The EU AI Act put a date on the calendar that internal audit functions could not ignore, and then, in 2026, moved it. That combination — a hard obligation with a shifting timeline — is exactly the situation where internal audit earns its keep: not by predicting the final regulatory text, but by getting the organization ready for the version that is coming regardless.
This is a practitioner's read on what the AI Act deadline means for internal audit, what actually changed in 2026, and how to build an AI-audit approach you can run now — including the awkward part nobody wants to talk about, which is how you audit your own AI-assisted audit work.
What the Deadline Actually Is (and What Moved)
The EU AI Act entered into force in 2024 on a staggered timeline. The prohibited-practices and AI-literacy provisions applied first, in early 2025. Obligations for general-purpose AI models followed in August 2025. The date most internal audit functions circled was August 2, 2026 — when the bulk of the high-risk system obligations (Annex III use cases such as employment, credit, and access to essential services) were scheduled to apply.
Then the timeline moved. In 2026 the European Commission advanced a simplification package that, as reported through its passage, defers most of the high-risk obligations — with the widely cited new application date landing in December 2027. Confirm the final effective dates against the adopted text before you rely on them in a plan; simplification packages change in negotiation.
Here is the trap. "The deadline moved" reads to a lot of executives as "we have more time, deprioritize it." That is the wrong lesson for two reasons. First, the prohibited-use and transparency obligations already apply — the deferral is about the high-risk tier, not the whole Act. Second, the work that makes an organization compliant by late 2027 — inventorying AI systems, risk-tiering them, standing up controls and documentation — takes quarters, not weeks. A function that treats the deferral as a reprieve rather than runway will be doing the same scramble in 2027 that it avoided in 2026.
Why This Lands on Internal Audit's Desk
Two data points frame the pressure. Gartner has reported that 97% of chief audit executives rank regulatory compliance among their top priorities for 2026 — and AI regulation is the fastest-moving piece of that compliance picture. KPMG's audit-committee research has found that a large majority — around 88% of audit committees — cite AI, cybersecurity, and supply-chain risk as areas demanding more attention. The board is asking. The regulator is signaling. The question lands on internal audit because it is the function positioned to give independent assurance over how the organization actually governs its AI — not how a vendor deck says it does.
Meanwhile the audit profession's own regulators are moving in the same direction. The PCAOB has been emphasizing professional skepticism and audit-documentation quality in a world where AI is increasingly in the workflow — the through-line being that AI in the process raises, not lowers, the bar for evidence and for showing your work. That is a useful frame for internal audit too: AI does not relax documentation standards; it makes them the whole game.
A Practical AI-Audit Approach You Can Start Now
You do not need the final regulatory text to start. The core of an AI-audit / AI-governance approach is stable regardless of which deadline sticks. Four moves.
1. Inventory the AI
You cannot audit what you have not inventoried — the same first principle as any audit universe. Build a register of where AI is actually used across the organization: procured tools with AI features, internally built models, embedded AI in SaaS platforms, and shadow AI (the marketing team's writing tool, the analyst's copilot). Shadow AI is where the real exposure hides, because it is unmanaged by definition.
2. Risk-Tier Each Use
Not every AI use is high-risk. Map each system to a risk tier using the Act's own logic as a starting rubric: prohibited, high-risk (Annex III use cases — employment, credit, essential services, biometrics), limited-risk (transparency obligations), and minimal-risk. The tiering drives everything downstream: a chatbot that suggests knowledge-base articles is not the same risk as a model that screens job applicants. Document the tiering rationale — that document is itself an audit artifact.
3. Assess Controls Over the AI
For the systems that matter, evaluate controls across the dimensions regulators and standards keep converging on:
- Data — provenance, quality, bias testing, and a lawful basis for the training and input data.
- Model — validation, performance monitoring, drift detection, and change management over model versions.
- Human oversight — is there a meaningful human in the loop, with the authority and information to override? "Human on the loop" that rubber-stamps is not oversight.
- Documentation — technical documentation, intended purpose, known limitations, and an audit trail of decisions. This is the deliverable the high-risk tier is built around, and the one most organizations are weakest on.
4. Audit the AI-Assisted Work — Including Your Own
This is the part that gets skipped. If your organization uses AI to make or support decisions, those decision trails are auditable. And if internal audit itself uses AI — to draft risk assessments, review evidence, summarize interviews — then your own workpapers are in scope for the same scrutiny. The standard to hold yourself to is the one you would hold an auditee to: an evidence trail showing what the AI produced, a human sign-off from someone accountable, and source citations so a reviewer can trace a conclusion back to the underlying evidence rather than to an unexplained model output. We wrote about staying inside the standards while doing this in how to use AI in audit without failing your next QAR, and about the deeper defensibility posture in the defensibility ledger for agentic AI in audit.
A Worked Example: Tiering One System
Take a single use — an AI model that ranks inbound job applicants — and run it through the four moves.
| Step | Applied to the applicant-ranking model |
|---|---|
| Inventory | Embedded in the ATS; owned by Talent Acquisition; vendor-supplied model retrained quarterly on the company's hiring data |
| Risk tier | High-risk — employment and worker management is an Annex III use case |
| Controls | Data: bias testing on protected classes, lawful basis for training data. Model: version change log, performance monitoring. Human oversight: a recruiter reviews the ranking and can override, with the override logged. Documentation: intended purpose, known limitations, and a per-candidate decision trail |
| Audit the work | Test that overrides actually happen (not rubber-stamped), that the vendor's documentation matches reality, and that rejected-candidate decisions are traceable |
The value of the walk-through is that it turns "we should govern AI" into a specific, testable list. Every cell in that table is a control you can evaluate and a document you can request.
Three Failure Modes to Avoid
- Treating the deferral as a reprieve. The high-risk date moved; the prohibited-use and transparency obligations did not, and the readiness work takes quarters. "We have time" is how 2027 becomes another scramble.
- Auditing the policy, not the practice. A governance policy on paper is not assurance. The exposure lives in shadow AI and in human-oversight controls that look real but rubber-stamp. Test the practice, not the intent.
- Exempting your own AI use. If internal audit uses AI and holds auditees to an evidence-trail standard it does not meet itself, the independence and skepticism argument collapses. Audit your own workpapers to the same bar.
What "Good" Looks Like by Late 2027
Work backward from the obligation. An organization in reasonable shape has: a maintained AI inventory; a risk-tiering applied and refreshed; assigned ownership for each high-risk system; documentation packages for high-risk uses; a human-oversight design that is real, not nominal; and an audit trail that lets someone reconstruct how an AI-influenced decision was made. Internal audit's role is to provide independent assurance that those things exist and operate — and to have flagged the gaps early enough to fix them.
The functions that will be calm in late 2027 are the ones treating the deferred deadline as the runway it is. The ones that will be scrambling are the ones that heard "it moved" and closed the file.
How Audvera Supports This
Audvera is not an EU AI Act compliance product, and it does not maintain an AI-system register for you — that governance work belongs with your risk, legal, and compliance functions. What Audvera does address is the posture the profession's regulators are increasingly signaling for AI-assisted work: evidence review that is source-cited and human-gated, with an immutable trail of who ran an AI assistant and who reviewed and signed off on its output. When internal audit uses AI in its own engagements, that is exactly the documentation standard the profession is converging on — the ability to show the work, not just the answer.
If you want to see what defensible, human-reviewed AI-assisted audit work looks like in practice, start with a free risk assessment and follow the trail from an AI-drafted risk through to a reviewer sign-off.