Gave an agent write access to fix one class time on our website and it rewrote our waiver.

Two weeks ago I asked our agent to move one class on the website, Thursday intro to bouldering at our east location from 6:30 to 7:00. It did, and while it was in there it also "tidied up" our membership freeze policy and rewrote the waiver paragraph for day passes so it "flowed better." Nobody asked it to touch either one. Worse, it found a 2024 price sheet PDF in the Drive folder it can read and decided that was the truth, so 3 youth class prices went back to the old ones. Front desk caught it two days later when a member showed up with a screenshot of the cheaper price. We honored it, obviously. It's n8n with Claude making the edits through the site's API, kicked off from our schedule spreadsheet. We're 4 locations and the schedule changes maybe twice a month, so this felt like a safe first thing to automate. Rolling back was its own mess. I didn't know what else it touched, so I restored the host's backup from before the edit, which wiped two days of form submissions, including 14 birthday party inquiries. The notification emails had contact info for 9 of them, the other 5 just say "new submission" so we're hoping they call back. The owner wants to keep it "with guardrails" and designing them is on me. So how are people scoping write access on anything customers see, down to a single field if that's even possible? And how do you stop it improving stuff nobody asked about? My prompt literally said only change the time. Does everyone just review every change? That kind of defeats the point of me setting it up. Is anyone running theirs against a copy of the site with a person hitting publish? Or is this overkill and I should just go back to doing it myself? Funny thing is the waiver it wrote reads better than ours, but our insurance broker wants a list of everyone who bought day passes online while it was up.