Shadow AI Audit: 23 Unsanctioned Accounts, One Policy, Zero Leaks

The situation

A 45-person accountancy practice bought two AI licences for the tax team and assumed that was their AI estate. The managing partner booked a Reality Check for a different reason entirely — the licences weren’t being used and he wanted to know why.

What the Reality Check found

The licences weren't being used because the staff had already solved the problem their own way. Week one's inventory found 25 AI touchpoints, not 2: twenty-three personal ChatGPT and Gemini accounts, a free browser extension that summarised email, and the two paid seats — one opened twice since purchase.

Three of the personal accounts had been used to draft client correspondence containing names, figures and tax positions. Nothing malicious — people hitting deadlines with the best tool they could reach. But client data was sitting in consumer accounts, outside the firm's control, outside its engagement letters, and outside UK GDPR defensibility.

The verdict

Kill the unused premium tier (it duplicated what the firm's practice suite already did). Fix everything else: the problem wasn't too much AI, it was AI without a perimeter.

The fix (Recovery Sprint, 3 weeks)

One sanctioned platform with enterprise data controls, chosen against the workflows people were actually doing in their personal accounts. A one-page AI use policy — what's allowed, what's never allowed, where to ask. Access through firm accounts only, personal-account use ended with an amnesty week rather than a witch-hunt. Ninety-minute hands-on session per team, built around their real documents.