Deal research
Gather, then distil — and they are separate passes on purpose.
The first pass researches the prospect, the company, the industry, the geography and the likely buying process, and comes back with prose and the sources it used. The second takes that, sets it beside what your CRM already holds about the deal, and distils both into five fixed sections plus a readable briefing. Splitting the work means the model that searches and the model that writes can be different, and it means a failure can name which half went wrong.
How it works
How a briefing is built
Two passes and one honest failure mode.
It starts from what you already know
The deal, its stage, the contact, the company and the buying committee around it are assembled first and stored with the briefing. The research is anchored to a real record rather than to a name typed into a box.
It gathers, with sources
A web-capable model researches the company and its market and returns findings with citations attached. That output is kept as it was returned, so what the second pass was working from is still there to inspect afterwards.
It distils, and admits which pass failed
The second pass produces the five sections and the briefing. If either pass fails, the stage it failed at is written onto the briefing before the error is re-raised — so a retry knows what it is retrying, and a person reading a failed briefing is told whether the research or the writing was the problem.
In the product
A briefing anchored to a real deal.
The mechanism
Five fixed sections, and what happens to anything that is not one of them
A briefing is a document a rep will read once and act on, so the useful engineering here is in what the pipeline refuses to pass through rather than in what it generates.
The section keys are a closed set, filled whether or not the model returns them
The five are fixed in code: the prospect and their company, the industry, the geography, the buying process, and what the buyer is likely thinking. The synthesis pass is asked for structured output, and what comes back is reduced to exactly those five. A key the model invented is dropped rather than rendered, and a key it omitted is present and empty rather than missing.
That symmetry is what makes the briefing safe to render as a template. A section that can silently vanish means a reader cannot tell an unanswered question from one nobody asked, and a section the model made up means the shape of the document changes between two deals for no reason a rep could explain.
The two passes are separately modelled and separately billed
Each pass names its own model in configuration, and each is logged to the usage ledger under its own feature key with its own token counts and cost. So the searching half and the writing half can be different models — and can be changed independently — without either becoming a guess about the other. The raw output of the first pass is kept exactly as it was returned, alongside its own citations and its own usage, which means what the second pass was working from is still there to read if the finished briefing says something surprising.
The CRM facts are assembled first and stored with the briefing
Before anything is asked of a model, the deal, its stage, the contact, the company and up to eight members of the buying committee around it are compiled into a snapshot and written to the briefing. That snapshot is the grounding for both passes and it is kept for audit, so a briefing can be read months later against what was actually known at the time rather than against what the record says now. Where a company or contact carries enrichment data, it is read here and only read — this pipeline consumes that data and never writes any back.
A failure names its half before it is raised
If the gathering pass fails, the briefing records that it failed at gather; if the distilling pass fails, it records synthesis. That write happens before the error is re-thrown, which is the detail that makes it useful — an error that propagates first and gets attributed afterwards loses the attribution whenever the process handling it is the thing that died. So a retry knows what it is retrying, and a person looking at a failed briefing is told whether the research or the writing was the problem instead of being shown an empty document.
Citations are merged rather than replaced
The final citation list is the deduplicated union of the sources the gathering pass returned and any source URLs the distilling pass attached to individual sections. Neither list wins. A synthesis that cites something the gather pass did not return is worth seeing, and a gather source that no section ended up using still evidences what was read.
Who it is for
For the half hour before a first meeting
This is the work a good rep does anyway and a busy one skips.
- An account executive walking into a discovery call with a company they met yesterday.
- A sales manager who wants preparation to be even across the team, not dependent on who is diligent.
- A rep selling into an unfamiliar industry or geography.
- Anyone who has opened a deal five minutes before the call and searched the company by hand.
Related
Questions
The things people actually ask.
Does it cite its sources?
The gathering pass returns findings with citations and that output is retained, so the briefing can be traced back to what it was built from rather than being taken on trust.
Why two models instead of one?
Because searching and writing are different jobs with different costs. Each pass names its own model in configuration, so a deployment can change either one without touching the other and without a release.
What happens if the research fails?
The briefing records which pass failed before the error is raised, so it can be retried knowingly and a reader is never left with a blank document and no explanation.
Is it researching my customers or the public web?
Both, and separately. What your CRM knows is assembled from your own records; the public research is about the company and its market. The second pass is what puts the two beside each other.
Does it decide anything about the deal?
No. It briefs. Stage movement, qualification and next steps stay with the person selling, on the deal record itself.
Brief one deal and read the sources.
A briefing you cannot trace is just a well-written guess.