Build an AI Agent That Finds Leads and Writes the First Email

Nobody replies to a cold email that sounds like every other cold email. So the fix is not a better template. It is an agent that never uses one.

It is 9 p.m. And Dana is still at her laptop. She runs a four-person managed IT support shop in Austin, the kind of business that keeps small law firms and clinics from losing their files. Dana is good at the actual work. Fixing a broken server, migrating an office to new machines, walking a panicked assistant through a ransomware scare. What she is bad at, and what she is doing right now instead of sleeping, is finding the next client.

She has a list of forty accounting firms and dental offices in a spreadsheet. Half of them do not have an IT contract yet. She knows this because she checked their websites, one at a time, on a Tuesday afternoon she will never get back. Now she is trying to write forty emails that do not sound like the same email forty times.

Why cold outreach eats a founder's week

Here is what prospecting actually costs a small business owner, in hours, not theory:

• Find businesses that fit your ideal client. Google, LinkedIn, directories, one tab at a time.
• Find a real person and a real email address at each one. Guess, verify, guess again.
• Read enough about each business to say something true and specific.
• Write an email that does not read like a template, forty separate times.
• Track who you already emailed, so you do not send it twice.

Do that for one prospect and it is due diligence. Do it for forty and it is a part-time job you never agreed to take. You are the bottleneck. Not because you are bad at sales. Because a human being cannot personally research forty strangers before breakfast.

What is an AI outreach agent, and how is it different from a mail merge?

A mail merge fills in a name and a company. Everyone can tell. An outreach agent does the part that used to require a human: it reads about the business first, then writes.

Picture the difference on one prospect, a 12-person dental office that just opened a second location. A mail merge sends: "Hi [Name], I noticed [Company] could use better IT support." An agent sends something closer to: "Saw you opened your second location this spring. Two sites usually means two sets of scheduling software talking to each other badly. Worth a 15-minute call?" One line, one real fact, one specific ask. That is the whole trick, and it is the one thing a spreadsheet formula cannot do.

Why this is possible now, and not two years ago

This is not a hypothetical. It is what builders in the AI-agent niche are shipping and testing this week, using Claude Skills, Anthropic's system for packaging a reusable task (research this, write that, send this) so an AI agent can run it on command instead of you re-explaining it every time. We covered the basics of Claude Skills in an earlier post; what changed is that people are now chaining two or three of them into a full pipeline.

A day before this post went live, marketer Corey Haines posted a companion piece to the same loop: a skill that researches one prospect, finds a real trigger (a launch, a hire, a review), and writes a 71-word email with one observation, one number, and one ask, then sends it. The post picked up 200 bookmarks and nearly 9,000 impressions in under a day. Bookmarks are a stronger signal than likes here. People were not just nodding along. They were saving it to build later.

Isn't this just spam with extra steps?

Fair question, and the honest answer is: it can be, if you let it run unsupervised and skip the research step. Two things keep it from becoming spam.

First, the one-observation rule. An email that references something true and specific about the business is not a blast, it is a note that happens to be written fast. Second, a human still reads and approves before anything sends. The agent drafts. Dana decides. That single checkpoint is the difference between outreach and spam, and it costs you thirty seconds per email, not thirty minutes.

One more honest note, because Notma does not do hype: the demo that pulled 31,000 views leans on the line "you don't have to pay for Apollo anymore." Viewers pushed back on that in the comments, and they had a point. The skill itself is free code. The scraping and email-verification tools it calls still often need a paid API key. You are not eliminating cost, you are trading a flat monthly subscription for pay-as-you-go usage, which is usually cheaper for a four-person shop sending forty emails a week, but it is not zero.

How do you actually build it? Two skills, one loop

You do not need one enormous tool. You need two small skills that hand off to each other, plus you in the middle.

1. Write your Ideal Client Profile in plain English. Not a persona deck. One paragraph: the size of business, the signal that means they need you now (new location, recent hire, a bad review, a tool they clearly do not have), and where you have won before.
2. Let a prospecting skill search the open web. Point it at LinkedIn, Google Maps, directories, or competitor customer lists, and have it return real names, real businesses, and a contact if one is public. In the demo above, the exact request was: "Find me 500 SaaS founders in Texas with their direct emails." Your version might be: "Find me dental and law offices in Austin with 5 to 20 staff that opened a second location in the last year."
3. Have a research skill pull one real fact per prospect. A launch, a review, a job posting, a news mention. One fact is enough. Ten facts is a report nobody reads.
4. Have a writing skill draft the email. One observation, one number if you have one (time saved, cost avoided, a stat from your own work), one clear ask. Cap it near 70 words. Longer reads like a proposal; shorter reads like a person.
5. You approve, then it sends and logs the follow-up date. This is the step that keeps it honest. Nothing goes out that you have not seen.

Here is the diagram version of that loop, the same five steps end to end:

What this replaces, concretely:

Run this for Dana's forty accounting firms and dental offices: the agent returns the list in minutes instead of an afternoon, drafts forty short, specific emails overnight, and she spends her actual working time reading forty drafts and hitting send on the ones that sound like her, not writing them from a blank page.

What you have after this

A week in, you have stopped treating prospecting as a task you do when everything else is done. It is a loop that runs in the background and hands you decisions, not busywork. A month in, you have a real sense of which one-line observation actually gets replies, because you are sending enough of them to notice a pattern. Three months in, the list of prospects you have researched and reached is bigger than anything you could have built by hand in your spare evenings, and you got your evenings back.

Same founder. Same client list you were always going to chase. A completely different Tuesday afternoon.