An AI agent finds customers by watching public sources for signals that a company needs what you sell, ranking those companies against a buyer you described, and naming the person who owns the problem. It does not guess contact details or send anything without your approval.
An AI agent finds customers by watching public sources for signals that a company needs what you sell, keeping only the companies that match the buyer you described, and naming the person who owns the problem. The finding is the work; the message is written afterwards, from evidence, and waits for your approval.
Most tools sold as lead generation are a search box over a database. You still do the hard part: deciding which of the four thousand results is worth a message this week. An agent that finds customers has to do that part, and be able to show why.
What finding a customer actually means
It is three questions in a fixed order. Does this company have the problem I solve? Is there a reason it matters this month? Who inside owns it? A list that answers only the first question is a directory.
Fit. The company sells to the people you serve, is the right size, and operates where you can deliver.
Timing. Something changed recently: a role was posted, a location opened, a review complained about the exact thing you fix.
Owner. There is a named person whose job includes the problem, not a generic inbox.
A customer is a company with a reason to talk this week, and a person who owns the reason.
Where the signals come from
A hiring post is the most honest signal available, because a company only pays for a role when the work has outgrown the team. But it is not the only one, and a good agent watches many places at once so it does not depend on a single feed.
Source type | What it tells you | Example
Job boards | Work has outgrown the team | Two operations roles open in a month
Review sites | A competitor is failing at something you do | Reviews mention slow replies
Communities | Someone is asking for a tool like yours | A thread asking for an alternative
Company sites | The offer, the pricing and the proof | A new service page
News and funding | Budget exists that did not last quarter | A round, or a new office
Every one of these is a public page. None of them requires an account, a key, or a login, which is why a source list is something you can switch on rather than a set of integrations to build. In the current build the monitor loop polls Reddit and X; the rest of the catalog is listed and switches on as each source is wired.
The loop, step by step
1. Describe the buyer once
Not a market, a person. Operations leads at 50-person logistics firms who are hiring for a second dispatcher is a buyer. Small businesses is not. The narrow version tells the agent where to look and, just as importantly, what to ignore.
2. Watch, and keep the reason
Each signal is stored with the page it came from, the quote that supports it, and the date. A signal that is a month old is not a reason to write today, so it expires. That record is also what stops the agent writing a claim it cannot support.
3. Rank against the buyer
Fit and timing are scored separately, so a perfect-fit company with no reason to act does not crowd out a good-fit company with a strong signal. The score is computed from stated criteria rather than a model's impression, which makes it reproducible and explainable.
4. Find the owner, and the address
The role that owns the problem comes from the company's own pages, a licensed provider, or the job post itself. The work email is revealed through a licensed provider under a daily budget, and verified before use. A miss costs nothing, and the same person is never charged twice in a workspace.
5. Write, then wait
The first message is built from the signal that was saved, in the channel where the person is reachable. It goes to Needs you. Nothing leaves until you approve it, and a social draft is a task you post rather than something the agent posts for you.
What it refuses to do
It does not scrape LinkedIn, Google Maps, Yelp, Facebook or Instagram, because those terms do not allow it. It does not store Google Places content. It does not guess an email address, and it does not claim to be a person. A page behind a login or a CAPTCHA is recorded as blocked and skipped rather than worked around.
Connecting it to the rest of your work
Finding customers is only useful if the result lands where you already work. Instead of an integration per app, the agent exposes one MCP endpoint. Claude, ChatGPT, Cursor, Notion, Slack and the rest can each add that server and read the same list, with one key for the workspace.
claude mcp add --transport http outbound https://app.example/mcp \
--header "Authorization: Bearer $OUTBOUND_API_KEY"
Read tools cost no actions. A tool that spends actions says so in its description. Approving a draft is a separate step from sending it.
What to expect in the first week
Day | What the agent does | What you do
1 | Reads your site, proposes the buyer and the proof | Correct two lines
2 | Watches the sources and scores the first companies | Say yes or no to examples
3 | Finds owners and verifies addresses | Approve the first drafts
4-7 | Keeps finding, keeps drafting | Review what is new each morning
The useful question after a week is not how many messages went out. It is how many of the companies it found you would have found yourself, and how many of those had a reason you did not know about.
How does an AI agent find customers? It watches public sources for changes that show a company has a problem you solve, such as a job post for a role your service replaces, a complaint in a review, or a launch. It keeps the companies whose signals match the buyer you described, then finds the person who owns that problem.
Does it invent contact details? No. A business email is revealed through a licensed provider, or read from a page the company published itself, and it is verified before any message is drafted. If no address can be verified, the person stays in the list without one rather than being guessed.
Where do the customers come from? Three places: a licensed company dataset with no contact details attached, your own lists and CRM, and businesses' own public websites and posts. Every source is recorded so you can see where a company came from, and each source has an allowed use that must include outreach.
How is this different from a lead database? A database gives you a list of companies that fit a filter. This gives you a shorter list of companies with a reason to talk this month, the source of that reason, and a draft written from it. The filtering is the product, not the raw list.
Do I have to approve every message? Yes. Every first message, whether it goes to email or a social channel, waits for you in Needs you until you approve it. The agent does not send outreach on your behalf without that click.
Can I use it with the apps I already have? Yes, over MCP. The agent exposes one endpoint, and an app that speaks Model Context Protocol, such as Claude, ChatGPT, Cursor, Notion or Slack, can read the same list and the same drafts. You paste one URL and one key instead of creating a developer account per app.