# How to Build Prospect Lists With AI Agents (Without Burning Your Domain)

> How to build prospect lists with AI agents, step by step. ICP, finding contacts, email verification, scoring, and the legal and bounce risks to avoid.

Last updated: 2026-09-28T10:00:00

Canonical URL: https://crevio.co/blog/build-prospect-lists-with-ai-agents

*Last updated: September 2026*

**Ask an AI model for the email addresses of 50 prospects and roughly three in four will bounce.** That is the headline of [Hunter's test of 13 AI models and apps](https://hunter.io/blog/llms-get-email-data-wrong): 76% of the outreach addresses the models produced on their own were invalid. Hunter sells email data, so weigh the source, but the pattern is easy to check yourself: ask a chatbot for a prospect's email and run it through a verifier. The same test found that when a model could call a real email-finding tool, its hit rate jumped from about 20% to 78%.

That gap is the whole lesson of this guide. You can build prospect lists with AI agents, and it can save you hours every week, but only when the agent does the research, sorting, and bookkeeping while real data tools supply the contact details. This guide is for founders and small B2B teams who do their own outbound and want a list that is accurate, legal, and won't wreck the domain they send from.

- **The agent is the researcher, not the database.** It reads company websites, applies your criteria, fills the spreadsheet, and removes duplicates. Names and emails come from a data provider
- **Every email gets verified before it is sent,** because B2B contact data goes stale at roughly [22.5% a year](https://www.hubspot.com/database-decay), and Gmail starts filtering senders whose spam rate reaches [0.3%](https://support.google.com/a/answer/81126?hl=en)
- **Cold B2B email is legal in the US and the UK if you follow the rules,** but scraping LinkedIn breaks its User Agreement, and a US court has enforced that
- **Start small:** 50 rows, reviewed by you, beats 5,000 rows nobody checked

## What an AI Agent Actually Does When It Builds a Prospect List

AI prospect list building is not one feature. It is a loop of small, boring jobs: search a directory, open a company website, decide whether the company fits, look up who runs sales there, paste the result into a sheet, check it against last month's sheet. Done by hand, it takes minutes per row. An AI agent with a web browser and access to your tools does the same loop without getting bored, and keeps a source link for every decision it makes.

What it should not do is *know* things. A language model has read a lot of the internet, so it will happily produce a plausible name and a plausible email like `jane.miller@acme.com`. Sometimes Jane exists, sometimes she left in 2023, and sometimes she never existed at all. The fix is structural, not a better prompt: the agent is only allowed to write down a contact detail that a tool returned, and it has to record which tool.

![Matrix of the seven prospect list steps showing which ones the AI agent owns, which ones a data tool or verifier owns, and which ones you own](https://crevio.co/vite/assets/prospect-list-pipeline-d2wv84vc.svg)

## Step 1: Write an ICP Brief the Agent Can Check

Your ideal customer profile (ICP) is usually a sentence like "B2B SaaS companies that need help with onboarding." That works for a person who can use judgment. An agent needs criteria it can check against a web page, one row at a time. Write your ICP brief with three parts:

- **Must-haves** it can verify: industry, country, headcount range, a product or tool they use, a page that must exist (a pricing page, a careers page)
- **Disqualifiers:** existing customers, competitors, agencies if you only sell direct, companies under a size you can serve
- **One timing signal:** hiring for a role you help with, a recent funding round, a new product launch, a job post that mentions a tool you replace

Here is a brief that works:

> Find US and UK software companies with 20 to 200 employees that sell to other businesses, have a public pricing page, and are hiring a customer success or onboarding role right now. Exclude agencies and anyone on the attached customer list. For each company, record the website, the headcount source, the job post URL, and one sentence on why it fits.

Notice the last sentence. Asking for a source link and a reason on every row is what lets you audit the list in five minutes instead of trusting it blindly. Our guide to [giving an AI agent a knowledge base](/blog/ai-agent-knowledge-base) covers where a brief like this should live so the agent reads it on every run.

## Step 2: Find Companies That Match

There are two ways to find companies, and the best lists use both.

**Database filters** are fast and broad. [Apollo](https://www.apollo.io/) lets you filter a database it says covers 240 million contacts and 30 million companies by industry, headcount, location, technology, and hiring activity. The catch is that every competitor filtering for the same buyer gets the same rows, and those people already get a lot of cold email.

![Apollo homepage, a sales platform with a large B2B contact database and a free plan](https://crevio.co/vite/assets/apollo-homepage-bqzxjdpf.png)

**Live web research** is where an agent earns its keep. Niche markets, local businesses, and anything that depends on a timing signal are poorly covered by databases. An agent can work through an industry association's member directory, a conference exhibitor list, a marketplace's partner page, or job boards, open each company's website, and check it against your brief. It is slower per row than a database export, but the rows are usually fresher and less contacted.

A good pattern: pull a broad first cut from a database, then let the agent visit each company's site to confirm the must-haves and find the timing signal. Rows that fail the check get dropped with a reason, not silently.

## Step 3: Find the Right People, Not a Guessed Name

Once you have companies, you need people. Decide the roles first: for most small B2B offers, that is the founder at companies under 30 people, and the head of the relevant function above that. Two contacts per company is plenty; ten makes you look like spam.

Then have the agent pull those people from a data tool, never from its own memory. [Hunter](https://hunter.io/) is the simplest option here: give it a company domain and it lists the people it has public email addresses for, with their role, department, and seniority, plus the pages where each address was found. Its free plan includes 50 credits a month, which is enough to test the workflow on a small list.

![Hunter homepage, an email finder and outreach platform with a free plan of 50 leads a month](https://crevio.co/vite/assets/hunter-homepage-fpbpr5ae.png)

If the data tool has nobody for a company, the right answer is "not found," not a guess. Tell the agent that explicitly in your brief. It is one line, and it removes the worst failure mode on this list.

## Step 4: Enrich and Verify Every Email

Enrichment fills the gaps: job title, LinkedIn profile URL, company size, location, tech stack. Verification answers one question: will this address accept mail? They are different jobs, and skipping the second is how people burn their sending domain.

### Enrichment with a waterfall

No single provider has everyone. [Clay](https://www.clay.com/) is built around this problem: it is a spreadsheet where each column can call a different data source, and a "waterfall" tries one provider after another until one returns a match. Clay says it connects to more than 150 providers. It has a free plan with 100 data credits a month and paid plans from $167 a month, per its [pricing page](https://www.clay.com/pricing). For a small team, the useful idea is the waterfall itself, even if you run it with two tools instead of twenty.

![Clay homepage, a data enrichment platform that runs waterfall lookups across many data providers](https://crevio.co/vite/assets/clay-homepage-c0389j14.png)

### Verification is not optional

A verifier connects to the recipient's mail server and checks whether the mailbox exists, without sending anything. [ZeroBounce](https://www.zerobounce.net/) is one of the established options: 100 free checks a month, and pay-as-you-go starts at $39 for 2,000 addresses, according to its [pricing page](https://www.zerobounce.net/email-validation-pricing). Hunter and Apollo both include verification in their plans as well.

![ZeroBounce homepage, an email verification service with a free email verifier](https://crevio.co/vite/assets/zerobounce-homepage-ben7pdey.png)

Verifiers return more than yes or no. The statuses you will see, and what to do with each:

| Status | What it means | What to do |
|--------|---------------|-----------|
| Valid | Mailbox exists and accepts mail | Send, in small batches |
| Invalid | Mailbox does not exist | Delete the row |
| Catch-all (accept-all) | Server accepts everything, so the check is inconclusive | Hold back, or send a very small test batch |
| Unknown | Server did not answer | Re-check in a day, never send blind |
| Disposable or role (info@, sales@) | Not a person | Drop for cold outreach |

Put the verification date in its own column. An address that was valid in March is a guess by September.

![Four levels of contact data trust, from a guessed email in model memory to an address a verifier confirmed this week, with only the top level safe to send](https://crevio.co/vite/assets/contact-data-trust-ladder-cgmbmm6d.svg)

## Step 5: Deduplicate Against Everything You Already Know

Duplicates are how you end up emailing the same person three times from two tools, or pitching a current customer. The matching is tedious for a person and ideal for an agent. Have it match on three keys, in this order:

1. **Email address,** lowercased
2. **Company domain,** stripped of `www.` and country variants
3. **Full name plus domain,** for rows where the email is still missing

Then check the new list against four sources: your existing customers, your CRM or previous prospect lists, anyone you have emailed in the last 90 days, and your unsubscribe and bounce lists. That last one matters legally, not just politely. Someone who opted out stays opted out, whichever tool you use next.

## Step 6: Score and Rank Before You Send

You will rarely email the whole list at once, so rank it. Keep the scoring simple enough that you can explain every score:

| Signal | Points |
|--------|--------|
| Meets every must-have in the brief | 3 |
| Timing signal found in the last 30 days (job post, funding, launch) | 3 |
| Contact is the decision maker, not an influencer | 2 |
| Email verified as valid this week | 2 |
| Catch-all or unverified email | minus 3 |

Anything scoring 8 or more goes in the first batch. Ask the agent to write the reason for each score in a column next to it. When a reply comes back, you can see which signals actually predicted interest, and adjust the weights after a few weeks.

## Step 7: Hand the List Off to Outreach

The handoff is where a good list gets wasted or pays off. Four rules:

- **Send from a separate domain,** like `getyourbrand.com` instead of `yourbrand.com`, so a cold campaign that goes wrong does not hurt delivery of your invoices and receipts
- **Start with 20 to 50 emails a day per mailbox,** and watch bounces and replies for a week before adding more
- **Write each first line from the row's timing signal.** "Saw you are hiring your first onboarding lead" beats any template
- **Review the first batch yourself,** row by row, before anything is sent

Most deliverability guides put the danger line at around [2% hard bounces](https://www.smartlead.ai/blog/how-to-reduce-email-bounce-rates-for-cold-outreach). Above that, pause, find out which source the bad rows came from, and fix it before sending another email.

## The Problems Nobody Puts in the Demo

Every AI prospecting demo shows a spreadsheet filling itself in. None of them show the next month. These are the parts that go wrong.

### Hallucinated people and emails

Covered above, but worth repeating because it is the failure that does the most damage. A model with no tool access fills every row confidently. A model with tool access can still paper over a failed lookup with a guess if you let it. Two safeguards: require a source column for every contact detail, and spot-check five random rows by opening the source link. If one of five is wrong, the whole list is suspect.

### Data freshness

People change jobs constantly. HubSpot's [database decay simulation](https://www.hubspot.com/database-decay), built on MarketingSherpa research, puts B2B contact decay at about 2.1% a month, or 22.5% a year. A list you built in January is roughly a fifth wrong by December. Treat a prospect list as a snapshot, re-verify before every campaign, and rebuild rather than reuse after about three months.

### Legality: CAN-SPAM, GDPR, and LinkedIn

This is not legal advice, but here is where the lines are.

- **United States.** The [CAN-SPAM Act](https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business) allows cold commercial email, including B2B, which gets no exemption. You need an accurate sender and subject line, a physical postal address, and a working opt-out honored within 10 business days. The FTC lists penalties of up to $53,088 per violating email
- **UK and EU.** Emailing someone at their work address generally relies on "legitimate interests" under GDPR, which means the email has to be relevant to their job, you must tell them where you got their data, and you must stop on request. The UK's [ICO guidance on business-to-business marketing](https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/business-to-business-marketing/) notes that sole traders and some partnerships count as individuals, so they need consent. EU countries add their own rules on top, and some, like Germany, are much stricter about unsolicited email. Check the rules for each country you target
- **LinkedIn.** Section 8.2 of LinkedIn's [User Agreement](https://www.linkedin.com/legal/user-agreement) bans using software, bots, or browser plugins to scrape or copy profiles. This is not a theoretical rule: in the long-running hiQ Labs case, a court found hiQ [breached LinkedIn's User Agreement](https://newmedialaw.proskauer.com/2022/11/11/court-finds-hiq-breached-linkedins-terms-prohibiting-scraping-but-in-mixed-ruling-declines-to-grant-summary-judgment-to-either-party-as-to-certain-key-issues/) through scraping and fake profiles, and the case ended in a [$500,000 judgment against hiQ](https://www.proskauer.com/blog/hiq-and-linkedin-reach-proposed-settlement-in-landmark-scraping-case). An agent with a browser can technically scroll LinkedIn. Don't let it. At best your account gets restricted; at worst you are the next case

![LinkedIn User Agreement section 8.2, which prohibits using software, bots, crawlers, or browser plugins to scrape or copy profiles](https://crevio.co/vite/assets/linkedin-user-agreement-n59kvjek.png)

### Bounce rates and sender reputation

Gmail's [sender guidelines](https://support.google.com/a/answer/81126?hl=en) ask every sender to keep spam complaints below 0.3% and recommend staying under 0.1%. If you send 5,000 or more messages a day to Gmail addresses, you also need SPF, DKIM, DMARC, and one-click unsubscribe. A bad list hurts twice: bounces tell mailbox providers you do not know who you are writing to, and irrelevant emails get marked as spam. Both show up in where your *next* email lands, including the ones to paying customers if you send from the same domain.

## AI Prospecting Tools Compared

| Tool | Best for | Free plan | Paid from |
|------|----------|-----------|-----------|
| [Apollo](https://www.apollo.io/pricing) | Broad database search plus built-in sequences | Yes | About $49 per user a month, billed yearly |
| [Clay](https://www.clay.com/pricing) | Waterfall enrichment across many providers | 100 data credits a month | $167 a month |
| [Hunter](https://hunter.io/pricing) | Finding and verifying emails by company domain | 50 credits a month | $34 a month, billed yearly |
| [ZeroBounce](https://www.zerobounce.net/email-validation-pricing) | Verifying a finished list | 100 checks a month | $39 for 2,000 checks |
| A general AI agent | Research, criteria checks, sheets, dedupe, scheduling | Varies | Varies |

Prices change often, so check each page before you buy. None of these replace each other. The common setup is one source of contacts, one verifier, and something (a person or an agent) doing the research and bookkeeping in between.

## How to Build a Prospect List With Crevio's AI Agent

![Crevio homepage, an AI business builder where you describe what you want and the AI builds and runs it](https://crevio.co/vite/assets/crevio-homepage-gabwec82.png)

[Crevio](https://crevio.co/) is an AI business builder: an AI agent that runs the day-to-day work of your business alongside you. For prospect list building, here is what it can do today, and what it can't:

- **Research the web.** It can search the web and use a real browser to open company sites, directories, and job boards, recording a link for each row
- **Find, verify, and enrich contacts itself.** The agent has built-in lead tools, no extra subscription needed. It can discover companies by industry, location, headcount, and technology; list the people at a company domain, filtered by department and seniority; find a named person's work email; verify an address; and enrich a person or a company with role, size, and other public details. Each lookup costs 2 AI credits
- **Use the tools you already pay for.** Crevio connects to 3,000+ apps. If your data provider or verifier is in the catalog, connect it and the agent can call it alongside the built-in lookups. By default, connected apps ask for your approval before the agent changes anything in them
- **Keep the list in a spreadsheet.** The agent can build and update spreadsheets on its own computer, and you can import prospects into Crevio as customers from a CSV
- **Write the outreach.** It has a built-in [cold email skill](/blog/ai-agent-skills) for first lines and follow-up sequences, written like a peer rather than a vendor
- **Refresh the list on a schedule.** A [recurring task](/blog/schedule-recurring-ai-agent-tasks) can re-check your criteria every Monday and add new matches

And the honest limits:

- **There is no dedicated list-building skill yet.** The ICP brief, the source-link rule, and the "not found beats a guess" rule are yours to write into the task
- **The built-in lookups are only as good as public data.** They return what they can find for a company or person, and "not found" is a common, correct answer for small firms. A lookup is also not a green light: have the agent run the verifier on every address the day you send, and hold back anything that comes back catch-all or unknown
- **It does not check whether you are allowed to email someone.** Crevio's email campaigns add unsubscribe links and skip addresses that hard bounced, marked you as spam, or unsubscribed, but there is no consent check for cold lists. The legal decision is yours
- **New tasks run Autonomous by default,** which means writes, including sending email, go ahead without asking. Switch a prospecting task to Supervised so you [approve each batch](/blog/human-in-the-loop-ai-agents), and know that a recurring Supervised task switches itself to Autonomous after 15 approved runs in a row
- **Lookups and research use AI credits.** At 2 credits per lookup, the free Starter plan's 20 AI credits a month cover about ten lookups before any browsing or writing, so treat Starter as a way to try the workflow. A weekly list of 50 verified contacts means a find and a verify per person, around 200 credits, which fits a paid plan

## FAQ

### Can AI build a B2B prospect list for me?

Yes, if it has tools. An AI agent can search the web, check companies against your criteria, pull contacts from a data provider, verify emails, remove duplicates, and rank the result in a spreadsheet. On its own, without a data tool, a model will invent plausible but wrong contacts, so never use addresses it produced from memory.

### Is it legal to email a prospect list built with AI?

In the US, cold B2B email is legal under CAN-SPAM if you use honest headers, include a postal address, and honor opt-outs within 10 business days. In the UK and EU it usually rests on legitimate interests and must be relevant to the person's job, with extra rules per country. How the list was built also matters: scraping LinkedIn breaks its User Agreement.

### How many prospects should be on my first list?

Start with 50 to 100 verified contacts at companies that match every must-have. That is enough to learn which signals get replies, small enough to review by hand, and low-risk for your sending domain. Scale only after a week of bounces under 2% and some replies.

### Which is better for prospecting, Apollo or Clay?

They do different jobs. Apollo is a contact database with outreach built in, good for a fast, broad first list. Clay is an enrichment workspace that combines many providers, good for filling gaps and building custom research columns. Many teams use Apollo to find contacts and Clay to enrich them.

## The Bottom Line

Building prospect lists with AI agents works when you split the job correctly: data tools supply the facts, the agent does the legwork, and you own the criteria and the send button. Get that split wrong and you get a beautiful spreadsheet of people who do not exist. The best prospect list is not the biggest one. It is the one where every row can tell you where it came from.

## Related Blog Posts

- [AI Agents for Marketing: What to Hand Over and Where They Fail](/blog/ai-agents-for-marketing)
- [Human in the Loop AI Agents: What to Approve and What to Let Run](/blog/human-in-the-loop-ai-agents)
- [How to Schedule Recurring AI Agent Tasks That Run Without You](/blog/schedule-recurring-ai-agent-tasks)