Agentic Lead Intelligence, scoring the market
Three agents in a line, a scout that finds, an analyst that judges, a database that remembers. Together they turn diffuse market signals into a ranked list of who to talk to and exactly why now.
Most lead lists answer "who exists." The harder, more valuable question is "who is about to need what I sell, and how do I open the conversation." This inquiry builds a three-agent pipeline that treats targeting as a signal-detection problem: it watches for leading indicators of growth and regulatory pressure, scores targets against a maturity framework, and surfaces a personalised hook for each one.
Hypothesis
The signal that a company is about to buy rarely shows up in firmographics. It shows up earlier, in movement, private-equity-backed expansion, a sudden run of offshore hiring, a shift in carrier or compliance mandates. These are public but scattered, and reading them is exactly the kind of judgement work a single prompt does badly and a small division of labour does well. The hypothesis: separating finding, judging, and remembering into distinct agents produces leads with traceable reasoning rather than a black-box score.
Method
- Scout. The first agent casts wide, pulling candidate organisations and the raw signals attached to them, like funding events, hiring patterns, and mandate changes, without judging relevance yet.
- Analyst. The second agent scores each candidate against a regulatory-maturity framework and weighs the signals, producing a ranked shortlist with an explicit rationale per target.
- Database. The third agent persists what was found, dedupes against prior runs, and tracks how signals age, so a target that was "too early" last month resurfaces when it ripens.
- Hook. For each surviving target the pipeline drafts a hyper-personalised opening tied to the specific signal that flagged it, not a template with the name swapped in.
Where it stands
The division of labour is the point. When one agent both finds and judges, its reasoning collapses into an unauditable score; splitting the roles keeps the rationale legible. You can read why the analyst ranked a target highly and disagree with it. The database stage turned out to matter more than expected: most of the value is not in any single run but in watching a signal ripen across runs.
The live research question is signal provenance. A hook is only as good as the signal it rests on, and some signals are noisier leading indicators than they first appear. The current work is grading the predictive value of each signal class so the analyst can weigh them honestly.
Open questions
- QWhich leading indicators actually predict intent, and which only correlate with companies that are already obvious?
- QHow do you keep a personalised hook on the right side of the line between "well-researched" and "uncomfortably surveilled"?
- QCan the analyst's scoring framework be made portable across industries, or is regulatory maturity inherently sector-specific?