How I Built an AI Sales Agent for My Directory in One Day With Claude Code

I run TapWaterData.com. It's a directory of 18,700 water utilities across the US -- every one of them has verified contacts, service areas, EPA compliance data. I've been running it for six months.

Companies that sell to water utilities -- pipes, chemicals, pumps, monitoring software -- they buy this kind of data from ZoomInfo and Apollo for thousands a year. Same data I have.

And yeah, I get inbound sales. Someone Googles "water utility contacts," finds my site, reaches out. It works. But it's unpredictable. One month I'll close a few deals. Next month, crickets. I can't control when someone decides to search.

I kept staring at this gap. I have 18,700 verified utility contacts. There are hundreds of companies who'd pay for that data. But I was just... waiting for them to find me.

After a particularly quiet month, I decided to stop waiting.

I opened Claude Code, described what I wanted, and it built the entire system in a single session. An AI sales agent -- 8,800 lines of code, 39 files -- that runs 24/7 on my Mac Mini. It finds companies that sell to water utilities, researches them, and writes personalized outreach using data only my directory has.

I didn't write the code. I wrote a spec. Claude Code did the rest.

Seven days later: 665 companies discovered. 306 personalized emails drafted. 60% open rate on sends. Total cost: $27.

Let me walk you through exactly what happened.

Here's what most directory monetization advice sounds like: sell featured listings, run display ads, charge for premium access.

I tried thinking about it differently.

My directory isn't just a website. It's 18,700 rows of data that sales teams at water industry vendors spend weeks building manually. Who runs procurement at each utility? How big is their service area? Any EPA violations? What conferences do their vendors go to?

When I send a cold email, I can reference all of that. A generic sales email says "we help companies like yours." Mine says "you just expanded into Texas -- I have procurement contacts for 1,142 Texas utilities."

That's not a cold email anymore. That's a warm intro backed by data nobody else has.

So I stopped thinking of TapWaterData as a website people visit. I started thinking of it as an unfair advantage in outbound sales.

This is the part that might surprise you.

I'm not a backend engineer. I didn't sit down and write 8,800 lines of TypeScript. I opened Claude Code -- Anthropic's coding agent -- and described what I wanted.

My spec was basically this: "I have a directory of 18,700 water utilities. Build me an agent that runs on a loop, finds companies selling to these utilities, classifies them, finds the right contact person, researches the company for timing signals, and writes a personalized cold email using my directory data as the value prop."

One session. A few hours. Claude Code built the whole thing. Database schema, API clients, scraping logic, LLM classification, email drafting, SendGrid integration, a React dashboard for reviewing outreach, and a PM2 config to run it 24/7.

I reviewed the output, tested it, and hit start.

I know that sounds too easy. And honestly, it kind of was. The hard part wasn't the code. It was knowing what to ask for. The spec -- the clear description of what each stage should do, what data it should use, what the output should look like -- that's where the real work happened.

Let me make this concrete. Here's what happens to a single company as it flows through the pipeline.

AquaPure Corp shows up on the AWWA conference exhibitor list. (Composite example, but this is exactly how the real pipeline works.)

The orchestrator -- the code that runs the loop -- is 147 lines. Each step is one function. They chain together in a loop that runs every hour.

Here's what I want to be honest about: the AI isn't the secret sauce. The data is.

Claude does two things in this system. It classifies companies ("do they sell to water utilities? A, B, or C tier?") and it writes emails. That's it. Everything else is normal code -- HTTP requests, database queries, API calls.

But here's why the emails get 60% opens instead of the industry average of 21%.

Before writing anything, the agent builds a dossier. Seven questions about the company:

1. What do they sell?
2. How do they sell it? Direct, distributors, reps?
3. What territory?
4. Growing? Hiring? New contracts?
5. What's their pain point in reaching utilities?
6. What data from MY directory helps them most?
7. What's the angle?

Then Claude writes the email using that dossier plus my 18,700 utility contacts as the offer.

Anyone can ask ChatGPT to write a cold email. Nobody else can say "I have procurement contacts for 1,142 Texas water utilities." That's the difference. The AI is the pen. Your directory data is the ink.

I let it run for a week. Here's everything -- good and bad.

Some context on those numbers. A human doing this work -- finding 665 qualified companies, looking up contacts, writing personalized emails for each -- that's about 3 months full-time. ZoomInfo would charge $500-2,000/month for comparable data. I spent $27 and used a Mac Mini that was already sitting on my desk.

I'm including this because every "how I built X" post that only shows wins is lying.

The core works. Find, classify, enrich, write, send, track. 60% open rate says the approach is right. Infrastructure just needs hardening.

I went with the simplest thing at every decision point. This is a solo operation, not a startup.

I keep coming back to this: the hard part wasn't building the agent. Claude Code handled that. The hard part was the spec -- knowing what to ask for.

And the spec comes from answering four questions about your directory:

1. Who would pay to reach your audience?

2. What data do you have that they don't?

3. How does that data help them specifically?

4. What does the pipeline look like? (Find, classify, enrich, research, write, send -- same structure, different data.)

Here's what I promised. The exact product spec I gave Claude Code to build this system. The ICP definition, the pipeline stages, the data model, everything.

If you run a directory and want to build your own version of this, the spec is the starting point. Adapt it for your niche, your data, your ICP. Then hand it to Claude Code.

This is week one. I'm tracking what happens next:

• Scaling sends. I've been the bottleneck. Time to batch-approve Tier A drafts.
• Fixing infrastructure. No more sleep mode. Firecrawl alerts. Monitoring.
• Tracking conversions. Replies, meetings, actual deals over 30 days.
• LinkedIn outreach. The system already generates LinkedIn messages. Adding delivery.

The real question isn't whether the agent can find prospects and write emails. It can. The question is whether directory-powered outbound converts better than generic tools.

I think it does. Because nobody else has my data.

I'll share the results as they come in.

-- Andy