The situation
The client sells complex engineering work. Every deal starts as a conversation, and every conversation contained the qualification - who the champion was, what the project actually was, whether budget had been discussed, what would happen next.
None of it reached the CRM. It reached a transcript nobody read, and a rep’s memory, which is not a database. The pipeline was built from whatever survived the drive home.
Nobody asked us to summarise the calls; a summary makes a nice demo and changes nothing. The ask was to turn the conversation into structured, queryable properties and let the system decide whether a deal exists.
What RevOps XL built
1. Capture, then a deliberate fork
A meeting recorder joins the call and transcribes it. From there the system splits, because two different problems need two different tools.
The straightforward extractions - buying role, intent rating, budget discussed, next steps agreed - run natively inside the CRM: keyword listening enrols the contact into a segment, and segment enrolment fires a workflow that writes the property. There’s no external dependency to maintain, and it survives anyone leaving.
The extractions a keyword can’t do - what this person actually needs, what’s driving their urgency, which services the work would require - go out to an orchestration layer, through a model, and back in via the API. That path exists because the native tool has a ceiling, and ignoring that ceiling is how these projects fall over.
2. Roughly fifty workflows, and one that matters
Nine workflows for buying role. Sixteen for project type. Thirteen for project stage. Three for intent. Two each for budget, next steps and whether AI is in scope.
Then one workflow that reads all of them and creates a deal, but only when high or medium intent coincides with agreed next steps and an identified champion. All three of those have to hold, or nothing happens.
That last workflow is the entire product; the other forty-eight just feed it.
3. The guardrails are the engineering
Anyone can feed a transcript into a model. The work is stopping it from ruining your CRM by Friday.
- Meeting-type isolation. A working session is not a sales call. Properties only populate on sales and discovery calls, or your delivery meetings rewrite the pipeline.
- Deal-flood prevention. A deal is created only if no open deal already exists on that account. Without it, every good call spawns a duplicate and the forecast doubles overnight.
- A ninety-day recency rule. Properties update only if empty or untouched for ninety days - so a throwaway line on call six doesn’t overwrite what was carefully established on call one.
- Internal-meeting bypass. Staff domains are excluded on both paths. Otherwise the team qualifies itself, enthusiastically, all week.
4. Identity resolution, before anything is written
Every write is preceded by a lookup: find the record by email, take its internal ID, then upsert against that ID rather than any string.
It sounds pedantic. It is the difference between enriching a contact and creating their evil twin.
Where it landed
The system is in production. Deals now appear because three qualification conditions were met in a real conversation, not because someone remembered to log one.
The count of new deals matters less than whether they hold up, and these ones do: every property on them traces back to a sentence somebody actually said.
How full the fields are
Counted from the client’s HubSpot deal export of 30 September 2026, 973 deals. The system only exists for part of that book, so the split is by creation date. Share of deals with the property filled in, rounded to whole percents.
Before 2026 almost nothing was recorded. Among 2026 deals that have a recorded meeting, about two in three now carry each of the qualification fields, and 92 in 100 carry an urgency driver. The remaining 2026 deals came in without a recorded meeting for the system to read.
Of the 42 deals the system opened itself, all 42 carry buying role, intent, project type, stage, AI scope, services and urgency, and 41 carry budget discussed. That figure is close to circular, because the run that opens a deal also writes its fields; it shows the writes land.
How the objects connect
42 of 42deals the system opened are attached to the contact on the call
38 of 42of those also carry a company
91%of all 973 deals are linked to a company, and 91% to a contact
12deals, 1.2% of the book, have neither a contact nor a company
Won deals are better linked than the book as a whole: 270 of 282, 96%, sit on a company. Contacts are the weaker side, with 80% of 3,455 attached to a company.
Which partners send deals
The referral trail lived in a free-text Lead Source field, where one referrer appeared under seven different spellings and nothing could be counted. We mapped it to eight referral types. 277 deals trace to a named person, firm or site, and 72 of them, 26%, came through partner firms. 17 of those 72 closed won.
Another 37 name a real person whose relationship to the company we couldn’t confirm from the record, so their type is blank and the partner count is a floor. It’s the first time the question of which partners actually send business has had an answer you can filter on.
