← Field Notes

Pitfalls

Why most contractor AI pilots die in month three

Josue Hernandez · September 2026 · 9 min read

The pattern is consistent enough to be boring. Month one: the demo was great, everyone's excited. Month two: it's working, mostly, and someone is quietly fixing its mistakes. Month three: that someone stops, nobody notices for two weeks, and the tool becomes a line item somebody cancels in Q4. It's almost never the model's fault. It's one of these six things.

01Nobody owns it

The owner bought it. The office manager is expected to run it. The techs are expected to use it. None of those three people had it in their job description, and the owner is the one least likely to look at it daily. So it drifts.

Fix: before the pilot starts, one named person owns it, and it's in their actual job. Fifteen minutes a day, on the calendar. If nobody can spare fifteen minutes, the shop isn't ready, and that's a fine answer.

02The data was never there

"Connect it to your customer list" assumes there is a customer list. In a lot of shops there's a phone, a QuickBooks file that's six months behind, and the owner's memory. The AI vendor's onboarding assumed a CRM. The pilot spent its first month being a very expensive way to discover that.

Fix: do the data cleanup as its own project, before any AI, and treat it as a win on its own. A shop with a real customer list, a shared calendar, and a current pricing sheet is a better business whether or not it ever automates anything. Most of the value people attribute to "AI" in the first month is actually this.

03The handoff rules were vague, so the first bad handoff killed trust

The agent booked a job outside the service area. Or told a customer with an active leak that someone would call back tomorrow. Or quoted a number. It happened once, the owner heard about it from an angry customer, and from that day forward the office manager double-checked everything the agent did, which meant it was now creating work instead of removing it.

Fix: the handoff list is the product. Write it before anything else, write it narrow, and start the agent in shadow mode where its mistakes cost nothing. One bad handoff in front of a customer costs more trust than a hundred correct ones earn. We wrote out how we sequence this in the 30-day plan.

04The office manager was routed around instead of brought in

This one is political and it's the most common. The person who has been answering the phone for eight years hears "AI is going to answer the phone" and, reasonably, hears "you're being replaced." They don't sabotage it. They just don't help it, and a tool that the most knowledgeable person in the office won't help is dead.

Fix: make them the owner from item one. The job isn't "answer the phone" anymore; it's "run the thing that answers the phone, catch what it misses, and handle the calls that actually need a human." That's a promotion, and it should be framed and paid like one. Every successful small-shop pilot we've seen had the office manager as its biggest advocate by month two.

05There was no number

Ask "is it working?" three months in and you get "I think so?" Nobody wrote down what calls-missed looked like before, so nobody can say it went down. Without a number, the pilot is judged by the last mistake anyone remembers, and there's always a last mistake.

Fix: pick two to four numbers before you start and write down what they are today. Calls missed, estimates booked, days-to-paid, hours on the phone. Then check them at day 30 and day 90. It doesn't have to be sophisticated. A notebook works.

06The demo was the product

Some of what's being sold to contractors right now is a demo with a pricing page. It handles the happy path beautifully and falls over on the first customer who calls from a noisy jobsite with a partial address and two problems. The vendor's answer is "that's an edge case." In the trades, that's every case.

Fix: test with your ugliest real calls before you sign. Play a voicemail of a customer with a thick accent describing a water heater problem over traffic noise. If the vendor won't run that test, that's your answer. And ask what happens when it fails: does it hand off, or does it guess? Guessing is disqualifying.

07The honest summary

Notice that none of the six is "the AI wasn't smart enough." Current models are more than capable of answering a plumbing shop's phone. The pilots die on ownership, data, handoffs, people, measurement, and vendor honesty. All six are fixable before day one, and all six are cheaper to fix before day one than after.

If you're evaluating a pilot right now, run down the list. If you can't check all six, don't start yet. Fix the ones you can't check; that work pays off whether or not you ever buy the tool.

This note draws on our own build work and on conversations with contractors and vendors in the Los Angeles area. The patterns are ours; the examples are composites, not any single client. We'll add named case studies as partner pilots complete and partners agree to share them.