AI in lead research: what it's actually good at

AI is genuinely good at the grunt work of lead research — reading thousands of company websites, classifying firms against your ideal client profile, and cross-referencing sources at a fraction of manual cost. It is poor at exactly two things: telling the truth without verification, and making the subtle fit judgements that decide whether a technically-matching firm is actually worth your time. Build the pipeline around that split and lead research stops being the bottleneck of outbound.

What does lead research actually involve?

Mechanically, five jobs: find companies that might match your ICP; find the right people inside them; fill in the fields that make targeting and copy possible (sub-vertical, size, signals, tech); verify that contact details are real; and prioritise the result. Done by hand, this is genuinely slow work — minutes per record, thousands of records — which is why most small firms either skip it (and send generic email to bad lists) or pay heavily for it. It is also the layer where campaigns are won or lost, because as I argue across AI Automation for B2B: what actually works, no downstream cleverness compensates for researching the wrong companies.

What is AI genuinely good at here?

The tasks that are reading-heavy, judgement-light, and checkable:

• Classification at scale. Given your ICP definition, a model can read a firm's website and answer "is this a commercial cleaning contractor or a domestic one?" thousands of times without fatigue. This single capability makes sub-vertical segmentation affordable for small firms — it used to be an intern-month per list.
• Cross-referencing. Joining Companies House records, websites, and directory data into one coherent record — catching that two entries are the same firm, or that a "20-person agency" filed accounts as a two-person LLP.
• Signal extraction. Hiring pages, new office announcements, tech visible on the site — turned into structured fields you can filter and sequence on.
• Research notes. A three-line summary of what the firm does, in plain language, attached to every record so whoever handles the reply has context instantly.

Notice the pattern: in every case the model processes text that exists, rather than asserting facts from memory.

Where does AI fail in lead research?

Three places, reliably. First, hallucination: ask a model what a company does without giving it the website and it will answer anyway, fluently and sometimes wrongly — every claim must be grounded in a fetched source. Second, staleness: people change jobs constantly; any contact detail a model "knows" is historical by definition, so email verification belongs to dedicated verification tools, not to the LLM. Third, nuanced fit: "matches the ICP on paper but their site screams they buy on price" is a judgement experienced humans make in seconds and models make badly — which is why a human should sample and prioritise the final list rather than rubber-stamp it.

What does the working pipeline look like?

The mechanism, step by step: when a segment is defined, then candidate companies are sourced from places that enumerate rather than guess — Companies House by SIC code and filing data, Sales Navigator, directories. When candidates exist, then an AI pass reads each website and classifies against the written ICP, attaching evidence for every claim it makes. When a record passes classification, then enrichment fills the targeting fields and finds the named decision-maker. When contacts are found, then every email address goes through verification before it is allowed near a sending inbox. When the list is assembled, then a human samples a few dozen records — and when the sample error rate is more than a few per cent, then the classification prompt gets fixed before the list ships, not after the campaign fails.

That pipeline is what lets a list be both large and precise — the combination that manual research prices out of reach. It is the reason we can build a verified, segmented database as a £950 standalone project rather than a five-figure one.

Should you rent this capability or own it?

Own it. The pipeline is a durable asset: the prompts encode your ICP, the workflow encodes your process, and the output feeds every future campaign. Renting it through an agency's black box means losing all three the day you part ways — the argument I make at length in Why you should own your automations. The components are affordable and sit comfortably inside the modest toolset I cost out in An AI stack for a ten-person firm, with costs.

And the pay-off arrives downstream: research quality is the quiet variable behind reply rates. Cold email still works — bad cold email is what died — and the difference between the two is decided here, before a single word is written.

---

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