Build log: an AI agent that builds and enriches a B2B database

This is a build log, not a pitch: the AI-assisted pipeline we run to turn an ideal client profile into a verified B2B database. ICP criteria go in one end; a deduplicated, enriched, verification-gated list comes out the other, ready to load into a sending tool. The agent does the volume work — sourcing, cross-referencing, enrichment — and humans make the judgement calls it cannot.

It automates the middle stages of the process described in the B2B database building guide; if you have not read that, it is the map this piece zooms into.

What goes in?

A written ICP: sub-vertical, headcount band, geography, and the role that owns the number we are selling to. This part is human work and stays human work — if you cannot name your buyer precisely, the agent has nothing to aim at, and it will happily build you a large, useless list. Garbage criteria in, garbage records out, faster than ever.

What does the agent actually do?

The pipeline runs in five stages. When the ICP criteria are loaded, then:

1. Company sourcing. The agent queries multiple data layers — the public register, commercial data platforms, directories — for companies matching the filters, and deduplicates across sources. Cross-referencing is the value here: a company that appears in one source with a plausible size signal gets checked against the others before it counts.
2. Decision-maker identification. For each surviving company, the agent scans public signals — team pages, LinkedIn, filings — to find the person most likely to hold the target role. It proposes one or two names per company, not ten.
3. Enrichment. Each proposed contact is enriched with confirmed role, seniority and a candidate work email address, built or found from known patterns and sources.
4. Verification. Every address is checked before load: hard-bounce addresses removed, catch-all domains flagged separately, remaining records scored for bounce risk. This gate is non-negotiable — verification before sending is what protects the sender reputation the whole campaign runs on.
5. Load. Clean records go into the sending tool, segmented by sub-vertical, with the flagged catch-alls held back or routed to lower-risk treatment.

What falls out along the way?

A lot, and that is the point. When a few hundred companies enter the pipe, then a meaningful share drops at each gate: some turn out to be dormant or mismatched on closer inspection, some yield no identifiable decision-maker, and some contacts fail email verification outright. I will not put fake precision on the attrition — it varies by sub-vertical, company size and how well-documented the industry is online — but it is entirely normal for the final loaded list to be well under half of the raw companies sourced.

Firms sometimes read that as waste. It is the opposite. Every record that falls out is a bounce that never happens, a spam complaint never triggered, a wasted send never made. The output number matters less than the survival rate's honesty: a 1,200-record list where every address verified beats a 5,000-record list of hopeful guesses in every metric that reaches a bank account.

What is AI genuinely good at here?

Two things, and they are the two things humans are worst at:

• Scale without fatigue. Checking the 400th company website with the same care as the first. A human researcher degrades over an afternoon; the agent does not.
• Cross-referencing. Holding four sources in view at once and noticing that the "operations director" on the website left according to LinkedIn three months ago. Humans can do this; they cannot do it two thousand times.

What still needs human judgement?

Three calls the agent does not make:

• The ICP itself. Choosing which sub-vertical to target is a strategy decision, made from knowledge of the client's best existing customers — not a data task.
• Edge-case fit. A company that half-matches the filter needs someone who understands the service to say yes or no. The agent flags; a human rules.
• Spot-checking. We sample the output of every build. Not because the pipeline is unreliable, but because unchecked automation drifts, and the cost of drift lands on deliverability weeks later.

This division of labour is the honest version of what AI does in outbound. It compresses the research grunt work from weeks to days; it does not know your business, and it will not fill your pipeline on its own. The firms that get burned are the ones that automate the judgement along with the labour.

Why publish the mechanism?

Because the mechanism is not the moat — execution and judgement are. Any firm could assemble a version of this pipeline. Most will not, and of those that try, most will skip the verification gate or the spot-checks, because those are the unglamorous parts. If you would rather run on a database built this way than build the pipeline yourself, that is the work we do.

---

Next step: the Growth System Audit — £450, seven days, credited against any build — maps where your growth system leaks and what to build first.