Operations

The Call From the Attic

By John J. BakerNovember 18, 20269 min read


When a newer tech gets stuck on a fault code and calls the senior tech, two trucks stop instead of one. The answer often already sits in the manufacturer manual, that unit's service history and the agreement terms. An AI assistant that reads those documents, and that the junior tech asks first, is one of the clearest early uses of AI in field service.

I want this piece to be informative to someone exploring this website and trying to understand what AI can actually do in a trades business. So I will start with the plain version, then the research, then the limits.

Here is the call. A second-year tech is in an attic, or on the roof of a strip mall, looking at a rooftop unit throwing a code they have not seen before. They try a couple of things. Then they call the senior tech, who is halfway through a repair across town. The senior tech puts down their tools, talks it through over the phone, and sometimes drives over. One problem, two trucks, and a customer waiting.

Today: the stuck call

  1. A second-year tech on a rooftop unit sees a fault code they have not seen before.

  2. They try a couple of things, then call the senior tech, who is on another job.

  3. The senior tech stops their own repair to talk it through, and sometimes drives over.

  4. Two trucks are stalled on one problem, and two customers are waiting.

One problem, two trucks.

Assistant first

  1. Same rooftop unit, same fault code.

  2. The tech asks the assistant with the model number, the code and what they have checked.

  3. It answers from the manufacturer manual, this unit’s past tickets and the agreement terms, and shows where each answer came from.

  4. The tech checks the source, does the work, and writes one line on the ticket about what fixed it.

Escalation point

Answer does not match what they see, or anything involves safety: call the senior tech, with the model, code, history and steps tried already in hand.

An illustrative sequence, not a measured job. The assistant reads only the documents the contractor loads: manuals, ticket history and agreement terms.

Why does a junior tech's stuck call cost two trucks, not one?

Because the person who knows the answer is also the person you most need working on their own job. The junior tech waits. The senior tech stops. The customer on the senior tech's job waits too. Nobody tracks this as a cost, because it looks like good teamwork, and it is. It is also the most expensive way to look something up.

Most of what the senior tech says on that call is not judgment. It is recall: what this code means on this brand, what we found on this unit last spring, whether this customer's agreement covers the part. That recall lives in their head because nobody ever put it anywhere else. I wrote about that pattern in the office in your business runs on what people remember. This is the field version.

It will not get easier on its own. The U.S. Bureau of Labor Statistics counts 440,900 HVAC and refrigeration mechanic and installer jobs in 2025, projects 11 percent growth from 2025 to 2035, and expects about 40,600 openings a year. BLS also notes that new hires typically need a lengthy period of on-the-job training to reach competency. More of your bench will be early in that training, and your senior techs will get more of these calls, not fewer.

What is an AI troubleshooting assistant, in plain terms?

It is a chat tool on the tech's phone that answers questions only from documents you give it. You load your manufacturer manuals, your past service tickets and your agreement terms. The tech types or says the model number, the code and what they have checked. The assistant finds the relevant pages and tickets and answers in plain language, with a pointer to where the answer came from, so the tech can check it.

The technical name for this setup is retrieval-augmented generation. The important part is the first word. The assistant retrieves from your documents before it writes anything, instead of answering from whatever it picked up on the internet. That is the difference between a tool you can hold accountable and one that sounds sure of itself.

It does not replace the senior tech. It replaces the first five minutes of the phone call.

What does the research say about AI and less experienced workers?

The best evidence so far says the biggest gains go to the least experienced people. Two studies stand out, and I will be honest about how close each one is to your work.

The first is a field experiment from ETH Zurich, published in the International Journal of Production Research in July 2025. The researchers built a troubleshooting assistant on project documents for train commissioning, which is complex, hands-on technical work. With 19 technicians completing 173 tasks, technicians using the assistant significantly outperformed those without it, and the benefits were more substantial among less experienced technicians. The gains also depended on how familiar and comfortable each technician was with AI. It is a small sample and it is not HVAC, but it is the closest published test I have found to a tech asking an assistant for help in the field.

The second is larger. Brynjolfsson, Li and Raymond studied 5,179 customer support agents, published in the Quarterly Journal of Economics in 2025. An AI assistant raised issues resolved per hour by about 14 percent on average, and by 34 percent for novice and low-skilled workers, with minimal impact on the most experienced. The authors suggest the tool spread the practices of the best workers to everyone else. That is a support desk, not a rooftop, so treat it as an analogy.

Support desk, 5,179 agents

Gain in issues resolved per hour with an AI assistant

  • All agents, average

    +14%

  • Novice and low-skilled agents

    +34%

Most experienced agents: minimal impact. Support desk work, so an analogy for the field.

Field experiment, 19 technicians, 173 tasks

Troubleshooting assistant built on project documents

  • Technicians with the assistant significantly outperformed those without it.

  • The benefits were more substantial among less experienced technicians.

Train commissioning, not HVAC, and a small sample. Gains also depended on each technician's familiarity with AI.

Sources: Brynjolfsson, Li and Raymond, “Generative AI at Work,” Quarterly Journal of Economics, 2025. Löwhagen, Schwendener and Netland, ETH Zurich, International Journal of Production Research, 2025.

Put together, the pattern is consistent: an assistant built on good knowledge helps the newer person most, and much of what it hands them is what your best people already know.

What would the assistant actually read?

Three kinds of documents, all of which you probably already have:

  • Manufacturer manuals and service bulletins for the equipment lines you service most. Fault code tables, wiring diagrams, startup and reset procedures, spec values.
  • Your own service history for that unit. Every past ticket: what was found, what was replaced, what the tech wrote. This is where "it always does that in July" lives.
  • The customer's agreement and warranty terms. What is covered, what needs approval, what the response commitment is.

That last one matters more than it sounds. In ServiceTitan's 2026 commercial survey of 1,020 mostly mechanical, electrical and plumbing contractors, fielded by Thrive Analytics, 69 percent named warranty coverage and agreement details as a top obstacle for technicians in the field, and access to equipment service histories came up as a challenge too. The survey is vendor-funded, so read it as a direction, not a precise measure. The information exists in your office. It is just not in front of the tech when they need it.

If your tickets are thin, the assistant will be thin too. The same tickets do double duty; I showed another use for them in the 30-day clock in the service ticket.

Which calls should still go straight to the senior tech?

Anything involving safety, judgment or the customer relationship. The assistant is for questions with an answer in a document. It is not for questions that need someone who has seen a hundred of these.

Ask the assistant first

Facts that live in a manual, a ticket or an agreement

  • What does this fault code mean on this model?
  • What are the spec values and the reset procedure in the manual?
  • What did we find on this unit last time?
  • Is this unit under warranty or a service agreement, and what does it cover?
  • Which part number fits this unit?

Call the senior tech first

Safety, judgment and the customer relationship

  • Anything that could hurt someone: a gas smell, signs of overheating or burned wiring.
  • Repair or replace: what we recommend to the customer.
  • The assistant’s answer does not match what the tech is looking at.
  • The same unit failed again after a recent repair.
  • An upset customer, or a question about cost.
  • Work outside the tech’s training or certification.
A working rule to adapt with your own senior techs, not sourced data. When in doubt, the tech calls a person.

The rule I would write down for the team is short: ask the assistant first for facts, call a person first for safety, and call a person any time the answer does not match what you are looking at. When the junior tech does call, they call with the model, the code, the history and what they have already tried. That makes the senior tech's five minutes count.

How do you keep junior techs learning instead of copying the answer?

By treating the assistant as a reference, not a crutch, and by keeping the senior tech in the loop on purpose. This is the part I care about most, because a tech who never struggles with a problem may never learn to solve it.

There is early research on exactly this worry. A 2026 interview study of 14 IT professionals by Abou Khamis, Assal and Matrawy, posted on arXiv and not yet peer reviewed, found that AI can shorten the path to expertise while also reducing exposure to the hands-on cycles of building, failing and debugging that have traditionally trained experts. They also describe pressure to move at AI speed even when slower manual checks are needed for safety or validation. It is a small, qualitative study in a different field, but the warning fits the trades.

A few habits keep the learning human:

  • Explain, then act. Have the assistant give the reason behind the step, not just the step, and have the tech write one line on the ticket about what they learned.
  • Weekly ride-along review. The senior tech reviews a handful of the week's assistant questions with the junior techs. It turns a lookup into a lesson.
  • Some calls stay manual. Early in training, certain diagnostics are done by hand first, with the assistant used to check afterward.

This is the same idea I wrote about in start it by hand: keep the human part of the work human, and let the tool amplify it.

What are the honest limits of an AI assistant in the field?

It is only as good as what you load into it, and it can be wrong. If the manual is for the wrong model year, or the tickets say "checked, OK" and nothing else, the answers will be weak.

It can also sound confident when it should not. That is why it should always show where its answer came from, and why the tech checks the source before acting on anything that matters.

It does not see the unit. It cannot smell a burned contactor or hear a compressor straining. The tech is still the one on the roof.

And the research is early. The field experiment is 19 people in rail. The large study is a support desk. The caution study is 14 interviews. None of it measures HVAC techs directly, so the right move is a small test in your own shop, not a big rollout. For more on which work should never be handed to a machine, see what not to automate.

How would you test this on one trade and one equipment line for 30 days?

Pick your most common commercial equipment line and two or three junior techs. Load the manuals for that line, the last two years of tickets for units of that brand, and the agreement terms for the customers who own them.

For 30 days, those techs ask the assistant before calling the senior tech. Ask the senior tech to keep a simple tally: calls received, and for each one, whether the assistant had already helped or whether it was a real judgment call. Review every week together.

At the end you will know three things: how many calls the assistant absorbed, where it was wrong, and what your tickets and manuals are missing. That last list is valuable on its own, whatever you decide about AI.

Where ClearOak comes in

ClearOak & Co. is a software and operations partner for trades businesses. We are servant leaders and operators first. We find the process costing you the most, bring in the right builders for the job, and train your team to own what we build.

For technician support, I would start by riding along with how a stuck call works today: who calls whom, what they ask, and where the answer actually lives. Then we gather the manuals, clean up the ticket history for one equipment line, write the escalation rules on one page with your senior techs, and run the assistant next to your team for a month before anyone relies on it. What you get back is time and margin: senior techs who finish their own jobs, and fewer second trucks on one problem.

If your senior techs spend their days on the phone with the rest of the bench, schedule a call and we will map one week of those calls together.

Prefer email? Reach me at john@clearoakconsulting.com and tell me which equipment line generates the most calls.

Sources

Have a workflow that is costing you?

Book a free call. We will tell you honestly whether we can help and what that looks like.