or Unacceptable level that lacks its category or written reasoning. It is not a finished classification until both are recorded. -
Reassessment required— when a material field changes after classification (such as the intended purpose, domain, provider, operator role or a special capability), Aegis keeps the recorded level but marks it as no longer describing the system (Article 25). The box names the changed fields and the date it was flagged. Re-classify the system to clear it. -
Classification history— a collapsible list, loaded when you open it, of every classification and every material change, each dated and showing what it was and what it became. It is an append-only trail; with nothing recorded it reads "No classification changes recorded".
Then come the tools, each a step towards an obligation:
-
Annex IV Technical Documentation— a guided wizard through the Annex IV sections (the technical file high-risk providers must keep). Each takes content and a completion status ofMissing,Draft,In review,CompleteorOutdated. A section can only be saved asCompletewhen it has content or at least one linked evidence item, and a completed section shows who reviewed it and when. TheReview & Exportstep counts the completed sections and offersExport as PDF; while any section has unsaved changes, it names them and asks you to save first so the PDF does not leave them out. -
Conformity Self-Attestation— records your Article 43 assessment withRecord attestationand tells you which route applies: Annex III point 1 systems (biometric or law-enforcement) need a named notified body, the rest self-assess. Outcomes runNot startedthroughPassed,FailedorExpired. APassedoutcome can only be recorded once the system's risk level is classified. Administrators additionally seeExport Declaration of Conformity: issuing the declaration is your organisation's own formal attestation under Article 47, so that control is reserved for Administrators (or a user explicitly granted Issue Declaration of Conformity through a permission group). Managers and Contributors record the assessment but are not shown an export button. The declaration can only be exported for a classified system with aPassedassessment that is still in date and whose Annex IV technical documentation is complete: every required section (all except Accessibility) must have its entries set toComplete. While any required section is missing, or has an entry inMissing,Draft,In revieworOutdated, the export button stays disabled, the incomplete sections are listed under it, and a link takes you to the documentation wizard to finish them. Downgrading a completed attestation asks you to confirm. - Conformity checklist and prohibited practices check — inline panels showing where this system stands on each.
-
FRIA wizard — shown only when Aegis derives that Article 27
requires one (the side column then reads
FRIA Required; otherwiseFRIA Not Required). It scores criteria for severity and likelihood on your default risk methodology's scale, then reportsLow Impact,Medium Impact,High ImpactorImpact Mitigatedwith recommendations. -
EU Database Registration— tracks the Article 71 filing throughDraft,Ready to file,FiledandAccepted, and exports an Annex VIII dossier. Filing is a manual submission to the European Commission — there is no API — so you record the identifier here once issued. - Transparency notice editor — drafts the Article 50 disclosure telling people they are dealing with an AI system, with notice types for chatbots, emotion recognition, biometric categorisation, deepfakes and generated content. Publishing a material revision bumps the version and re-prompts everyone who acknowledged the last one.
Screen for prohibited practices
Prohibited Practices Check opens
/compliance/eu-ai-act/prohibited, headed
Prohibited Practices Screening — a cross-system view of the
Article 5 bans, the practices the Act forbids
outright. A banner at the top reads green ("All n systems passed
prohibited practices screening") or red with a count when systems are flagged.
The table below lists System Name, Risk Level,
Emotion Recognition, Biometric Categorization,
Deepfake Capable and Status. A green tick means the
practice is not flagged; a red Flagged warning means it is and
needs review, and the row is tinted. Selecting a system name opens its detail
page. With nothing registered you get an empty state and a
Go to AI System Inventory link.
The screening surfaces the flags recorded against each system. It is not an independent detector — a system nobody has marked as emotion-recognising reads as clear here. The table lists up to the first 100 systems; the flagged count in the banner covers the whole inventory.
Tips and limits
Two of the four navigation cards lead to signposts rather than workspaces:
| Card | What you get | State |
|---|---|---|
AI Systems |
Inventory, registration and the system workbench | Built and in daily use. |
Prohibited Practices Check |
Cross-system Article 5 screening table | Built and in daily use. |
Technical Documentation |
A signpost page explaining Annex IV, listing what you can do per
system (the guided wizard, PDF export, retained revisions), with a
Go to AI system inventory button
|
No aggregate view; the per-system wizard and PDF export are built. |
Transparency Obligations |
A signpost page listing what is still to come — reusable notice
templates, Article 50 conformity tracking, deepfake labelling rules,
an affected-user communication audit trail — and a
Go to AI Systems inventory link
|
Labelled "Coming soon" there; the per-system notice editor is built. |
Aegis stores the risk level, conformity outcome and FRIA result that
you enter. A Passed conformity status is your own
attestation, not a legal certification, and no status here makes a system
lawful. These fields are your organisation's record of decisions taken by
people.
Deleting a system warns you first that its documentation items, transparency
notices and assessments will no longer be reachable. And if the inventory fails
to load it shows a red error — "Could not load your AI systems. This is a
temporary problem — your registered systems are safe" — with a
Try again button rather than an empty table, so a passing fault
never reads as though your records had vanished.
Where this connects
- Compliance frameworks — the EU AI Act also appears as a framework alongside CyFun, GDPR, NIS2 and DORA, mapped to controls.
- Risks — record and treat what a high-risk AI system could cause.
- Vendors — when the system comes from a supplier, keep that supplier's assessment there.
- Evidence and Audit readiness — attach the proof behind each obligation and gather it when an auditor asks.
- Action items — where readiness follow-ups are tracked to completion.
- AI dashboard — the credits your readiness runs consume, and every AI action taken across Aegis.