If your organization uses Microsoft Copilot, Copilot Studio agents, or any third-party AI inside operations that touch people in the EU, the EU AI Act almost certainly applies to you — as a deployer, not a builder. A deployer is any organization that uses an AI system under its own authority, and that is the role most enterprises occupy the moment they turn Copilot on. The Act is also extraterritorial: a U.S. or UK company whose AI outputs affect EU employees or customers is in scope regardless of where it is headquartered.

This guide is written for the IT directors, CISOs, and operations leaders who have to answer a board question that is now overdue: "Is our Copilot deployment regulated, and what are we actually required to do about it?" It maps the deployer obligations, the deadlines that moved in 2026, and the specific points where a Microsoft-stack rollout crosses from unregulated to high-risk — in plain terms, without legal boilerplate.

Key Takeaways

• Most enterprises are deployers, not providers. Using Copilot, Copilot Studio agents, or ChatGPT under your own authority makes you a deployer, which carries its own obligations even though Microsoft carries the provider's.
• Risk is set by use case, not by the tool. The same Copilot instance is minimal-risk for drafting email and high-risk the moment it screens job candidates or influences credit decisions.
• Two deadlines already bind you. AI literacy obligations (Article 4) have applied since 2 February 2025, and transparency obligations (Article 50) apply from 2 August 2026.
• The high-risk deadline moved — but the work did not. The Digital Omnibus deferred Annex III high-risk compliance to 2 December 2027, giving deployers margin, not a reprieve.
• Fines are larger than GDPR. Prohibited-practice violations reach up to €35 million or 7% of global turnover; high-risk and transparency breaches reach up to €15 million or 3%.
• The first deliverable is an inventory. Every credible compliance path starts by cataloguing the AI systems in use and classifying each by risk and by your role.

The short answer: are you a deployer, and does the Act reach you?

Under Article 3 of Regulation (EU) 2024/1689, a provider develops an AI system and places it on the market under its own name; a deployer uses an AI system under its own authority. Microsoft is the provider of Copilot and the Copilot Studio platform. When you configure agents, point them at your data, and put them in front of employees or customers, you are the deployer — and the Act assigns you a distinct set of duties.

The reach is broad. You are in scope if you deploy AI to users in the EU, or if the outputs of your AI system are used in the EU, even when your company sits entirely outside it. A U.S. enterprise running an AI recruitment screen for EU employees is regulated. So is a SaaS company whose LLM-powered feature is used by EU customers. If any of that describes you, the only open question is which risk tier your specific uses fall into.

Deployer vs. provider: which obligations are yours?

The distinction matters because it decides which controls you own and which you verify. Providers carry the heavy build-side duties; deployers carry the use-side duties and the responsibility to confirm the provider did its job.

Provider vs. deployer responsibilities compared

Provider (e.g., Microsoft)

• Owns: Risk management system, data governance and documentation, technical transparency, conformity assessment, and automated logging capability.
• Your action: Verify it through vendor due diligence before and during deployment.

Deployer (e.g., your organization)

• Owns: How the system is used — inventory, human oversight, workplace notification, incident reporting, and (for high-risk uses) fundamental rights impact assessments.
• Your action: Build and evidence these controls yourself; the provider cannot do it for you.

Being a deployer is lighter than being a provider, but it is not a pass. The obligations attach to your configuration choices and your context of use — the two things Microsoft has no visibility into.

When does Microsoft Copilot or an AI agent become "high-risk"?

This is the pivot that most rollouts miss. A general-purpose tool like Copilot starts as minimal risk and becomes high-risk based on how you deploy it. The legal trigger is the purpose and context, not the brand of tool or the sensitivity of the data.

The practical test is whether the AI materially influences a decision about a person in one of the Annex III domains: employment and worker management, access to essential services such as credit and insurance, education and grading, critical infrastructure, law enforcement, or migration. Two illustrative pairs make the line concrete.

• Copilot drafting marketing emails: Not high-risk — no decision about an individual's rights or access.
• Copilot or an agent ranking job candidates: High-risk — employment and worker management under Annex III.
• Copilot summarizing a code review: Not high-risk — internal productivity, no individual impact.
• An agent approving or scoring credit applications: High-risk — access to essential financial services.

The dangerous version of this is what practitioners call shadow high-risk: an HR team wiring a custom-prompted assistant to filter applicants, or a marketing team using AI to segment customers in ways that affect their access to an offer. These are legitimate business tasks that quietly create high-risk deployments no one classified. Without an inventory, you cannot see them — and the Commission's May 2026 draft classification guidelines read Annex III expansively, so more systems qualify than a literal reading suggests.

What are the EU AI Act deadlines that actually matter in 2026?

The Act phases in across several dates, and the Digital Omnibus amendment adopted in mid-2026 moved the high-risk dates. Here is the timeline that governs a Microsoft-stack deployer today.

EU AI Act deadlines for deployers

2 February 2025 — already in force

• What applies: Prohibited practices (Article 5) and AI literacy obligations (Article 4).
• Fine exposure: Up to €35 million or 7% of global annual turnover for prohibited practices.

2 August 2025 — already in force

• What applies: Obligations for providers of general-purpose AI models (GPAI): technical documentation, transparency, copyright.
• Fine exposure: Provider-side; deployers verify via due diligence.

2 August 2026 — now live

• What applies: Transparency obligations (Article 50) — chatbots and assistants must disclose they are AI; AI-generated synthetic content must carry machine-readable labels. Penalty enforcement by national authorities also begins.
• Fine exposure: Up to €15 million or 3% of global annual turnover.

2 December 2026 — watermarking backstop

• What applies: For systems on the market before 2 August 2026, the Omnibus cut the watermarking transition from six months to three, making the effective synthetic-content deadline 2 December 2026.
• Fine exposure: Same transparency tier as above.

2 December 2027 — high-risk (deferred)

• What applies: Annex III use-based high-risk systems (recruitment, credit scoring, education, critical infrastructure, law enforcement, migration). Deferred from August 2026 by the Omnibus.
• Fine exposure: Up to €15 million or 3% of global annual turnover for high-risk obligations.

Two practical notes. First, the deferral is margin, not amnesty — until the Omnibus was formally published, the original dates technically remained in force, and compliance leaders were advised to keep preparing as though the earlier deadline applied. Second, machinery-embedded AI now falls under the Machinery Regulation rather than the AI Act directly.

What must a deployer actually do? The obligation checklist

For any deployment you classify as high-risk, the deployer duties are concrete and evidence-based. A widely used enterprise breakdown lists eight core obligations, and two more — AI literacy and transparency — already bind every deployer regardless of risk tier.

• AI system inventory and classification: Maintain a live catalogue of AI systems with vendor, use case, risk tier, data flows, and integrations — including shadow AI discovered through tooling.
• Data governance: Where you control the input data, ensure it is relevant and sufficiently representative for the system's intended purpose.
• Human oversight and continuous monitoring: Keep real-time oversight of system behavior against its intended purpose, with alerting on anomalies.
• Logging and retention: Keep automatically generated logs for a minimum of six months, longer for financial institutions, capturing usage times, inputs, and who verified the results.
• Workplace notification: Inform employees and their representatives before deploying AI that monitors or evaluates work performance.
• Incident reporting: Establish procedures to notify providers and authorities when a system causes serious harm to health, rights, infrastructure, property, or the environment.
• Fundamental Rights Impact Assessments (FRIAs): Conduct a FRIA before deploying high-risk AI, mandatory for public bodies and for uses such as credit scoring and insurance risk assessment.
• Vendor due diligence: Verify the provider's compliance status, assess their controls, and negotiate SLAs for AI incidents.

Layered on top, two duties apply broadly. AI literacy (Article 4) has been binding since February 2025 and requires that staff using AI understand the tool, its limits, and when human oversight is needed. Transparency (Article 50) requires that people are told when they are interacting with AI and that synthetic content is labeled.

The deadline moved, but the obligation to know what AI is running in your business, and why, did not.

Most of this is governance work, not model work — and it is exactly the layer that breaks when Copilot is switched on across a tenant without an operating model underneath it. BabyBots runs fixed-fee AI governance and Copilot readiness assessments that inventory every AI system in your Microsoft estate, classify each by EU AI Act risk tier, and hand back a prioritized remediation plan — so the obligations above become a tracked program rather than a board-meeting surprise.

What this means specifically for a Microsoft-stack deployment

Microsoft, as the provider, has committed to EU AI Act compliance and publishes guidance on the provider-side controls it carries. That protects you on the build side of Copilot and the platform, but it does not classify your use cases, notify your workers, or run your FRIAs. Those remain yours.

The highest-leverage move for a Microsoft estate is to treat classification as a tenant-wide exercise. Copilot Studio agents are the most likely to drift into high-risk territory because they are purpose-built for specific workflows — an agent that screens applicants, triages benefits eligibility, or scores customers is a different regulatory object from a Copilot that summarizes meetings. Pair the AI inventory with your existing data-governance and shadow-AI controls so that classification, oversight, and logging ride on infrastructure you already operate rather than a parallel compliance stack. For teams standing up agents at scale, the same discipline that keeps autonomous agents governed in production is what makes EU AI Act evidence a byproduct of good operations instead of a separate project.

Frequently Asked Questions

Does the EU AI Act apply to a US company using Microsoft Copilot?

Yes, if the AI's outputs are used in the EU or you deploy it to EU users. The Act is extraterritorial, so a US-headquartered enterprise that runs Copilot for EU employees or customers is a deployer in scope, regardless of where its servers or headquarters sit. Location does not exempt you; the location of the people affected by the AI is what matters.

Is Microsoft Copilot a high-risk AI system?

Not by default. Copilot is a general-purpose tool that starts at minimal or limited risk, and it becomes high-risk only when you deploy it for an Annex III purpose such as screening job candidates, scoring credit, or making decisions about access to essential services. The classification follows your use case, not the product name.

What is the difference between a provider and a deployer under the EU AI Act?

A provider develops an AI system and places it on the market under its own name; Microsoft is the provider of Copilot. A deployer uses an AI system under its own authority; your organization is the deployer. Providers carry build-side duties like conformity assessment, while deployers carry use-side duties like human oversight, workplace notification, and fundamental rights impact assessments.

When is the EU AI Act deadline for high-risk AI systems?

Following the Digital Omnibus amendment in 2026, Annex III use-based high-risk systems must comply by 2 December 2027, and Annex I product-embedded systems by 2 August 2028. These were deferred from the original 2026 dates. Transparency obligations under Article 50, however, already apply from 2 August 2026.

What are the fines for EU AI Act non-compliance?

Violations of prohibited-practice rules carry fines of up to €35 million or 7% of global annual turnover, whichever is higher. Breaches of high-risk and transparency obligations carry fines of up to €15 million or 3% of global annual turnover. These caps are larger than GDPR's, which is why compliance leaders are treating AI classification as a board-level risk.

What is the first step toward EU AI Act compliance?

Build an inventory. Before any legal assessment, catalogue every AI system in use — including embedded AI in third-party software and shadow AI adopted by teams — and record your role and the risk tier for each. An accurate inventory is the prerequisite for classification, prioritization, and every obligation that follows.

Where this is heading

The Omnibus deferral bought deployers time, but the direction is fixed: AI governance is becoming an operational capability that runs continuously inside the platform, not a policy document reviewed once a year. The enterprises that treat the 2027 high-risk deadline as a planning horizon — inventorying now, classifying Copilot and agent use cases, and wiring oversight into the tenant they already run — will absorb each future obligation as a small increment. The ones that wait will discover their shadow high-risk deployments the hard way, during an incident or an audit.

Map your Copilot estate against the EU AI Act

If you are deploying Microsoft Copilot or Copilot Studio agents and cannot yet name which uses are high-risk, that is the gap to close first. Book a BabyBots EU AI Act readiness assessment — in a single working session we inventory the AI running across your Microsoft estate, classify each use by risk tier, and hand back a prioritized obligation-by-obligation plan mapped to the deadlines that apply to you. Governance-first, built to last, and owned by your team when we leave.

Sources

• NeuralTrust — EU AI Act Enterprise Compliance Guide 2026 (June 2026)
• Sentra — EU AI Act Compliance: What Enterprise AI Deployers Must Know (February 2026)
• Gruppo RES — AI Act and AI Governance: Deadlines and Changes from August 2, 2026 (July 2026)
• Regulation AI — EU AI Act Omnibus 2026: New Compliance Deadlines Explained
• Microsoft Trust Center — EU AI Act Compliance