AI mold inspection: what it can and cannot tell a homeowner

Nobody inspects a building with a language model. What AI can do is read what a homeowner reports, look at the photos they took, and write a careful preliminary assessment of what is visible — the same triage a good estimator does over the phone, except written down, consistent, and available at eleven at night.

The distinction matters, because the moment such a document claims more than that, it becomes a liability for the company whose logo is on it.

Inspecta generates a structured report rather than a paragraph of prose. The homeowner is shown a headline, a concern level with a short summary, what is observable in their photos, the moisture indicators and affected materials that go with it, a plain-language explanation written for someone who is not in the trade, and a recommended next step. If they used your remediation estimate widget, the preliminary range sits alongside it.

More is generated than the homeowner is shown. Possible contributing factors, the services that would typically address the condition, an urgency level, the assessment’s own confidence and its stated limitations are all recorded on the lead, together with a contractor-facing summary of the submission written for somebody who prices this work for a living.

An assessment produced from photographs cannot responsibly diagnose anything, so this one is built not to try. It will not:

• Confirm that mold is present, or diagnose what a discoloration is.
• Name a species, or use the phrase "black mold".
• Make health, toxicity or exposure claims.
• Declare a home safe, unsafe or uninhabitable.
• Guarantee an outcome, or make claims about your company it cannot support.
• State a dollar figure — pricing is calculated separately, by code, from your own rates.

What it does instead is describe. Possible microbial growth. Indications of moisture intrusion. Conditions consistent with what was reported. Anything behind a wall or under a floor is treated as unknown, because it is.

Instructing a model to be careful is not a safeguard, because instructions are advisory. So every field of every report — the parts the homeowner reads and the parts only your team sees — is scanned before anything is shown.

Species claims, toxicity and health claims, habitability verdicts, definitive diagnoses and dollar amounts are rejected outright at that stage. Genuine disclaimers survive, because the narrow set of sanctioned disclaimer phrasings is allowed through deliberately. If a report fails the check, generation is retried once, and if it fails again the homeowner is shown a conservative fallback that recommends professional evaluation instead.

The remaining rules — no guarantees, no claims about your company that cannot be supported — are set in the instructions the assessment is written under, and the hedged, observational style they produce is what the scan then holds the finished report to.

There is a temptation to think that a bolder report sells more work. In practice the opposite reads as more professional. A homeowner who is told “you have toxic black mold” by a website knows they are being sold to. A homeowner who is told what is visible, what it might indicate, what cannot be determined from photographs, and that a certified professional should evaluate it in person, is being treated as an adult.

The recommendation at the end of every assessment is the same recommendation you would make: have somebody look at it properly. That is the appointment you wanted, and the homeowner arrives at it already informed.

An AI assessment is triage, not an inspection. It sorts the water stain from the wall that needs opening, it gives the homeowner something credible to read while their concern is fresh, and it hands your team a file with photos, answers and a summary already attached.

The judgement stays with the person who walks through the door with a moisture meter. The software just makes sure that person walks into the right houses.