Is AI safe to use for contract notices?

A contract notice is a legal act on a clock: served late, to the wrong address or with the wrong content, it can cost the whole entitlement, so the question is not whether AI writes fluent notices, it is what happens when it is wrong. Used well, AI sits inside the process: drafting a notice from a template and the facts you give it, checking a draft against the contract's requirements, summarising a chain of events into the factual background, flagging that an instruction looks like it should trigger one. Every one of those uses has a person who knows the contract reading the output before anything is sent, and the deadline itself lives in a register or diary, never in a model's memory. What is not safe is the unsupervised version: asking a general tool what notice to serve and trusting the answer, letting it decide dates, or pasting a confidential contract into a public tool against your own policies. AI also does not know your amendments: it reasons from standard forms unless you hand it the actual clause, and construction disputes are fought on amendments. Treated as a fast, tireless assistant with a supervisor, it takes real hours out of notice administration; treated as the administrator, it is an unpriced risk with a deadline attached.