B2B SaaS / HR Tech
Discovery calls, recorded, read, and acted on by machine.
How a French B2B SaaS trebled SQL-to-opportunity conversion by feeding recorded discovery calls into a structured LLM pipeline and rewriting every downstream step around what it found.
CASE / 11
French HR SaaS
The company had a 14-person sales team, strong inbound demand from a well-ranking content function, and a pipeline that leaked badly between SQL and first-closed-won. Win rates against the top three competitors were fine. Win rates inside the same bracket but with different buying committees were inconsistent in ways nobody could explain. The VP Sales had been running win-loss interviews manually for a year and was out of bandwidth.
GEO
France · Benelux
Setup
Revenue ops + sales enablement
Duration
11 weeks
Shipped
Q1 2026
The data was already there, unread
The data was already there, unread
The team had 18 months of recorded discovery calls in Gong. Nobody had time to listen to more than the weekly coaching sample. An enormous amount of signal was sitting on S3, unread. The first engagement decision was to treat the recording archive as a data source, not a coaching tool.
We built a structured extraction pipeline: each call's transcript was passed through Claude with a schema requiring the model to identify the buying committee composition, stated priorities, objections raised, competitor mentions, technical prerequisites, decision timeline, and budget signalling. Output was persisted as structured columns in Snowflake, joined to the Salesforce opportunity record. Three weeks of ingestion gave us 4,200 calls indexed against outcomes.
What the pipeline found
What the pipeline found
The biggest single finding was about committee composition. Deals where the first discovery call included an IT or DPO representative had a 31% higher close rate and a 23% shorter sales cycle than deals that surfaced compliance concerns in stage three. The pattern was obvious in hindsight: dragging a compliance conversation in late was creating rework and trust damage. It was not visible until we had structured data on who was in which call.
Second, the content function was ranking for questions the sales team was not prepared to answer on the first call. Prospects were arriving primed by blog content about integration depth; AEs were spending the first twenty minutes of discovery on generic qualification. The mismatch cost conversions without anyone being able to point at the cause.
Discovery, rewritten
Discovery, rewritten
We rebuilt the discovery playbook in three specific ways. First, the pre-call brief was generated automatically from the prospect's CRM record, site behaviour, and any prior contact, with suggested primary questions aligned to the buying committee we believed would be in the room. AEs could override; most did not. Second, the post-call summary was generated from the recording and populated Salesforce automatically, which killed the 35 minutes per day AEs were spending on data entry and, more importantly, meant the data was actually complete. Third, the stage-three gate was rewritten to require committee composition data, not just MEDDIC scores.
What we did not automate
What we did not automate
We did not automate outbound messaging. The team had a clean inbound motion; grafting on AI-written cold email would have cannibalised the content function's trust signal. We also did not automate objection handling on live calls. Real-time AI prompting to AEs during calls was tested and rejected: it distracted more than it helped, and the objections the team was losing on were not the ones real-time tools surfaced anyway.
Launch With Us