How the AI works
W-Scrut uses Google's Gemini 2.5 Vision model to evaluate warranty claims. Here's what happens from the moment a customer sends a message to the moment they get a reply.
The evaluation pipeline
Risk Gate
Before the AI even looks at the claim, two risk checks run: (1) Has this customer made a concurrent claim in the last 10 minutes? (2) Does this customer have a suspicious historical return rate (>15%)? Either condition escalates to a human reviewer immediately.
Policy Fetch
W-Scrut fetches your return policy for the specific SKU from your GitLab repository. If no policy exists for that SKU, it falls back to the W-Scrut Standard Policy v2.1.
Evidence Resolution
The customer's evidence photo is downloaded and converted to the format Gemini expects. For Telegram, this means resolving the file_id into the actual image bytes.
Gemini Vision Analysis
Gemini 2.5-Flash receives three inputs: (1) the photo, (2) your SKU policy, (3) the customer's written claim. It analyzes whether the visible defect matches the description and is covered under your policy.
Adjudication
The model outputs a decision (APPROVE/REJECT/ESCALATE) and a confidence score (0–1). If confidence ≥ 0.78, the decision is automatic. Below that threshold, the claim is escalated to your human review queue.
Notification
The customer receives an immediate reply in the same channel they used. Your dashboard is updated in real time. If escalated, the claim appears in your review queue.
Confidence threshold
The 0.78 threshold is calibrated to minimize both false approvals (fraud getting through) and false rejections (valid claims blocked). You can adjust this per-tenant — contact support if you need a custom threshold.
What makes a good claim photo
- Clear focus on the defective area (not blurry)
- Good lighting — natural light is best
- Shows the whole product, not just a close-up
- Multiple angles if the defect is subtle