AI agent for B2B prospecting
Context
B2B prospecting done well consumes hours: researching the company, understanding its moment, finding the decision-maker and writing an approach that doesn't read like spam. Done by the founder, that routine competes with delivery and client work. We built an AI agent to do this work and pointed it at our own funnel.
What the agent does
An autonomous three-stage pipeline, with human oversight at the point that matters:
- Discovery: finds 10 to 15 companies per run, prioritizing timing signals (expansion, hiring, operational change) in public business registries and professional networks.
- Qualification: researches each approved company (size, digital presence, decision-maker), classifies it into the right service track and builds a PDF dossier.
- Outreach: drafts a personalized message with a specific hook from the research, sends it and keeps automatic follow-ups on cadence.
Every decision is recorded in a traceable funnel: every discovered company has a history of why it advanced or not.
Quality gates
The differentiator isn't volume: it's the agent refusing to send a bad message. Three gates block the pipeline:
- Signal revalidation: if the hook that justified the discovery is gone, the lead doesn't advance.
- Anti-generic gate: without a specific hook coming from the research, the message doesn't go out. It's held for human review.
- Pre-flight: before each run the agent checks the environment. If anything is down, it aborts and alerts, never sends halfway.
A message only goes out with approval: the human stays in the loop exactly where the cost of a mistake is reputation.
Architecture
Dossier generated by the agent
Technical decisions
Why human in the loop?
A generic or wrong message costs reputation, and reputation doesn't come back with a retry. The agent handles research, writing and cadence on its own, but a message without a strong hook stops in a holding state until review. Reviewing takes minutes; recovering a burned decision-maker takes months.
Why our own agent and not a sales engagement SaaS?
Cadence tools fire sequences, but they don't research or write with context. The expensive work in prospecting is understanding the company and formulating a real reason to reach out. The agent does the whole job; sending is the easy part.
Why a funnel in traceable files?
Every lead has an auditable history: when it was discovered, why it was approved or rejected, which touches it received. No lead disappears into an opaque database, and the agent's behavior can be inspected and corrected by reading the record itself.
Results so far
- ~12 weekly hours of prospecting recovered: research, dossier and draft that took ~2 hours per lead now take minutes of review.
- ~100 companies researched, each with a funnel record; the qualified ones get their own dossier.
- Follow-up never forgotten: automatic cadence with every touch recorded.
Stack
How we execute
Same method we sell in Track 2: locked scope, incremental delivery, agent in production since the first week. The difference is that the client was us, and the agent runs in our sales operation every day.
“I wake up to leads researched, qualified and with a message ready to review. My job became deciding, not typing.”