To build a lead list with an AI agent that is actually usable, you need four things the agent cannot supply on its own: a tight definition of who you want, a list of public places where those companies show up, a verification rule for every row, and a decision about what you will do with the list that keeps you inside privacy law. With those in place the agent does the slow part, reading hundreds of pages and filling a table, in an hour.
Here is the process I use for lead research for small B2B products, with the prompts and the checks.
Step 1: define the profile in checkable terms
"SaaS companies" is not a profile. A profile is a set of facts an agent can confirm from a public page:
- Company type and size signal: for example, "software companies with 10 to 50 people on their team page"
- A trigger: hiring for a role, a recent launch, a tech they use, a funding announcement, a product listed on a marketplace
- Geography, if it matters
- Exclusions: competitors, enterprises, agencies
Write it as five lines. Every row in the final table must satisfy all five.
Step 2: pick sources the agent can read
Public, no login, allowed by the site's terms:
- Company directories and marketplaces in your niche (app stores, integration directories, startup lists)
- Job boards: a company hiring for a role is a strong signal they have the problem your product solves
- Conference speaker and sponsor lists
- Public tech-stack signals on company sites (what their careers page says they use)
- News and launch announcements
- The companies' own sites for the contact page, the team page and the about page
Not sources: scraped personal data from social networks, purchased lists, anything behind a login you agreed not to scrape.
Step 3: extract one row per company
Ask the agent for a fixed set of columns: company, website, what they do (one line, from their own site), the trigger that qualified them with its URL, size signal with its URL, location, a generic contact route (contact page or published company email), and the date checked.
Prompt: "For each company, fill the columns from the pages you read. Put the URL you read next to each fact. Leave a cell blank if you did not find the fact; do not guess."
The URL-next-to-every-fact rule is what makes the list checkable.
Step 4: verify before scoring
Pick ten rows at random and open the URLs. If more than one of the ten has a fact that does not match the page, the extraction is unreliable and you need to tighten the prompt or the sources before trusting the rest. Common failures: company size taken from an old article, a trigger that is three years old, a website that now redirects to an acquirer.
Then ask the agent to re-check the whole list for two things: the website still resolves, and the trigger URL still shows the trigger.
Step 5: score with a rule, not a feeling
A simple score the agent can apply the same way to every row:
- Trigger is less than 90 days old: 2 points
- Size within the target range: 1 point
- Uses a tool your product integrates with: 1 point
- Contact route is a named role or a contact form rather than a generic inbox: 1 point
Sort by score. The top 30 are the ones worth a personal message; the rest are a watch list.
Step 6: keep it legal and polite
- Company-level information from public pages is the safe ground. Personal emails harvested from profiles are where privacy rules in the EU, UK and several US states apply, and where the reputational cost lands.
- Write to a role or a contact form first. If you do write to a person, say where you found them and make opting out one reply away.
- Store the list with the source URLs and the date; delete rows that go stale.
- Never let an agent send outreach on its own. It drafts; you read and send.
What the agent is good and bad at
Good: reading 300 pages without getting bored, filling columns consistently, re-checking URLs, drafting a first-line for each company from its own site.
Bad: knowing whether a company is a good customer (that is your profile's job), knowing whether a fact is current (that is the verify step), and judging tone in outreach (that is you).
Deliverable
A CSV with the columns above plus score, and a short note: profile used, sources read, how many candidates were found, how many passed verification, the date. Thirty scored rows with URLs beats three hundred unverified names.
Related reading: How to find a niche with an AI agent: demand signals first for choosing who to sell to before building the list, and When to trust AI agent output and how to verify it for the verification habit.
FAQ
Can an AI agent build a lead list? Yes, from public sources, with a precise profile and a verification step. It reads and fills tables well; it does not know which companies are good customers or whether a fact is current.
Where should an AI agent look for leads? Niche directories and marketplaces, job boards, conference lists, launch announcements and the companies' own sites. Nothing behind a login, nothing scraped from personal profiles.
How do I verify an AI-generated lead list? Require a source URL next to every fact, spot-check ten rows by hand, and have the agent re-check that websites resolve and triggers are still visible.
Is it legal to use AI for lead generation? Company information from public pages is generally fine. Personal data is covered by privacy law in many places; write to roles and contact forms first, and let the agent draft, never send.
How many leads should a list have? Thirty verified, scored rows is a good first batch. Quality drops fast past that, and thirty is what one person can follow up properly in a week.
What columns should a lead list have? Company, website, one-line description, trigger with URL, size signal with URL, location, contact route, date checked, score.
If you want a verified list built this way, the lead list research gig on my store delivers it for $9: https://rmgo.sites.tab.bot/leads?s=blog
Latest comments
0