> Enrich.chat gives AI agents verified work emails and mobiles. To use it from an agent, read https://enrich.chat/agents.md . This is the Markdown version of https://enrich.chat/blog/ai-email-finder/

# AI email finder: what the AI does, where the email comes from, and what a miss costs

The label covers search boxes, domain guessers and chat agents. We read eight products' own pages on October 16, 2026 to see what the AI does, where each address comes from and what happens when nothing is found.

By Enrich.chat Editorial. Updated October 16, 2026.

**Short answer:** An AI email finder is an email finder with an AI layer on top. The AI writes the query or the summary, while the address comes from a data source, a pattern guess that gets checked, or a model's memory. Memory is the weak one: in [Hunter's own 2026 study](https://hunter.io/blog/llms-get-email-data-wrong/), 76% of the addresses language models produced were invalid.

## Key takeaways

- In an AI email finder, the AI writes the query or the summary. The address comes from a data source, a pattern guess that gets checked, or a model's memory.
- Of eight products read on October 16, 2026, only Clay describes its AI as finding the email itself, by reading the web. Elsewhere the AI searches, tags or writes, and Kaspr's pages make no AI claim.
- Asked from memory, models do badly. In a 2026 workshop paper, Gemini answered 477 of 500 cold requests and got 82 wrong. In Hunter's vendor study, 76% of 1,604 checked addresses were invalid.
- Hunter, Voila Norbert, Skrapp and Prospeo each say a search with no usable email costs nothing, and Apollo says it charges only for emails it can verify. Kaspr's pages do not say.
- Through an agent, Enrich.chat returns a work email as verified, catch_all or unknown. A pattern guess comes back in a separate guess field that the agent must not present as an email.

## What is an AI email finder?

An AI email finder is an email finder with an AI layer on top. The AI writes the search or the summary. The address itself comes from somewhere else: a data source, a pattern guess that gets checked, or the memory of a language model. Figure 1 shows what each source can prove.

**Figure 1: Where the AI sits and where an email address comes from.** An AI layer at the top writes the query and the summary. Below it are three places an address can come from. A data source gives sources or a status and should be shown as found, with its status. A pattern guess applies a format such as first.last to a name, has no proof until it is checked, and should be shown as a guess kept apart. Model memory gives whatever the model recalls or invents, has no proof, looks real and often bounces, and should never be shown as data.

*The AI layer and the three places an address can come from. Editorial diagram. The memory risk is described in [Kalai et al., 2025](https://arxiv.org/abs/2509.04664), and the found-versus-inferred split in [Hunter's API reference](https://hunter.io/api-documentation/v2).*

The label is loose. [Clay's explainer](https://www.clay.com/blog/ai-email-finder), dated April 21, 2026, says an AI email finder uses machine learning and predictive analytics to find and check addresses. Product pages say less. On October 16, 2026, neither [Kaspr's homepage](https://www.kaspr.io/) nor its [pricing page](https://www.kaspr.io/pricing) made an AI claim. Kaspr ranks first for the phrase with [a blog post](https://kaspr.io/blog/email-finder-tools) that lists 19 tools. Voila Norbert ranks second. Its [finder page](https://www.voilanorbert.com/email-finder/) does not use the term, and its [homepage](https://www.voilanorbert.com/) mentions AI only for email writing in its sequences product.

So the phrase is a search keyword more than a product type. Two questions are more useful. What does the AI do in this product, and where does each address come from? The rest of this page answers both for eight products, then shows how an agent can use Enrich.chat without presenting a guess as an email. For the wider market, see our guides to [AI lead finders](https://enrich.chat/blog/ai-lead-finder/) and [AI prospecting tools](https://enrich.chat/blog/ai-prospecting-tools/).

## Can an AI model find an email address from memory?

Sometimes, for people with a large web footprint. But a model also writes addresses that look right and are not, and it does not tell you which is which. A model's memory is not a data source, so no product should pass you an address from it unchecked.

Researchers have known for years that models keep some of their training text. A [2020 paper by Carlini and colleagues](https://arxiv.org/abs/2012.07805) pulled hundreds of verbatim passages out of GPT-2, including names, phone numbers and email addresses. A [2022 study by Huang, Shao and Chang](https://arxiv.org/abs/2205.12628) found that models leak personal information through memorization but are weak at linking a name to its address. [Kalai and colleagues](https://arxiv.org/abs/2509.04664) argue that models guess because training and testing reward a guess over admitting uncertainty.

A [workshop paper from February 27, 2026](https://www.ndss-symposium.org/wp-content/uploads/madweb2026-20.pdf) shows both sides. Amsalem and Berger asked three chat apps 500 times each for the address of 100 security researchers, with no web page supplied. Figure 2 shows the result. ChatGPT declined 335 times, answered correctly 156 times and wrongly 9 times. Gemini answered almost every request: 395 right and 82 wrong. Claude declined 452 times. The paper is work in progress, covers people who publish their addresses and tested one model setting each, so it does not predict results for ordinary sales contacts.

**Figure 2: What language models returned when asked for email addresses.** Two bar charts. The top chart shows 500 cold requests per model for the address of security researchers, with no web page supplied. ChatGPT with GPT-5 was correct 156 times, wrong 9 times and declined 335 times. Gemini with Flash 2.5 was correct 395 times, wrong 82 times and declined 23 times. Claude with Sonnet 4.5 was correct 46 times, wrong 2 times and declined 452 times. The bottom chart shows Hunter vendor study results: of 1,604 addresses that models produced and Hunter checked, 76 percent were invalid, 10 percent were accept-all and 14 percent were valid.

*Top: requests for the addresses of 100 security researchers who publish them, five trials each, from [Amsalem and Berger, MADWeb 2026](https://www.ndss-symposium.org/wp-content/uploads/madweb2026-20.pdf), a work-in-progress paper. Bottom: [Hunter's study](https://hunter.io/blog/llms-get-email-data-wrong/), run by a vendor that sells email data and checked with its own verifier. The two charts use different tests and are not comparable.*

Hunter ran a larger test. [Its study, updated September 1, 2026](https://hunter.io/blog/llms-get-email-data-wrong/), set 305 "find this person's email" tasks at 125 real companies to six API models and three open-weight models, plus a smaller hand-run test of four chat apps. It then ran each returned address through Hunter's Email Verifier. Of the 1,604 addresses that got a firm verdict, 76% were invalid, 10% were accept-all and 14% were valid. GPT-5.5, the strongest model in the test, returned an address for 42% of asks and a valid one for 19.5%. With Hunter's tools connected, a valid address came back for 78.1%.

Hunter sells email data and used its own verifier, so read this as a vendor study. It measured Hunter's connector and says nothing about any other tool, including ours. Its closing advice still holds for any stack: let the model do the thinking, and take the address from a source. Our guide on [whether ChatGPT can find email addresses](https://enrich.chat/blog/can-chatgpt-find-email-addresses/) covers the three routes it has.

## Where do the emails come from in products sold as an AI email finder?

In the pages we read, the address comes from a data source or from a pattern guess that gets checked. The AI sits on the search, the tagging, the domain detection or the outreach writing. The table sets each vendor's AI claim beside its own account of where addresses come from. These are the vendors' statements, not our findings.

*What each vendor says its AI does and where its addresses come from, from its own pages read on October 16, 2026.*

| Product | What the vendor calls AI | Where the address comes from, per the vendor |
| --- | --- | --- |
| [Hunter](https://hunter.io/pricing) | An AI Assistant that searches its company database, an AI writing assistant and an MCP server | Public web pages it crawls. With no public source, it builds an address, tags it Inferred and verifies it |
| [Kaspr](https://www.kaspr.io/pricing) | None on its homepage or pricing page | A LinkedIn browser extension and, it says, more than 150 sources |
| [Voila Norbert](https://www.voilanorbert.com/pricing/) | AI email writing in its sequences product. None on the finder page | Not stated. You enter a first name, last name and domain |
| [Apollo](https://www.apollo.io/ai/assistant) | An AI Assistant that builds lists, researches accounts and writes sequences | Apollo's own contact database. It says it charges only for emails it can verify |
| [Skrapp](https://skrapp.io/pricing) | AI tagging of seniority, department and buyer intent, and an AI Enrichment plan feature | Not stated. The finder takes a name and a domain and labels results Valid, Catch-all, Invalid or Unknown |
| [Prospeo](https://prospeo.io/pricing) | AI Search, AI Lookalikes and AI data fields | Its own database. Its MCP page says each record goes through a 5-step verification |
| [Clay](https://www.clay.com/blog/ai-email-finder) | Claygent, an AI agent that researches the web and, Clay says, can find emails from a prompt | Claygent's web research, or a waterfall across more than 50 data providers |
| [n8n template](https://n8n.io/workflows/11681-linkedin-email-finder-with-ai-domain-detection-using-google-sheets-and-hunterio/) | Two AI agents, set up with a Gemini credential, that find a company's domain from a contact's profile text | Hunter.io's API |

**Hunter draws the clearest line.** [Its data page](https://hunter.io/our-data) says contacts come from public web pages and that the Email Finder shows whether a result has public sources or is a guess. [Its help article](https://help.hunter.io/en/articles/1844277-email-finder-find-the-email-of-a-specific-person) says an address with no source carries an Inferred tag, was built from public data at the domain level, and was then verified. [Its API reference](https://hunter.io/api-documentation/v2) goes further: the `email-finder/found` call returns only addresses found on the web, and generated addresses are never returned from it. The AI in Hunter is the [AI Assistant in Discover](https://hunter.io/pricing), which searches for companies and is capped at 10 AI searches a month on Free and 50 on Starter, plus a writing assistant.

**The n8n template shows AI doing a side job.** The [n8n template](https://n8n.io/workflows/11681-linkedin-email-finder-with-ai-domain-detection-using-google-sheets-and-hunterio/) that ranks for this phrase is titled "LinkedIn email finder with AI domain detection." Per its page, an AI agent reads each contact's profile text to find the company domain, and a second agent picks the best domain from Google search results when the first one fails. Then the workflow asks Hunter.io for the address. The AI finds the domain. Hunter finds the email. A wrong domain from the AI means a lookup at the wrong company. [Zapier's Hunter page](https://zapier.com/apps/hunter/integrations) follows the same split. Its Find Email action generates the most likely address from a domain or company name and a person's name, and its Zapier MCP option lets an AI client call Hunter actions. The email is again Hunter's.

**Clay is the one case where AI is described as finding the email.** [Clay's explainer](https://www.clay.com/blog/ai-email-finder) says Claygent can find emails from simple prompts, and that Clay can also run a waterfall across more than 50 data providers. Our [Claygent guide](https://enrich.chat/blog/claygent/) notes that Claygent reads the web and is not a contact database, so it returns no email status unless you add a validation step.

The rest use AI around the email. [Apollo's AI Assistant](https://www.apollo.io/ai/assistant) works inside Apollo's database. [Skrapp](https://skrapp.io/) uses AI to tag enriched rows. [Prospeo's MCP page](https://prospeo.io/mcp) says the point of connecting an AI tool is real records instead of hallucinated results, and Hunter's [MCP page](https://hunter.io/mcp) says the same about invented addresses. Both vendors describe the problem that Figure 2 charts.

## What do they cost, and what does a miss cost?

Five of the eight products have a monthly entry plan between $39 and $65 per account, user or seat: Skrapp, Hunter, Voila Norbert, Prospeo and Apollo. Kaspr bills €59 per user. Clay starts at $185 and n8n sells workflow executions. Hunter, Voila Norbert, Skrapp and Prospeo each say a search with no usable email costs nothing. Apollo says it charges only for emails it can verify. Kaspr's pages do not say.

*Entry price on monthly billing, free allowance and miss rule, from each vendor's own pages, checked October 16, 2026. Dollars unless marked. Enrich.chat is our product.*

| Product | Entry plan, billed monthly | Free allowance | A miss, per the vendor |
| --- | --- | --- | --- |
| [Hunter](https://hunter.io/pricing) | Starter $49 a month, 2,000 credits, unlimited users | 50 credits a month | No credit when no email is found |
| [Kaspr](https://www.kaspr.io/pricing) | Starter €59 per user a month, B2B email credits listed as unlimited under a fair use policy | 15 B2B email credits a month | Not stated on its pricing or credits pages |
| [Voila Norbert](https://www.voilanorbert.com/pricing/) | Valet $49 a month, up to 1,000 leads | 50 credits once, at sign-up | Not counted: it counts only successful finds |
| [Apollo](https://www.apollo.io/pricing) | Basic $65 per seat a month, 2,500 credits | 75 credits per seat a month | Credits only for emails Apollo can verify |
| [Skrapp](https://skrapp.io/pricing) | Professional $39 a month, 2,000 credits, 2 users | 50 credits, 1 user | Not charged for Invalid or Unknown results |
| [Prospeo](https://prospeo.io/pricing) | Starter $49 per user a month, 2,000 credits | 100 credits a month | 1 credit per verified email, 0 if none is found |
| [Clay](https://www.clay.com/pricing) | Launch $185 a month, 2,500 data credits and 15,000 Actions | 100 data credits and 500 Actions a month | Clay's blog says waterfall data credits go only to successful searches. Each AI prompt uses an Action |
| [n8n template](https://n8n.io/pricing/) | n8n Starter 20€ a month billed annually for 2,500 executions, plus Hunter credits and your own Gemini and Google keys | Hunter's 50 credits a month. n8n lists a free trial and a self-hosted edition | Hunter: no credit when no email is found. The AI step still runs |
| Enrich.chat | Growth $29 a month for 100,000 credits, unlimited seats | 10 lookups a day, no signup | 0 credits when found is false |

The AI has its own meter, separate from the email. Apollo says its AI Assistant [does not use credits itself](https://www.apollo.io/ai/assistant), but actions it runs, such as enrichment, may. Hunter caps AI searches by plan. Clay charges an Action for each AI prompt plus data credits that depend on the model, per its [AI pricing doc](https://university.clay.com/docs/ai-pricing). The n8n template needs your own Gemini credential, so that usage lands on your own Google account. Our [AI lead finder guide](https://enrich.chat/blog/ai-lead-finder/) compares two-meter pricing in more detail, and [best email finder tools](https://enrich.chat/blog/best-email-finder-tools/) prices a month of work emails across nine finders.

Credits are the small cost of a miss. The large one is a bad address that gets sent. In [Hunter's own test](https://hunter.io/email-verification-guide/accept-all-catch-all/), 100 accept-all addresses and 100 valid ones got the same campaign. The bounce rate was 27% for the first group and 1% for the second. Hunter suggests keeping accept-all addresses to 2% to 5% of a campaign. An unlabelled guess has had no check at all. Our guides to [bounce rates](https://enrich.chat/blog/email-bounce-rate/) and [catch-all addresses](https://enrich.chat/blog/what-is-a-catch-all-email/) explain the limits that mailbox providers set.

## How can you tell whether an AI email finder is guessing?

Ask five questions before you trust a product or write an agent prompt. A good finder answers them in each response, not in its marketing.

1. **Does each address say where it came from?** Hunter shows source pages and dates, or an Inferred tag. The other vendors' pages we read do not describe a source for each address.
2. **Does it carry a status you can act on?** Skrapp labels results Valid, Catch-all, Invalid or Unknown and charges only for the first two. Hunter returns valid, accept_all or unknown with a confidence score. Enrich.chat returns verified, catch_all or unknown.
3. **Is a guess kept apart from a find?** Hunter does it with a separate call. Enrich.chat does it with a separate `guess` field. If a tool mixes both in one list, ask the vendor how to tell them apart.
4. **What does a miss cost?** The table above gives each vendor's answer. If a page does not say, run one lookup that cannot match, such as a made-up name at example.com, and compare your balance before and after.
5. **Can the model change the address?** If your agent copies the address field from the tool result, it cannot alter it. If it retypes the address into a summary, it can. Tell it to quote the field, and strip the guess in code before the model sees it, as our [lead generation agent guide](https://enrich.chat/blog/ai-agent-for-lead-generation/) does.

Then check five people whose addresses you already know. The [15-minute test](https://enrich.chat/blog/ai-lead-finder/) in our AI lead finder guide shows how, and the free [catch-all checker](https://enrich.chat/tools/catch-all-checker/) shows whether a domain accepts every address.

## How does Enrich.chat return an email without presenting a guess as data?

Enrich.chat is the data step. Your agent, such as ChatGPT, Claude or Cursor, writes the query and the summary. `GET /v1/email` takes a name and a company website and returns either a work email with a status label or `found: false`. If a pattern guess exists, it comes back in its own `guess` field and never in `work_email`. Enrich.chat has no chat window or agent of its own.

Each work email carries one of three labels: `verified`, `catch_all` or `unknown`. A label is not a delivery guarantee, so treat every contact as a strong lead. The table shows what each response means and what an agent may say about it.

*Responses from `GET /v1/email` and what an agent may report, from the Enrich.chat agents.md and openapi.json.*

| Response | What it is | What the agent says | Credits used |
| --- | --- | --- | --- |
| found true, status verified | A work email labelled verified | The address, labelled verified, plus a note that it is a label and not a delivery guarantee | 10 |
| found true, status catch_all | The domain accepts every address, so the mailbox cannot be confirmed | The address, labelled risky | 10 |
| found true, status unknown | A work email that is not verified | The address, labelled unconfirmed | 10 |
| found false, with a guess field | A pattern guess. It is not data | Not found. The guess is left out or listed apart as an unconfirmed guess | 0 |
| found false, no guess | No result | Not found | 0 |

Here are two requests for fictional people on example.com. The responses are illustrative.

```
GET https://enrich.chat/v1/email?name=Marta%20Okafor&domain=pier.example.com

{
  "found": true,
  "work_email": {"value": "marta.okafor@pier.example.com", "status": "verified"},
  "pattern": "first.last@pier.example.com",
  "name": "Marta Okafor",
  "title": "Head of Operations",
  "credits_used": 10,
  "free_lookups_remaining": 9
}

GET https://enrich.chat/v1/email?name=Jonas%20Weir&domain=pier.example.com

{
  "found": false,
  "guess": {"email": "jonas.weir@pier.example.com", "basis": "first.last format"},
  "credits_used": 0,
  "free_lookups_remaining": 9
}
```

The first response is an email the agent can report with its label. The second is a miss. The guess is the company's most common address format applied to the name, as [agents.md](https://enrich.chat/agents.md) describes it, and it is not a find. The agent writes not found, and either leaves the guess out or lists it apart as an unconfirmed guess. Figure 3 marks the fields.

**Figure 3: Two Enrich.chat email responses and what an agent reports for each.** Two illustrative responses from GET /v1/email for fictional people. In the first, found is true, the work email marta.okafor at pier.example.com has the status verified, which is a label and not a guarantee, the pattern is first.last, and credits_used is 10. The agent reports the address with its status label. In the second, found is false, and a guess field holds jonas.weir at pier.example.com marked as not data, and credits_used is 0. The agent reports not found and keeps the guess out of the list.

*Two illustrative `GET /v1/email` responses for fictional people on example.com. Field names follow [agents.md](https://enrich.chat/agents.md). The status is a label, not a delivery guarantee.*

Credits follow the same split. A work email found uses 10 credits and a lookup that finds nothing uses 0. A list of 40 people where 32 return an address uses 320 credits. That fits the Growth plan at $29 a month, with unlimited seats and no annual contract. The free tier gives 10 lookups a day with no signup, and a miss does not count against them. See [pricing](https://enrich.chat/pricing/) for every plan.

Paste this prompt into an agent that can reach Enrich.chat. The lines about not found, the guess field and memory do the work. In Hunter's study, a prompt that told the model not to leave the email blank cut refusals to 1%. A model told to fill every row will fill every row.

```text
Use www.enrich.chat to find the work email for each person below from their name and company website: [one name and website per line]. Send one request at a time, about one a second. For each person, show the address exactly as returned, its status (verified, catch_all or unknown) and credits_used. If found is false, write not found. If a guess field comes back, do not present it as an email that was found: list it apart as an unconfirmed guess, or leave it out. Never type an address from memory or build one from a pattern. End with the total credits_used.
```

To connect, add the MCP server at `https://enrich.chat/mcp` to your client, or let an agent that can send web requests read [enrich.chat/agents.md](https://enrich.chat/agents.md). The [install page](https://enrich.chat/install/) has steps for each app, and the free tier needs no key. For one lookup in your browser, use the free [email finder](https://enrich.chat/tools/email-finder/). The [email finder API](https://enrich.chat/api/email-finder/) page lists every field, and the [email format finder](https://enrich.chat/tools/email-format-finder/) shows a company's address pattern. Enrich.chat never logs in to LinkedIn, and a LinkedIn URL works as a lookup key on `/v1/person`.

## What can't Enrich.chat do, and when is another tool the better choice?

Enrich.chat is a data layer for your own agent, not a hosted AI email finder. These are its gaps:

- **No standalone verifier.** It labels the work emails it finds. It cannot check an address you already hold, which Hunter's Email Verifier does. See our guide to [email verification tools](https://enrich.chat/blog/email-verification-tools/).
- **No source pages or score.** A response carries a status label, not source links, dates or a confidence score like Hunter's.
- **A website, not a name.** The email lookup needs a company website such as pier.example.com. A company name alone is not supported, and the email input on `/v1/person` takes work emails only, not personal mailboxes such as Gmail.
- **No app, extension or sending.** There is no Chrome extension, saved list, bulk export on self-serve plans, sequencing or AI writing. Pace is about one request a second.

Another product is the better choice in these cases:

- **Hunter** when you want source pages and dates, a verifier for any address, and sequences in one app. See [what Hunter is](https://enrich.chat/blog/what-is-hunter-io/) and [Hunter alternatives](https://enrich.chat/alternatives/hunter-io/).
- **Apollo** when reps want a database, an AI Assistant and sequences in one seat. Its Free plan lists 5 AI Assistant chats a month. See [Apollo pricing](https://enrich.chat/blog/apollo-pricing/).
- **Clay** when you want a table with a web-research agent and several data providers behind it. See [Clay pricing](https://enrich.chat/blog/clay-pricing/).
- **Voila Norbert** when you want a plain capped plan with bulk CSV search and unlimited team members.
- **Prospeo** when you want AI search and enrichment per user. See [Prospeo pricing](https://enrich.chat/blog/prospeo-pricing/).
- **Skrapp** when you want a two-user plan at $39 a month with credits that roll over, per its pricing page.
- **The n8n template** when you already run n8n and want an AI step to find missing domains before a Hunter lookup. See our [n8n lead enrichment guide](https://enrich.chat/blog/n8n-lead-enrichment/).

You stay responsible for the law that applies to your outreach, such as CAN-SPAM, GDPR and TCPA. This is general information, not legal advice. Anyone can ask Enrich.chat to remove their data at [enrich.chat/opt-out/](https://enrich.chat/opt-out/). For more on AI around the email, read [AI prospecting tools](https://enrich.chat/blog/ai-prospecting-tools/) and [Claude Code for sales](https://enrich.chat/blog/claude-code-for-sales/).

## Frequently asked questions

### What is an AI email finder?

An AI email finder is an email finder with an AI layer on top. The AI writes the search, tags the results or writes the summary. The address itself comes from a data source, a pattern guess that gets verified, or a language model's memory. Memory is the weak option. Ask any product which source it used, and what it charges when it finds nothing. For the lead side, see our guide to [AI lead finders](https://enrich.chat/blog/ai-lead-finder/).

### Can I use AI to find email addresses?

Yes, but let the AI call a tool instead of recalling an address. A model asked from memory sometimes returns a correct address for a well-known person and sometimes one that looks right and bounces. In [Hunter's own study](https://hunter.io/blog/llms-get-email-data-wrong/), 76% of the addresses models produced were invalid. Connect your assistant to a data tool and tell it to write not found on a miss. See [can ChatGPT find email addresses](https://enrich.chat/blog/can-chatgpt-find-email-addresses/).

### Is there a free AI email finder?

Several products have a free allowance, as of October 16, 2026. [Hunter](https://hunter.io/pricing) gives 50 credits a month, [Skrapp](https://skrapp.io/pricing) 50 credits, [Prospeo](https://prospeo.io/pricing) 100 credits a month, [Apollo](https://www.apollo.io/pricing) 75 credits per seat a month and [Voila Norbert](https://www.voilanorbert.com/pricing/) 50 credits once. Enrich.chat gives 10 lookups a day with no signup. The units differ, so compare what one found email costs.

### Do AI email finders charge when they cannot find an email?

Most say no. On October 16, 2026, [Hunter](https://hunter.io/pricing), [Voila Norbert](https://www.voilanorbert.com/pricing/), [Skrapp](https://skrapp.io/pricing) and [Prospeo](https://prospeo.io/pricing) each said a search with no usable email costs nothing, and [Apollo](https://www.apollo.io/pricing/about-credits) said it charges only for emails it can verify. [Kaspr's](https://www.kaspr.io/pricing) pricing page does not say. Enrich.chat uses 0 credits when found is false.

### Can an AI email finder find a Gmail address?

Most B2B finders target work addresses. Enrich.chat does not look up personal mailboxes such as Gmail, and the email input on its person lookup takes work emails only. A chat model asked for a personal address is guessing, because nothing in the model checks that the mailbox exists. Some vendors sell personal emails as a separate credit, such as [Kaspr's direct email credits](https://help.kaspr.io/en/articles/8865905-what-are-credits). Read the terms and the law that applies to you. This is not legal advice.

### How do I stop my AI agent from making up emails?

Tell the agent to take every address from a tool and to write not found on a miss. Do not push it to fill every row: in [Hunter's study](https://hunter.io/blog/llms-get-email-data-wrong/), a prompt that said not to leave the email blank cut refusals to 1%. Keep any guess field out of the final list. The prompt on this page does that for Enrich.chat, and our [lead generation agent guide](https://enrich.chat/blog/ai-agent-for-lead-generation/) enforces it in code.

## Sources

1. [LLMs Get 76% Of Emails Wrong, But Hunter Can Easily Fix It](https://hunter.io/blog/llms-get-email-data-wrong/)
2. [Hunter plans and pricing](https://hunter.io/pricing)
3. [Hunter MCP](https://hunter.io/mcp)
4. [Hunter: our data](https://hunter.io/our-data)
5. [Email Finder: find the email of a specific person](https://help.hunter.io/en/articles/1844277-email-finder-find-the-email-of-a-specific-person)
6. [Hunter API reference v2](https://hunter.io/api-documentation/v2)
7. [What are accept-all email addresses](https://hunter.io/email-verification-guide/accept-all-catch-all/)
8. [19 of the best email finder tools for prospecting](https://kaspr.io/blog/email-finder-tools)
9. [Kaspr plans and pricing](https://www.kaspr.io/pricing)
10. [Kaspr homepage](https://www.kaspr.io/)
11. [What are credits? (Kaspr Help Center)](https://help.kaspr.io/en/articles/8865905-what-are-credits)
12. [Voila Norbert email finder](https://www.voilanorbert.com/email-finder/)
13. [Voila Norbert pricing](https://www.voilanorbert.com/pricing/)
14. [Voila Norbert homepage](https://www.voilanorbert.com/)
15. [Apollo pricing](https://www.apollo.io/pricing)
16. [Apollo AI Assistant](https://www.apollo.io/ai/assistant)
17. [About Apollo credits](https://www.apollo.io/pricing/about-credits)
18. [Skrapp homepage](https://skrapp.io/)
19. [Skrapp pricing](https://skrapp.io/pricing)
20. [Prospeo pricing](https://prospeo.io/pricing)
21. [Prospeo MCP server](https://prospeo.io/mcp)
22. [AI Email Finder: benefits, tips and how it works](https://www.clay.com/blog/ai-email-finder)
23. [Clay pricing](https://www.clay.com/pricing)
24. [How AI is priced](https://university.clay.com/docs/ai-pricing)
25. [LinkedIn email finder with AI domain detection using Google Sheets and Hunter.io](https://n8n.io/workflows/11681-linkedin-email-finder-with-ai-domain-detection-using-google-sheets-and-hunterio/)
26. [n8n plans and pricing](https://n8n.io/pricing/)
27. [Hunter integrations](https://zapier.com/apps/hunter/integrations)
28. [Work-in-progress: Deobfuscating Academic Email Addresses: A Security Evaluation of LLMs](https://www.ndss-symposium.org/wp-content/uploads/madweb2026-20.pdf)
29. [Are Large Pre-Trained Language Models Leaking Your Personal Information?](https://arxiv.org/abs/2205.12628)
30. [Extracting Training Data from Large Language Models](https://arxiv.org/abs/2012.07805)
31. [Why Language Models Hallucinate](https://arxiv.org/abs/2509.04664)

## Try this prompt

Find work emails without guesses

```text
Use www.enrich.chat to find the work email for each person below from their name and company website: [one name and website per line]. Send one request at a time, about one a second. For each person, show the address exactly as returned, its status (verified, catch_all or unknown) and credits_used. If found is false, write not found. If a guess field comes back, do not present it as an email that was found: list it apart as an unconfirmed guess, or leave it out. Never type an address from memory or build one from a pattern. End with the total credits_used.
```

## Related

- [AI lead finder: what it can and cannot do](https://enrich.chat/blog/ai-lead-finder/)
- [Best email finder tools (2026): 9 compared on price and AI agent access](https://enrich.chat/blog/best-email-finder-tools/)
- [Can ChatGPT find email addresses?](https://enrich.chat/blog/can-chatgpt-find-email-addresses/)