A tool we were testing looked at a photo of a water heater and offered a finding: corrosion at the TPR valve, recommend evaluation by a licensed plumber. It was right that there was corrosion. It was wrong about the valve, which was fine. The rust was on the fitting above it.

Close enough to be useful. Nowhere near close enough to publish.

That gap is the entire argument about AI in this industry, and most of the marketing around it walks straight past.

What it is genuinely good at

Sorting comes first. A four hundred photo inspection has structure hiding inside it, and software can put each image with the system it belongs to faster and far more consistently than a tired person at nine in the evening.

Drafting comes second. The paragraph about missing GFCI protection has been written by every inspector in the country. Having it arrive already drafted, in your wording, ready for you to correct, takes away typing without taking away thought.

Surfacing is the third and the most valuable. A model can look at a photograph, notice something that resembles a defect, and put it in front of you. Treat that as a prompt, never a verdict.

Where the line sits

Three things it should not be allowed to do.

Decide severity. Whether a crack is cosmetic or structural depends on things no single photograph contains. The age of the house. The soil. What the neighbor two doors down is dealing with. What you felt underfoot when you walked the floor. Severity is a judgment call and it belongs to the person who made the site visit.

Publish without review. Any system that can send a report to a client before a human has read it has confused speed with value. The minutes you save are not worth the one report that goes out with a finding you would never have written.

Invent what it did not see. A language model will fill a gap with something plausible if you let it. In an inspection report, plausible and observed are completely different categories, and only one of them is defensible.

Who signs

Your license sits on that report. Your insurance sits behind it. If a buyer moves into a house and finds something the report missed, no one is going to be interested in which part of the document was drafted by software.

That is not an argument against using AI in this work. It is an argument about where the tool stops and the professional starts, and about being honest that the line is there at all.

The test

When you look at any tool that puts AI near your reports, ask one question. What happens to a finding the model gets wrong?

If the answer is that you catch it in review, because the product is built to surface potential findings and wait for you to accept, correct, or throw each one out, then the AI is doing work worth having. If the answer is vague, or the demo shows a report writing itself while somebody talks about hours saved, then the tool is asking you to underwrite its confidence with your license.

We built AEYES to surface potential findings and hand them to the inspector. Every one gets accepted, edited, or discarded by a person before a report goes anywhere. That is slower than full automation. It is the only version of this we would put our own name on.