What you get
“What do prospects actually say about our contract length?”
Somebody in your company knows the answer to that. They know it as an impression formed across thirty calls they half-remember, and they will give you a confident summary that is mostly right and impossible to check.
The calls are recorded. The answer should be quotable.
Answering
Cited from what was retrieved, never from the answer.
The citations are the records the answer was assembled from — which means a citation cannot be fabricated, because it is not something the model produced.
And there is a floor. Below a set level of relevance the material is not close enough to be worth citing, and it says there is no evidence rather than generating something plausible. The reasoning is written down: better to say nothing than to invent.
- Halvorsen Group · 14 Aug22:08“Procurement will not sign anything over three years.”
- Vestra Freight · 2 Sep08:41“Three years is fine, the annual review is what we need.”
- Meridian · 9 Sep36:12“We got stuck on term length last time and lost a quarter.”
- Ask across everything, or about one relationship —
All calls, or scoped to a contact, a lead, an account or a deal.
- A citation jumps to the moment —
Not to the call. To the second.
- Three different kinds of nothing —
Found nothing, the model is unavailable, or the search ran in a reduced mode — distinguished, so the screen can tell you which rather than showing one shrug.
- Follow-up questions are suggested, and capped —
Rather than an open field and a blank stare.
Across many calls
A saved question, run over a filtered set, with the inputs recorded.
Define which calls — by outcome, by meeting type, by date, by what they were about — and write the question once. "Why do our pricing wins succeed?" is a saved analysis rather than a thing somebody does by hand each quarter.
Three decisions make it trustworthy. Only fully processed calls are eligible, because shipping a half-processed call into the question produces an answer nobody can defend. The number of calls is capped, so a match-everything filter cannot sweep a thousand conversations into one prompt. And every run records exactly which calls it saw, so the answer can be audited rather than believed.
- A failed run is still a record —
It lands with its error, so the screen can say the analysis failed rather than showing nothing and leaving you to guess.
- A filter matching nothing says so plainly —
The instruction is explicit: state that no calls matched, and do not fabricate.
›Does it remember the conversation?
Each question is answered on its own today, so put the context you need into the question rather than relying on what you asked before.
›Can a saved analysis run on a schedule?
Not today — they are run when you ask. The field exists and nothing acts on it, so treat scheduling as absent rather than as something to configure.
›Can it answer about calls I am not allowed to see?
It answers from what you can already reach. The same permission governs both.