AI lead generation: build a verified B2B lead list

A useful AI-generated B2B lead list is a sequence of evidence-backed steps, not one broad search. Define the ideal customer profile, discover more candidates than you need, verify the company and current role, enrich only the fields required for qualification, find work emails without guessing and retain provenance for human review before outreach.

Start with a testable ICP

Translate the request into row-level acceptance criteria before running paid work. For a list of recently funded fintech founders, an accepted row might require a unique person, a current founder role, a company that meets the agreed fintech definition, a Seed or Series A announcement inside a fixed date window, a supporting source and a company domain.

Define ambiguous terms up front. Decide whether former founders count, which geographies are in scope, whether one company may contribute several people and what evidence is sufficient for a funding event. Ask the agent for a small sample before scaling; correcting the contract at five rows is cheaper than rebuilding fifty.

Discover companies and people separately

Use the live company-discovery capabilityto create a broad candidate set, then normalise company names and domains before deeper research. Candidates are not yet accepted leads. Over-collect because some will fail the date window, company-fit test, role check or contact-data stage.

Verify material company claims independently. The company-news capability can help locate an attributable funding announcement; record the source, date and round rather than accepting vague language such as “recently funded”. Then use professional-profile searchto identify a current founder or co-founder. A matching name alone is not enough: distinguish current founders from former founders, investors, advisers and people at similarly named companies.

Enrich only what you need to qualify

Enrichment should answer the acceptance criteria, not fill every possible spreadsheet column. Use company enrichment for the firmographic or funding fields that materially affect qualification. Keep the retrieved value, source or tool, and date checked so later reviewers can distinguish evidence from the model’s interpretation.

Treat missing data as missing, not as a negative fact. An absent funding record does not prove that no funding occurred, and an incomplete employee count should not become a precise estimate invented by the model. Flag conflicts for review instead of selecting the most convenient value.

Find and verify contact data

When the brief requires work emails, use person-email discovery with the accepted person and company identity. If no address is returned, leave the field blank; never infer an address from a domain pattern and present it as found.

Pass a returned address through the email-verification capability and retain the status and verification date separately from the address. A deliverability signal is time-bound and is not consent to contact someone. Users remain responsible for the laws, platform terms and outreach policies that apply to their use of professional data; the UK ICO’s B2B marketing guidanceis one useful jurisdiction-specific reference.

Review provenance before outreach

Stop when the required number of rows passes the evidence contract, not when a spreadsheet happens to reach that length. Keep accepted, incomplete and rejected candidates separate. The final export should include company and domain, person and current role, relevant qualification fields, evidence URLs, contact status, tools used and research date.

Run a duplicate check across normalised domains, people and profile URLs, then sample the evidence manually. “Research verified”, “email verified” and “commercially qualified” are different claims; label them separately.

A reusable brief for your agent

Build a list of 50 current founders or co-founders at fintech companies that
announced Seed or Series A funding in the 12 months before today.

Before paid work, use Scrollport discover and inspect to select each tool from
the live catalog. Accept a row only when company fit, funding round, funding
date and current founder role are supported by recorded evidence.

If work emails are requested, never guess an address. Find it with a published
tool, verify it separately and retain the verification status and date.

Return accepted, incomplete and rejected rows separately. Stop when 50 unique
rows satisfy the acceptance criteria, with a source for every material claim.

This workflow can later become an agent-side skill: the catalog tools stay individually discoverable and inspectable, while the skill preserves the decisions, evidence gates and output format that make the process repeatable. Start with the leads and prospecting catalog and inspect every selected tool immediately before use.