Everyone on the instance is asking it questions it can already answer — the difference is only how long it takes them to get there. Kaddiya is the same agent for all three seats below; what changes between them is the roles the person signed in with, and therefore what the agent can see. What doesn't change is the last step: somebody reads the draft and approves it.
Forty open tickets, and the answer to a third of them is already somewhere in the instance — in a resolved case from March, in a knowledge article nobody links, in the journal of the ticket sitting two rows down.
Ask what's on your plate. Kaddiya pulls each case, reads the whole journal rather than the short description, looks for the last time someone fixed the same thing, and hands back a draft reply. Claude does the reading. You keep the judgment, and you press Send.
It does not resolve or close your tickets, reassign them, or reply to anyone on its own. The draft sits there until you send it, and if you never do, nothing was written.
You know the platform. What you don't know is this instance's particular forty customizations — which field was renamed, which business rule fires first, what the last person's update set actually captured.
Schema truth instead of guessing. Encoded queries you can lift straight into your own work. The update set checked before anything is proposed. Complete scripts written for your review, grounded in the documentation for the release your instance is actually running — with the topic it used cited underneath.
It does not commit code. A script arrives as a proposal card showing the exact payload; creating the record and creating an update set are each their own click, on your credentials.
Breaches by group. Ageing P2s. Who holds which role, and who quietly still holds it after changing teams. Why an ACL denied that row when the user swears they have the role.
Questions that used to be a report request, or a favour asked of the one person who writes good encoded queries, become a sentence. Every query it ran is on screen while it runs, so the number in the answer is one you can check — or re-run yourself tomorrow without asking anyone.
It is not a BI tool and keeps no dashboard. Each answer is computed from queries against your instance at the moment you ask, which is also why the queries are shown rather than summarised.
There is no service-desk edition and no developer edition. It is one agent, and the reason it behaves differently for the three people above is that they signed in as themselves: their roles, their groups, their user criteria. A service-desk agent sees the queues they're on. A developer with personalize_dictionary sees the dictionary. Neither of them sees anything the other's roles wouldn't have shown them anyway.
That is also why there is nothing to provision. You don't map people to seats, license them by persona, or maintain a second directory — the instance already knows who everyone is, and we ask it rather than duplicating it.
The permission model is not a feature we built. It is your instance, doing what it already does.