Short answer: AI is genuinely useful during a live inspection for reading labels, recalling documentation, drafting notes, and suggesting likely causes. It is not reliable for determining code compliance, judging structural adequacy, or issuing a verdict. Treat it as a well-read assistant that must cite its source, and keep the sign-off with a qualified human.
What works well today
| Task | Why it works |
|---|---|
| Reading model and serial plates | OCR on a full-resolution still is fast and accurate. See OCR extraction |
| Surfacing the right troubleshooting article | Retrieval over your own knowledge base with a match score |
| Suggesting next diagnostic steps | Common failure patterns are well documented |
| Drafting the report narrative | Summarizing notes and captures the human then edits |
| Flagging missing evidence | Comparing captures against a required shot list |
Where it fails
- Code compliance. Requirements vary by adopted code edition, local amendment, and occupancy. A confident answer from a general model is frequently wrong for your jurisdiction.
- Structural and safety judgment. Adequacy questions need training, liability, and often instruments.
- Concealed conditions. A model cannot infer what is behind the drywall from a photo of the drywall.
- Scale and measurement. Without a reference in frame, dimensional estimates are guesses.
- Rare failure modes. Retrieval finds the common answer; the unusual case is exactly where it misleads.
The design that keeps it safe
- Ground every suggestion in your own documents. Retrieval over an approved knowledge base beats open-ended generation for field accuracy.
- Show the citation and the match score. If the assistant cannot point to the article it drew from, the inspector should not act on it.
- Never auto-issue a verdict. Suggestions populate fields; a human selects pass, fail, or corrections required.
- Log what the assistant said. The record should show what was suggested and what the human decided.
- Control frame processing. Explicit consent for camera frames sent for analysis, with clear retention and deletion.
Accountability
The licensed inspector, adjuster, or technician owns the outcome. AI shortens the path to the right information; it does not transfer responsibility. Programs that state this in policy — and whose tools make the human decision explicit — get the productivity benefit without the liability exposure.
See how this is implemented in AI Assist, with knowledge-base grounding, in-call citations, and consent controls.
A capability map, honestly drawn
The useful question is not "is AI accurate?" but "accurate at what, with what input?" Vision models on a live call see a compressed, moving, badly lit frame chosen by someone who does not know what matters. That constrains the task list sharply.
| Task | Realistic today | Why |
|---|---|---|
| Read a nameplate, model, or serial number | Strong | Text recognition on a still frame is a solved problem when the shot is close and steady |
| Identify a component or equipment class | Strong | Large visual variety but distinctive shapes |
| Spot an obvious visible defect — corrosion, scorching, a disconnected line | Good | High-contrast visual signature |
| Confirm presence and completeness against a checklist | Good | Binary, and the model is told what to look for |
| Grade severity or remaining life | Weak | Requires history, load context, and touch |
| Diagnose intermittent faults | Very weak | The symptom is usually absent on camera |
| Anything defined by sound, smell, heat, or vibration | Out of scope | The sensor is a phone camera |
The failure modes worth designing around
- Confident wrong answers. A model will name a part it has never clearly seen. Suggestions need a visible confidence signal and a citation, not a bare assertion.
- Frame selection bias. The model sees what the guest pointed at. If the fault is behind the panel, no amount of reasoning helps.
- Compression artefacts. Video streams degrade fine detail — hairline cracks and small corrosion are exactly what gets smoothed away. Grab a full-resolution still before judging.
- Anchoring the human. The most expensive failure is not the model being wrong; it is a technician accepting a wrong suggestion because it arrived first and sounded certain.
- Missing history. "Is this worse than last time?" is unanswerable without the previous inspection, which is why prior media matters more than a bigger model.
The pattern that works: suggest, cite, defer
Treat the model as a fast, tireless assistant with no accountability. In practice that means three rules:
- Suggest, never decide. AI output is a prompt for the human, never a verdict written into the record on its own.
- Cite the source. Every troubleshooting step should point at the specific knowledge-base article it came from, with a match score, so the technician can check the reasoning in two seconds instead of trusting it.
- Defer on low confidence. A well-designed assistant asks for a better shot — "move closer to the data plate" — rather than guessing from a blurry frame. Requesting a better input is a feature, not a failure.
That is how AI Assist works inside Virtual Support Pro: it observes the live frame, interprets it against your own indexed documentation, and surfaces the matched article alongside the suggestion, with the human on the call making the call.
Where the value actually shows up
Because the wins are rarely the dramatic diagnosis. They are administrative and they compound:
- Nameplate capture that fills the model and serial fields without dictation errors.
- Checklist completeness prompts before the call ends, which is what prevents the second visit.
- Retrieval — pulling the right page of the right manual in seconds instead of minutes.
- Draft report structure from session notes and captures, reviewed and signed by a person.
- Triage: deciding whether this needs a truck at all, which is the single biggest cost line in field service.
None of that requires the model to be a better diagnostician than your technician. It requires it to be faster at looking things up and never bored.
Governing it
If AI output can influence a decision that affects a customer, write down how it is used: what data leaves the session, whether frames are retained, who can turn the assistant on, and how disagreements are recorded. Frame-capture consent should be explicit, and the retention of AI-processed frames should follow the same schedule as everything else — see GDPR and CCPA. Suggestions that were shown and rejected belong in the audit trail too; that record is what protects the technician who made the right call against a confident machine.
FAQ
Can AI grade damage severity?
It can propose a category from your rubric. Anyone accepting it without review is the risk, not the model.
Does using AI affect the evidentiary value of the media?
No, as long as originals are untouched and AI output is stored as annotation rather than modifying the source. See chain of custody.
What about hallucinated part numbers?
This is why OCR from an actual captured plate beats a model recalling a catalogue.
Can AI diagnose equipment faults from a video call?
It can reliably read plates, identify components, and flag obvious visible damage. Severity grading, intermittent faults, and anything sensed by sound, heat, or smell remain human work.
How accurate is AI on live inspection video?
Accuracy depends far more on the frame than on the model. A close, steady, well-lit still gives strong results; a compressed moving stream of a dim crawlspace does not. Always judge from a captured still, not the live feed.
Should AI write the final inspection verdict?
No. Draft the narrative, then have a qualified person review and sign it. The verdict carries professional liability that a model cannot hold.
How do you stop AI from giving confidently wrong advice?
Ground it in your own documentation, show the matched source and score with every suggestion, and design it to ask for a better image when confidence is low.
Does using AI on calls create privacy issues?
It can. Get explicit consent for frame capture, control what is retained, and apply the same retention and deletion rules you use for all session media.
Keep reading
- AI-generated inspection reports — what to automate and what to sign.
- Remote video support ROI — where triage pays for itself.
- Reduce support calls and misdiagnosis — seeing the problem first.
- C2PA content credentials — provenance in an era of generated images.
- Why FaceTime fails as an inspection tool — the missing record.