← Field Notes

Playbook

Your first AI employee: a 30-day plan for a five-person trade shop

Josue Hernandez · September 2026 · 12 min read

Most shops should not start with an "AI strategy." They should start with the phone. Here is the sequence we'd run to get one agent answering calls, booking estimates, and chasing quotes inside a month, without breaking anything that already works.

00Before you start: pick the leak, not the tool

Every small trade shop leaks money in the same handful of places. Calls that go to voicemail while the owner is under a sink. Estimates that get sent and never followed up. Invoices that sit for 45 days because nobody likes calling about money. Reviews that never get requested because the job's done and everyone's moved on.

An AI employee is only worth it if it plugs one of those leaks. So before you look at a single vendor, answer one question honestly: which one of these costs you the most jobs or the most cash? For most residential trade shops the answer is the phone. A missed call is a lost job, and it goes to the next plumber on the list. That's the leak this plan fixes. If yours is different, the structure still holds; swap the workflow.

One more rule. You are hiring one employee with one job. Not a "platform." Not a "transformation." One agent, one workflow, one number you can check at the end of the month.

01Week 1: write the job description

You wouldn't hire a person without telling them what to do. Same here. Spend the first week writing, in plain English, exactly what this agent handles and exactly what it doesn't. Ours for a plumbing shop looks like this:

• Answers every inbound call and text within seconds, 24/7, with the shop's name.
• Books estimates into the shared calendar, respecting service area, business hours, and the two-hour buffer around existing jobs.
• Collects name, address, phone, problem description, and a photo if the customer will send one.
• Hands off to a human, immediately, for: anything involving water actively leaking, gas smell, a customer who asks for a person, a price negotiation, or a complaint.
• Never quotes a price, promises a time it hasn't booked, or says "we'll be there today" unless the calendar says so.

The handoff list is the most important part. The failure mode of every contractor AI pilot we've seen is not the agent doing too little; it's the agent doing something it should have handed to a human. Write those rules first and write them narrow.

Also this week: get your data in one place. Your calendar, your customer list, and your pricing sheet need to be somewhere a piece of software can read them. If they're in three phones and a legal pad, that's the real week-one project, and it's worth doing even if you never touch AI.

02Week 2: wire the phone, run it in shadow mode

Get a new number for the agent. Do not forward your main line yet. Give the new number to nobody. Instead, have the agent listen on the main line in shadow mode: it hears every call, drafts what it would have said and what it would have booked, and a human reviews the log every evening.

You are looking for three things in the shadow log:

• Calls it would have handled correctly. This should be most of them.
• Calls it would have handled wrong. Every one of these becomes a new line in the job description.
• Calls it would have handed off that it didn't need to. This is fine early; being too cautious costs you nothing.

Two to three days of shadow mode is usually enough to find the obvious gaps. Common ones: the agent doesn't know your service area boundary, it doesn't know that "the guy who came last time" means a specific tech, it books a Sunday.

03Week 3: go live on overflow only

Now forward the main line to the agent only when a human doesn't pick up in four rings, and after hours. Your team still answers when they can. The agent catches what would have gone to voicemail. This is the lowest-risk way to go live: the worst case is exactly what was already happening.

This week, every evening, someone reviews:

• Every booking the agent made. Is it real, is it in the right slot, did it get the address right?
• Every handoff. Did the human actually get the message, and how fast?
• Every call that ended without a booking or a handoff. Why?

Expect to adjust the handoff rules again. Expect the office manager, if you have one, to be suspicious. That's healthy. Give them the review job; they'll find the problems faster than you will, and once they trust it, the whole shop will.

04Week 4: add the follow-up, and measure

With the phone stable, give the agent its second task: follow up on every estimate that hasn't been answered in 48 hours. One text, one call, polite, then it stops. This is the highest-return, lowest-risk automation in the trades because the alternative is usually nothing at all.

Then measure. You need exactly four numbers from the month, compared to the month before:

• Calls answered vs. calls missed.
• Estimates booked.
• Estimates that turned into jobs.
• Hours the owner or office spent on the phone.

If the first three went up and the last one went down, you have a working AI employee. If not, you learned which part of the job description is wrong for about the cost of a new drill.

05What this costs, honestly

There are three real costs. The setup, which is mostly the job-description and data work in week one whether you do it yourself or pay someone. The monthly software and telephony, which for a single agent is in the range of a few hundred dollars. And the review time, which is real: budget 20 minutes a day for the first month, then it drops.

Compare that to a part-time office hire, which in Los Angeles runs well over $2,000 a month before you factor in that they also go home at five.

06What we'd skip

• Letting it quote. Not in the first 90 days. Pricing is where you make or lose the job; keep a human on it.
• Letting it dispatch. Booking an estimate is one thing. Moving a crew is another. Later.
• Multiple agents at once. One employee, one job, one month. Add the second one when the first is boring.
• Anything that requires your customers to download an app. They won't.

This note is based on our own build work and on patterns we've seen across small contractors. We'll update it with real numbers from partner shops as those pilots complete. Where we cite a figure, we say where it's from; where it's our judgment, we say so.