The need
Partner prospecting was done by hand by a co-founder, searching LinkedIn profile by profile. About 20 hours a week produced only 10–15 potentially qualified prospects, with generic outreach messages and no structured tracking. It was a direct bottleneck on the growth of the partner network.
What I owned
The design and the build of the pipeline, its migration to production, and its documentation.
The solution
The sales team never leaves Notion, which is the database, the interface and the CRM at once. A director adds a card, asks for a run in a chat, and each prospect comes back with verified company data, a persona, a score from 1 to 100 and a qualification report on its own page.
Key modules
Enrichment
Fully deterministic, with no model calls. A cascade of sources gives legal identity, location, financials and reputation, and degrades gracefully when the paid source is down.
SIREN matcher
Refuses to guess. Attaching the wrong company with the same name would stick, because human-entered values are never overwritten, so an empty flagged field is better than a plausible wrong one.
Qualification
Sub-agents work in batches of 5 with 8 running in parallel, fill the missing data, assign one of three personas, and write a score and a report.
Protected stages
Cards from "Contacted" onward are live sales work and are never touched. The rule is enforced in the query filter, in a shared list, and in a check that raises an error before every write.
How it was built
The move to production was sequenced on purpose: 14 local tasks first, then changes to the real database one by one, each validated before the next. A complete snapshot of all 374 pages, including page bodies, was taken before the first change, with targeted backups before every destructive step.
Version 2 simplified the system deliberately: no scheduled task, no test database, one schema and one code path, and a first phase with no AI at all.
Results
- 216 prospects enriched and qualified in production, on the real database.
- A rough prospect name becomes a full candidate profile with no manual research.
- The value targets set for the wider project are 150–200 leads a month against 10–15 today, about 0 hours of weekly prospecting against 20, and a direct tool cost of about 0.59 per qualified lead. They depend on automated lead discovery, the next phase of the project.