Why does my AI agent make up emails, and how do I stop it?

Ask an AI agent for the work email of a named person and, if it has no tool for the job, it will often give you one anyway. The address looks right. It follows the company's pattern. It may even be correct. But the agent did not find it, it composed it.

That difference matters the moment you send. This post explains why agents guess, how to tell a found address from a guessed one, and the instruction that makes the agent call a finder instead.

Why does an AI agent make up an email address?

An agent answers from what its model has learned and from the tools it can call. With no email tool in reach, the only way to answer is to produce text that fits the request. A well-formed address at the right domain fits the request, so the agent produces one.

Nothing in that path checks that the mailbox exists. The model has no signal that it is guessing, so it reports the address in the same confident tone as a fact.

How can I tell a found email from a guessed one?

Ask where the address came from and what its status is. A found address comes with a source and a verification status. A guessed address has neither.

A real finder returns a status such as deliverable, risky or unknown, and often a confidence score. If your agent gives you an address with no status, treat it as a guess. A company's email format is a useful clue, but a pattern is not a confirmed mailbox.

What is the fix?

Give the agent a tool for the job and an instruction that makes it use the tool. The tool is a finder. The instruction is a skill. With looot, the finder is a catalog endpoint the agent reads the price of before it runs.

The skill matters as much as the tool. A tool the agent is free to skip is a tool it will sometimes skip. The skill says: ask for the name and the domain, call the finder, read the status, and say so when nothing came back.

What does the skill tell the agent to do?

Four rules, taken from the professional email skill.

1. Ask for the full name and the company domain, and ask again if only the company name was given.
2. Read the price of the endpoint and tell the user before the first run.
3. Run the finder and read the run status, not the web status. A run can fail with a normal response.
4. Report the email, its verification status and the confidence score. Never present an unverified address as confirmed.

When the first provider fails, the skill moves to the next one. When the input was wrong, it fixes the input and does not retry elsewhere.

Should I verify an address I found?

Yes, for a list you plan to send to. A found address has been seen or inferred by a provider, which is not the same as a mailbox that accepts mail today. A verification skill groups addresses as deliverable, risky and undeliverable before you send.

Try it: a prompt for your agent

What it cannot do

• It cannot find an address that no provider has. A person with no public footprint returns nothing.
• It does not make a found address current. Mailboxes change when people move jobs.
• It does not decide whether you should email someone. Cold email rules vary by country, and that is yours to check.

Last checked 2026-10-04 against the looot skill library. Author: the looot team.