Install QC is a post-install workmanship check: once a tint, PPF, ceramic, detailing, or wrap job is done, you drop photos of the finished work and the tool flags the likely defects that are visibly present in the shot — film issues like bubbles, lifting edges, debris under the film, gaps and uneven cuts, and scratches; ceramic issues like high spots, streaks, and holograms. Each finding comes with a one-line fix hint, and the whole check rolls up into a single pass / review / rework call, so problems get caught on the bench instead of in front of the customer.
Where to find it
Install QC lives at /app/ai/install-qc, but it isn't a standalone sidebar item — and there's currently no dedicated "AI Tools" entry in the sidebar at all. Reach it by opening Growth & Marketing from the sidebar (under "Run the shop"), then clicking its AI Assistant tab, which loads the AI Tools hub inline (that tab also has an "Open standalone" link straight to `/app/ai`); Install QC is one of the cards there, filed under the "Inspect & protect" section alongside Intake Damage Scan and Claim Triage. Plain Cmd+K opens a different feature ("AI Help," a KB-grounded Q&A popup) — despite what the hub page's own code comments suggest, it does not jump you to this tool, and the Cmd+Shift+K page-search palette doesn't have Install QC registered yet either. Intake Damage Scan is the check-in counterpart (pre-existing damage before you touch the car); Install QC is the check-out counterpart (your own workmanship after you're done).
Running a check
Drop up to 8 clear, well-lit photos of the panel you just finished into the dropzone — the page notes multiple angles help the model catch more. Tagging a service (Tint, PPF, Ceramic, Detailing, Wrap) is optional; note that the service buttons are toggles, so clicking an already-selected service again deselects it rather than switching to a different one. The "Run install QC" button stays disabled until at least one photo is dropped, and shows a spinner labeled "Checking…" while the request is in flight.
Reading the verdict
The result panel shows a verdict banner — Pass ("No defects flagged — looks ready to release"), Review ("Worth a look before the customer sees it"), or Rework ("Likely defect to fix on the bench first") — plus a confidence percentage, and any findings grouped into severity chips (severe / moderate / minor). If findings exist, a separate Findings card lists each one with its issue label (Bubble, Lifting edge, Debris, Gap, Uneven cut, Scratch, High spot, Streak, Hologram, Other), the location on the panel, and the fix hint. If the photo is too unclear for the model to call confidently, the tool doesn't guess — it degrades honestly and forces the verdict to "review," with a note telling you to re-shoot or eyeball it yourself. A failed analysis (unreadable image, etc.) shows a plain error asking for a clearer, better-lit shot instead of a verdict.
What it will and won't do
This is read-only and advisory by design: nothing you upload is ever saved, and the photo itself is never stored — it's read once for the check and discarded. The pass/review/rework call is derived deterministically from the findings rather than decided by the model directly, so the tool can never auto-fail a job on its own; you always make the final release-or-rework call, and the UI repeats that reminder under the findings list. Watch the badge in the top right of the page: "● Live" means it's reading your actual uploaded photos with a real model; "● Stub mode" (with an amber banner telling you to set ANTHROPIC_API_KEY) means no API key is configured and whatever result you see is sample data, not an analysis of your photo — useful for seeing how the tool behaves, but don't rely on a stub-mode verdict for a real job.