The Copilot Compliance Gap:
Why Professional Services Firms
Need an AI Audit Before It Bites

The AI compliance industry is busy auditing the wrong thing. Vendors and consultancies are queuing up to assess "high-risk AI systems" — credit scoring, biometric tools, HR screening algorithms. Meanwhile, the AI that actually touches your client work every day arrived through a Microsoft 365 licence renewal, switched on across the whole firm, with no audit framework behind it at all.

Microsoft 365 Copilot now sits inside Word, Excel, Outlook and Teams at thousands of law firms, accountancy practices and consultancies. It drafts client letters. It summarises matter files. It builds the first cut of advice. And in most mid-sized firms, nobody has asked the questions a regulator — or a claimant's solicitor — will eventually ask.

Why Copilot Slips Through the Compliance Net

Copilot is not classified as a high-risk AI system under regimes like the EU AI Act. It is a general-purpose productivity tool, and it inherits Microsoft's enterprise security posture, which gives IT teams a comfortable answer: "it's covered by our Microsoft tenancy."

That answer covers infrastructure. It does not cover what your people do with the output. The compliance exposure in professional services does not come from the model — it comes from AI-generated content entering advice, accounts and filings that carry your firm's professional duty of care. Three gaps appear in almost every rollout we see:

• No output verification standard. Who checks AI-drafted client advice, and against what standard? "A fee earner reads it" is not a standard — it's a hope. When Copilot summarises a 200-page disclosure bundle and misses the limitation date, the negligence claim doesn't name Microsoft.
• No data governance mapping. Copilot retrieves from everything the user can access. Years of over-permissive SharePoint and shared-drive sprawl mean it can surface one client's confidential material while drafting for another. Most firms have never tested what their permission model actually exposes.
• No usage records. If a piece of advice is challenged two years from now, can you show whether AI contributed to it, who reviewed it, and what they changed? Professional indemnity insurers are starting to ask exactly this question at renewal.

The Professional Negligence Angle Nobody Is Pricing In

For regulated professionals, the duty of care is non-delegable. A solicitor who relies on a Copilot summary, or an accountant who files figures from an AI-assisted reconciliation, remains fully responsible for the result. Courts in several jurisdictions have already sanctioned lawyers for filing AI-generated material containing fabricated citations — and in each case the professional, not the tool, carried the consequences.

In the UK, the SRA and ICO are actively building joint guidance and sandbox initiatives for AI use in regulated firms, and the ICO is now using its own AI auditing tools to test systems for bias and transparency. The direction of travel is clear: "we didn't know the AI was wrong" will not be a defence, and "we had no framework for checking" will make it worse.

What a Copilot Compliance Audit Actually Covers

The good news: auditing a Copilot deployment is a bounded, practical exercise — far smaller than the conformity assessments being built for genuinely high-risk AI. A proper audit for a mid-sized professional services firm covers five areas:

1. Access and data exposure review. Test what Copilot can actually retrieve under real user permissions. Identify confidential client material reachable across matter or engagement boundaries, and fix the permission model before the tool finds it.
2. Output verification policy. Define which categories of work product may use AI assistance, what review each category requires, and who signs it off. Client advice, court filings and statutory accounts sit at the top of the verification ladder.
3. Usage logging and audit trail. Configure the tenancy so AI involvement in client work is recorded and retrievable. If you cannot reconstruct it, you cannot defend it.
4. Staff competence and AI literacy. Document training on what Copilot is reliable for, where it fails, and how to verify output. For firms with EU exposure, this also maps directly onto the EU AI Act's AI literacy obligations under the general framework applying from 2 August 2026.
5. Client transparency position. Decide — deliberately, in writing — what you tell clients about AI use in their matters, and align engagement letters with it. Silence is a position too, just not one you chose.

The Questions to Ask This Week

You don't need a six-month programme to start. Ask these four questions of your practice and see how many produce a confident answer:

• Can Copilot, under any fee earner's login, retrieve material from a client they don't act for?
• Is there a written standard for reviewing AI-assisted client work before it goes out?
• Could you tell a regulator, insurer or court whether AI contributed to a specific piece of advice?
• Has anyone outside IT — risk, compliance, or a partner — formally signed off the Copilot deployment?

Two or more "no" answers means your firm is carrying unaudited AI risk through every client engagement. That is fixable — and fixing it now, while regulators are still in guidance-and-sandbox mode, is dramatically cheaper than fixing it after a claim.