EU AI Act compliance Microsoft 365
The AI Act asks fifteen things of you. Why are only four of them documents?
An organization running Copilot, Copilot Studio agents and OAuth-consented AI apps is a deployer in AI Act terms. Across the eleven frameworks 1Security evaluates, the obligations collapse into 15 shared requirements - and exactly four of them are documents a human writes: the AI impact assessment, the usage policy, the incident procedure, the risk documentation. The other eleven are measured live from the tenant: the agent inventory with accountable owners, captured agent instructions, restricted data kept unreachable, logs held past the six-month floor. Attest the four, watch the eleven, export the dated evidence pack.
The scenario
The obligations arrived before the tooling did.
The AI Act has been arriving in phases: prohibited practices and the AI literacy duty since February 2025, general-purpose AI obligations since August 2025, and the high-risk regime now set for December 2027 after the Digital Omnibus moved the wall. That date is not a reprieve - it is the window in which every deployer stands up AI governance that actually operates, with penalties scaling to 35 million euro or 7% of worldwide turnover for the prohibited end and 15 million or 3% for most of the rest.
The obligations read like a security backlog, not a legal one. Who oversees each AI system, and are they competent and authorized? What data can each system reach, and is that controlled? Are the logs kept, and for how long? A policy document answers none of it; a tenant, read properly, answers all of it.
This is why the Compliance screen treats the EU AI Act as one of eleven frameworks scored from a single set of 15 requirements. Eleven of those requirements are measured continuously from tenant data, and the four that are attested are precisely the AI documents the Act expects. Each article takes the worst status of the requirements it cites, and "not evaluated" is never counted as a pass - the arithmetic is deliberately conservative, because the number ends up in front of an auditor. Fix agent ownership once and the EU AI Act, the Polish KRiBSI act and ISO/IEC 42001 all move in the same pass.
The walkthrough
Four obligations, answered from the tenant.
Each beat maps to an article family - and each is a screen, not a binder.
- 01
Inventory every AI system
The Agents screen holds every agent from Copilot, Copilot Studio, Azure AI Foundry, Entra Agent ID and third-party vendors - with origin, blueprint, reach and activity. On the Compliance screen this is a measured requirement, not a claim: instructions captured per agent, and shadow agents surfaced precisely because a registry of the registered cannot show them.
- 02
Assign competent oversight
Every agent carries an accountable person - a native owner or sponsor, or one assigned in 1Security - and the requirement panel shows the exact progress, in the shape of "12 of 15 agents have an accountable person". That is Article 26(2), oversight by natural persons with competence and authority, as a measured verdict instead of an org-chart annex. Some agents covered reads At risk; none reads Not met.
- 03
Control the data exposure
Declare the sensitive types your AI must not touch - no-AI-access on payment cards, health data, payroll - and the "restricted data types stay unreachable" requirement measures every agent, app and user still reaching them, with the declaration recorded. No restrictions declared reads Not evaluated, never Met: data governance as a control with a count, not a paragraph.
- 04
Keep the logs, export the evidence
Log retention is its own measured requirement against the 183-day floor, with three years of history on standard licenses in practice. The whole status document is snapshotted every Monday and on demand, and the framework view exports a per-framework evidence pack - the same document the read-only API serves - so "what was our status on this date" has a real answer.
What makes it work
Three parts of the platform behind the answers.
The same 15 requirements score all eleven frameworks - NIS2, GDPR, DORA, both Polish acts, SOC 2, NIST CSF 2.0, HIPAA and both ISO standards alongside the AI Act - so one fix is counted everywhere it applies.
Compliance monitoring
Readiness per framework measured from live tenant data, a deduplicated worklist ranked by articles unlocked, Monday snapshots and evidence packs.
Explore the feature →AI agent inventory
Every agent across every ecosystem with owners, blueprints and reach - the population every AI Act answer starts from.
Explore the feature →AI data restriction
The data-governance control: declared restrictions with blast-radius previews, violation lists and recorded history.
Explore the feature →
FAQ
Common questions.
Are we a provider or a deployer under the AI Act?
An organization that runs Copilot, builds Copilot Studio agents for its own use, or consents third-party AI apps into its tenant is a deployer for those systems. Deployer obligations - oversight, data control, log retention, transparency - are exactly the ones this page describes. Building and selling an AI system makes you a provider, which is a different and heavier set.
Is Copilot a high-risk AI system?
In itself, almost never - risk class follows usage, not branding. Drafting emails and summarizing meetings sits outside Annex III entirely. Feed the same assistant into hiring decisions, worker evaluation or credit scoring and you have deployed a high-risk system with the full duty set attached. The honest question is not "is Copilot high-risk" but "which of our uses are" - and that starts with the inventory.
Which deadlines actually matter?
The Act applies in phases: prohibited practices and AI literacy since February 2025, general-purpose AI obligations since August 2025, and the high-risk regime of Annex III from December 2, 2027, after the Digital Omnibus deferral. The practical reading: the oversight, logging and data-governance groundwork is already due, and the 2027 date is the window to make it operational rather than documentational.
What are the four documents we still have to write?
An AI impact assessment, an AI usage policy and register, an incident response and reporting procedure, and risk management documentation. They are the only manual requirements of the fifteen - the compliance officer records each attestation once, with owner, review date and evidence link, and it is counted in every framework that cites it. The other eleven requirements are measured; the data decides, not a person.
Does this replace legal counsel?
No. Scoping decisions - which uses fall under Annex III, whether an impact assessment is required for a given deployment - belong with counsel. What 1Security replaces is the evidence gathering underneath those decisions: the inventory, the oversight status, the exposure lists and the retention attestation that counsel and auditors otherwise request by spreadsheet.
Do we need E5 or Copilot licenses for this?
No. The inventory, oversight assignments, restrictions, log retention and the evidence pack run on the standard read-only connection and the licenses you already own. Where a specific Microsoft control needs its own license, the product says so on the row rather than assuming it.
Walk into the AI Act audit with the tenant on your side.
Connect read-only and the inventory, oversight gaps and data-exposure answers land the same day. Attest four documents, and the evidence pack builds itself from there.
Or answer the deployer questions from a policy binder.