The State of AI in the Trades: 2026

For most of the last decade, "AI is coming for your job" was aimed at knowledge workers. The trades were treated as the safe harbour: you cannot code-review a rooftop unit or ask a language model to bend conduit.

That framing missed the point. AI did not come to the trades to replace the work. It came for the 20 minutes before the work — the lookup, the calculation, the second opinion, the "am I sure about this" pause that happens on every job.

What changed

Three things converged.

Models got good at citing instead of guessing. The biggest blocker to professional adoption was never intelligence, it was accountability. A tradesperson cannot act on an answer that might be invented. Retrieval-grounded systems that point at a specific article or table changed the risk calculus entirely.

Phones became the right interface. No one is carrying a laptop up a ladder. The whole interaction has to work one-handed, in bad light, with gloves half-on.

The reference material got structured. Codebooks, manufacturer charts, and standards are dense, but they are also highly structured. That structure is exactly what makes them tractable for retrieval.

What techs actually use it for

The pattern we see across our own products is consistent, and it is not what most people predict.

• Lookup replacement. Not "do my job" — "tell me what the book says, right now, without me climbing down."
• Math verification. The calculation was never hard. Doing it correctly at 4pm on the sixth call of the day is hard.
• Second opinion on unfamiliar equipment. Every tech has a range. A good assistant extends the edges of it.
• Teaching apprentices. This one surprised us. Senior techs use it to explain the reasoning, not just the answer.

What it is still bad at

Being honest about this matters more than the upside.

AI cannot see the job. It does not know the previous installer did something creative behind the drywall. It does not feel the vibration in a blower motor or smell a failing capacitor. It will confidently describe the textbook case while you are standing in front of an exception.

The correct mental model is a very well-read colleague who has never been on site. Extremely useful. Never the final authority.

Where this goes

The next step is not smarter chat. It is narrower tools that know one trade completely. A generalist assistant that is decent at everything is worth less to a working tech than a tool that knows the mechanical code cold and nothing else.

That is the bet we made with our field products, and 2026 is the year it stopped being a contrarian one.

Building something for a trade? Get in touch — it is the work we most like doing.