A chargeback hits. AI surveys your evidence and drafts the response.

Deep dive · 01

It only recommends evidence you actually have

Before the AI sees anything, a deterministic query counts what's on file for this customer: how many invoices, quotes, job photos, message-thread entries, warranties, and past appointments exist. That inventory is what the AI reasons over — and on the way back out, any 'include' flag for a section with a zero count is force-disabled in code. So the recommendation can never tell you to attach photos you never took or a warranty that was never registered. The pack reflects reality, not a hopeful template.

• Real DB counts run first; the AI reasons over the inventory, not guesses
• Reason code shapes the pick: fraud claim → identity / signature / on-premises proof; quality complaint → photos + warranty; not-received → completion proof + customer presence
• Post-processing strips any section whose evidence count is zero
• Output shape matches the evidence manifest exactly, so 'Apply' is a clean one-tap write

Drafted cover note — review before apply

A factual 2-3 sentence cover note in your shop's voice, with a one-line rationale.

Edit it, then click Apply to write it into your draft manifest. You still submit the response yourself.