Analytics
The useful question is usually the one no card was built for.
“Which roles have had no client movement in a fortnight?” is not a dashboard. It is a question you have on a Tuesday, and by the time somebody builds the report it does not matter.
Asking
Plain English, answered from your records.
Ask in a sentence and get an answer grounded in your own hiring data, with the records attached so checking it takes one click rather than an act of faith. Past questions are kept, so the one you asked last month can be re-run.
Where the data does not support an answer, it says so. That is the part that makes the rest usable.
- Which of our roles have slowed down since August?Four roles. Three stalled at the client's first response; one has had no new submissions in eleven days.
- Which sources cost us the most per hire?I cannot answer that — what each source costs is not recorded, so any figure I gave you would be made up.
Being told
Anomaly sweeps run whether you asked or not.
A sweep looks for changes worth a human glance — a stage that suddenly stopped converting, a reviewer who has not opened anything in nine days, a source whose volume collapsed, a role sitting still while others move. Alerts can be dismissed, and a sweep can be run on demand.
The point is the asymmetry: you only query what you already suspect.
- Scoped and permissioned —
It answers from what you are allowed to see.
- It will decline a question it cannot ground —
Rather than producing a confident aggregate from a dimension the records do not carry. A refusal you can read beats a number you would have repeated in a board meeting.
- The same idea across the platform —
There is an ask surface over the call corpus, and one over the tenant's own documents and policies.
Asked of any AI answer surface
Where it helps, and where it stops.
›How do we know an answer is right?
You do not have to take it on faith — it answers from your own records, past questions are kept so one you asked last month can be re-run, and it declines what it cannot ground. A surface that answers everything is one whose answers you have to check, which is slower than not asking.
›Can it see things a user should not?
No. It answers from what that person is already allowed to see, so the answer changes depending on who asks.
›What is an anomaly sweep actually looking for?
A change worth a human glance — a stage that stopped converting, a reviewer who has not opened anything in days, a source whose volume collapsed, a role sitting still while others move. You only query what you already suspect; this is the other half.
›Can we turn the alerts off?
Individual alerts can be dismissed, and a sweep can be run on demand rather than waited for.
›Is this a differentiator?
The plain-language surface is not — plenty of products have one. The part we would defend is that it refuses rather than guesses, and that the sweep runs without being asked.