About this agent
Built for auto glass — where every lead starts with "can you send a picture?" and the picture, read well, is most of the estimate.
In glass work the photo is the inspection: chip or crack, in the driver's sightline or not, spreading from the edge or stable — a trained eye prices most jobs from a clear picture. The bottleneck was never the judgment, it was that the judgment had office hours.
This agent gives the judgment a 24/7 shift. A customer's photo gets analyzed for damage type, size, and location; the repair-or-replace call gets made with a confidence level; and the estimate comes from your pricing table. Clean cases get answered in minutes. Ambiguous ones get flagged to a human with the analysis attached — which is exactly when a human should look anyway.
What it actually does
Trigger: Photo uploaded in chat / manual
- 1
Receives the photos
chat_image_uploadFires when damage photos land in chat — from the lead-triage conversation or sent directly.
- 2
Reads the damage
analyze_damageType, size, and position — chip versus crack, sightline, distance from the edge.
- 3
Makes the repair-or-replace call
judge_repairabilityWith a stated confidence — low confidence routes to a human instead of guessing.
- 4
Prices it from your table
estimate_costYour configured rates by damage class and glass type — never invented numbers.
What you get
Photo in, priced assessment out — severity, repairability, and cost from your rates, with the uncertain cases routed to a human.
A run, as you’d see it
Agent runs land on a timeline — what fired, what the agent found, and the action waiting for a human. This is that screen.
Damage photos received
2 photos · via lead chat
Details
Assessment
10in crack, driver side — replace
Summary
Crack extends from the edge into the driver's sightline — repair not viable. 2022 F-150 windshield replacement at your configured rate: $389 plus recalibration. Recommend booking; crack will spread.
Assessments are drafts for your team by default; sending the price to the customer is your call — or your rule, once the accuracy has earned it.
By trade
Same agent, configured to how your vertical actually works.
The original: chip-versus-crack, sightline rules, ADAS recalibration flagging.
The same pattern reads shingle damage photos for storm-response triage.
Questions, answered
How accurate is photo-based estimating?
On clear photos of common damage, very — and the agent states its confidence per assessment, routing anything ambiguous to a human with the analysis attached. The failure mode is a human looking anyway, which was the old process for everything.
Where does the pricing come from?
A pricing table you configure — damage classes, glass types, recalibration add-ons. The model judges the damage; your table prices it. It never invents a number.
Can this work for trades other than auto glass?
The template is glass-specific, but the pattern — photo in, classified assessment out, priced from your table — configures for any trade where photos carry the diagnosis. Ask us on a call.