AI Lead Generation: Agents, Add-ons, Prompts

For lead generation, use AI to find companies that fit a clear profile, read their public pages, score fit and prepare a clean list, then let a person decide who gets contacted. Pair a search add-on (Exa, Brave or Tavily) with a reader (Firecrawl or Apify) and a CRM connector. The model cost is small; search credits, data rules and review time decide the real cost. Checked 31 October 2026.

The job in one table

What the job is

Lead generation is the work of finding people and companies who might buy, before a salesperson speaks to them. It has three hard parts. Finding candidates takes searching. Knowing whether a candidate fits takes reading. Keeping the list honest, with no duplicates, no stale rows and no one who asked to be left alone, takes discipline.

AI is strong on the reading. An agent can open fifty company sites in the time you open three, and write two lines on what each one does. It is weak on judgement and dangerous on scale: a bad rubric run across ten thousand rows produces ten thousand bad leads, and an outreach tool that sends without review can harm your brand and break the law.

Which agent fits

The pattern is the same as in our sales build guide: the agent prepares and a person decides. Start with the research and scoring parts. Leave sending to people.

Add-ons that help

Pages: Exa, Brave Search, Tavily, Firecrawl, Apify, Fetch, Airtable, HubSpot, Zapier. Compare search servers in best browser MCP servers and browse the whole list at MCP servers.

Prompts and kit items

Use the prompt library for prospecting prompts by task. In the kit, the plan-before-coding prompt helps you scope the script before an agent writes it, and the CLAUDE.md starter is where your sourcing rules live: which sites are allowed, which fields to store, what to never collect.

A scoring rubric works best as a table the model must fill in:

"Evidence required" is the key column. Without it the model fills in scores by feel.

A safe workflow, step by step

1. Write the profile and the rubric by hand, with five good and five bad examples.
2. Choose sources you are allowed to read, and write down their terms.
3. Search: run the add-on query, capped at a number of results.
4. Read: fetch each page and extract the facts into fixed fields.
5. Score against the rubric, with a quote for each point.
6. Review: a person scans the top and the bottom of the list, plus a random sample from the middle.
7. Enrich only the survivors, with only the fields you need.
8. Hand off to a rep or into the CRM as a draft list.
9. Log every source and date, so you can show where each row came from.

What it costs

Model. On Anthropic's published prices, Sonnet 5.5 is $2 per million input tokens and $10 per million output; Haiku 5.5 is $0.10 and $0.50.

Worked example (the token counts are our assumptions). Reading one company page of 6,000 input tokens and writing a 300-token summary and score:

Search and scraping. This is usually the larger line. At the add-on prices above, 1,000 companies that each need one search call and one page read use about 2,000 credits or requests. On Brave at $5 per 1,000 searches that is $5 for the search half. On Firecrawl's Hobby plan, 5,000 credits cover the page reads for $19 a month. Firecrawl's pricing page says a basic scrape or crawl costs 1 credit per page, and a search costs 2 credits per 10 results.

People. Reviewing a list of 1,000 at 20 seconds a row is about 5.5 hours. That is the number to watch.

See Claude API pricing for model rates.

Risks and how to review them

The FTC guide for US marketing email lists accurate headers, honest subject lines, an ad disclosure, a valid postal address, an opt-out and honouring opt-outs within 10 business days. Other regions differ and several are stricter about cold outreach and about storing personal data. Check each market's regulator for its rules on B2B outreach and personal data. Your company is responsible for how the data was collected, whichever tool did it.

Measure the results

Compare with a list built the old way for the same period.

Choosing between the search tools

The three search add-ons differ in what they are good at, so match the tool to the question.

You rarely need all of them. Start with one search tool and one reader, run twenty companies by hand, and add a second tool only when the first misses something you can name.

Example: a week of prospecting (illustrative)

Suppose you sell a scheduling tool to dental clinics. On Monday you write the profile: independent clinics, two to ten dentists, a booking page that is a plain contact form. You write five clinics you would love and five you would not. On Tuesday the agent searches for clinics in one city, reads each site and fills your fixed fields: number of dentists named, booking method, last site update, any online booking tool named. On Wednesday it scores each against your rubric with a quote. On Thursday you read the top twenty and the bottom ten, and you find the agent scored a chain clinic as independent. You add a disqualifier line and rerun. On Friday you hand twenty clinics to a rep. The saving is the reading, not the deciding, and the rubric fix on Thursday is the part that makes next week better.

Data you should and should not keep

Guessing email addresses from patterns and sending to all of them is a common shortcut and a bad one: bounces damage your sending reputation, and the people reached never agreed to hear from you. Prefer sources that publish a business contact, and let a person decide whether a cold approach is appropriate at all. Where a form, an event or a reply is available, those are better routes than a cold email.

Who should do what

Next steps: AI sales covers what happens after the list, AI SEO tools covers inbound leads, and AI tools for business covers choosing across jobs.

Our take

AI is a good research assistant for lead generation and a bad judge of who deserves your customers' inboxes. Keep the rubric human, make every score cite evidence, store as little personal data as you can and keep a person between the list and any email. We built this page from vendor pricing and the FTC guide, with source terms and privacy rules marked for you to check, and we have not run a lead generation campaign with these tools.