AI governance for in-house legal teams: five minimum controls
By Asım Serdar Yılmaz · Published · Last updated · 2 min read
A team six months into generative AI asks to scale it, and nobody in the room can say what the rules are. That is the ordinary case rather than the alarming one. Use among US in-house respondents went from 23 to 52 percent in a year in the Association of Corporate Counsel’s 2025 survey, while Thomson Reuters’ 2025 report found 52 percent of organisations across industries still had no policy at all, and the gap between those two findings is where this starts.
Before a team scales, I want five controls written down on one page. Five is an arbitrary number, and it is simply what has fitted on one page without anything important falling off the bottom.
1. Data handling
First, data handling. Write down what may enter a prompt, and write it before anyone opens a chat window. Client identities, personal data and privileged material each need a rule of their own, because one sentence stretched across all three tends not to hold.
2. Accountability
Second, accountability. Every AI-assisted output has a named owner: the model drafts it, a lawyer signs it, and that lawyer is the one who answers for it afterwards.
3. Human review
Third, human review. Review depth should follow the impact and reversibility of the output rather than whether it happens to leave the team, which is the line most teams reach for first. High-impact or hard-to-reverse work goes to a qualified reader with authority to reject it, and a spot check after the fact doesn’t meet that bar.
4. Logging and records
Fourth, logging and records. Keep the metadata and the review trail: which tool, which version, who approved. Set a retention period and a deletion rule rather than storing full prompts by default. When a regulator asks how a particular document was produced, nobody’s memory will do the work a record does.
5. Error handling
Fifth, error handling, and here the instinctive order is backwards. When an AI-assisted output fails, the first three moves anyone reaches for are to roll it back, clear the session and rotate the key, and each of those can erase the record you will need in order to explain what happened. Preserve first: the prompt and the output you chose not to retain by default, the model version, the logs. Contain after that, unless the harm is still running, as with a misused credential or data still leaving; then contain at once and capture what you can alongside. Then name, on the same page, who may switch the workflow off and who tells the business.
These five are the floor for a pilot rather than the finished house. None of them needs a new platform or a standing committee; one page that links out to each control, signed first by the general counsel, will carry a team a long way. What the five do not settle is what an agent may reach while it reads a document a stranger wrote.
Sources
Next: KVKK standard contracts: what the first full year’s 2,497 notifications say →