The private office that says it has not adopted AI is usually describing procurement, not reality.
Staff summarise documents in consumer tools. Advisers use models to draft correspondence. Meeting platforms transcribe conversations. search, email and office software ship AI features by default. Vendors classify records and screen transactions using systems the office never selected directly.
Shadow AI is attractive for a reason
It removes tedious work quickly. It can turn a long pack into a decision note, translate correspondence, compare documents or produce a first draft before a formal system would even be approved. A policy that only prohibits use drives it into less visible channels.
Inventory the decisions, not the brands
Tool lists age quickly. Map what people are using AI to do: read, classify, summarise, draft, recommend, monitor, authenticate or act. Then record the information involved, consequence of error, external retention, human authority and evidence retained.
A transcription tool touching a sensitive family conversation may matter more than a sophisticated model drafting public marketing copy. Risk follows context and consequence.
The minimum operating boundary
- No private identity, health, family, legal or transaction material enters an unapproved public service.
- AI output does not become an instruction, fact or decision without a named accountable person.
- High-consequence use retains the source, prompt context, output and final human change.
- Vendors disclose model providers, retention, training use, subprocessors and deletion behaviour.
- Every approved use has a shutdown owner and a manual alternative.
The position
Do not begin with an AI strategy deck. Begin with a two-week discovery of real use and the decisions it touches. Approve useful patterns, close dangerous ones and give staff a safe route that is faster than improvisation.
AI is already inside. Governance begins by becoming honest about where.