Overview
Predicts which of your PayPal payments will get held, before they do, and fixes the cause.
PreventDiagnoseFixCode decides. AI explains. You approve.
Account hold risk
Code-scoredRisk points by cause
open ordersRefund rate by product
vs shop averageRecent activity
View audit logOrders needing attention
All ordersOrders
Every live PayPal order, scored by deterministic rules. Click a row for the evidence.
| Order ID | Age | Item | SKU | Amount | Risk | Top factor | Fix status |
|---|---|---|---|---|---|---|---|
Forecast
See how fixes and sales growth move your hold risk, using the same rules the agent uses.
Deep prevention scan
Every fix combination against sales growth, plus AI root cause per problem product. Fans out as parallel tasks on Render Workflows, with automatic retries.
Run the scan to see which fixes keep you safe as sales grow.
Risk per order, before and after the agent
Hollow dot = before, filled = nowWhat-if simulator
DeterministicProjected outcome
vs currentHold-risk score-
At high risk of a hold-
Days to limitation-risk line-
Limitation-risk line. Unhold projects your refund rate forward at the current pace. The line is the point where refunds plus disputes reach a level that commonly triggers account reviews. It is a heuristic estimate, not a PayPal threshold.
Agent & Rules
Code decides. AI explains. You approve.
1ScoreDeterministic code
2InvestigateRead orders and notes
3FixTracking, refunds
4ReportExplain and prevent
Live activity
IdlePress Run Unhold agent to start. A run takes about 1 to 2 minutes; you can switch pages while it works.
Needs your approval
Nothing to approve.
Product root cause
AI inferenceRun the agent to analyse refund notes.
Prevention checklist
Run the agent to build your checklist.
Buyer message drafts
AI draftRun the agent to draft messages.
Rules engine
Read-onlyRules are deterministic code (risk.js); the AI cannot change scores or amounts.
Audit & Settings
Every action, the exact payload hash, and how this environment is wired.
- Code decides. Risk scores, amounts and fixes are computed by deterministic rules. The model only reads them.
- AI explains. Explanations, root causes and buyer drafts are labelled as AI inference and quote your own data verbatim.
- You approve. Any action that moves money (refunds) waits for your explicit approval.
- Hash-bound approvals. Each approval is bound to a sha256 of the exact {order, capture, amount} payload; a changed payload cannot be executed.
- Low-risk writes only. Without approval the agent can only add tracking numbers that already exist in your shipping export.
- Full audit trail. Every assessment, tool call and PayPal write is logged with its payload hash.
- Cannot release holds. Unhold does not know or influence PayPal's internal risk decisions.