AI agent refund verification: proof the refund really happened.
Your support agent tells the customer the refund is done. Pruvz reads the billing system of record, independently and read-only, and confirms that exactly one succeeded refund matches the approved amount, the currency, and the payment it was meant for. When the agent's report and the billing system disagree, the case goes to human review with the complete evidence attached.
Refunds are where agent self-reporting breaks first: the money is real, the customer is waiting, and the agent's success message is not the same thing as a settled refund. This is the exact scenario Pruvz was built around, and it is the workflow you can watch running end to end today.
Four ways "refund issued" turns out to be false.
Each of these produces a successful-looking tool call and a wrong business outcome. Only a read of the billing system itself tells them apart.
The refund is still pending
The API accepted the request, and settlement happens later. Counting success at execution time records an outcome that has not happened yet.
The refund silently failed downstream
A fraud hold, a gateway rejection, or a processor error after the 200 OK. The agent's log says success; the billing system says nothing happened.
The wrong amount or currency moved
An integration bug or a unit mismatch refunds 89.00 instead of 8.90. The action happened, but not the action that was approved.
The refund ran twice
A retry after a timeout executes the business action again. Each call looks successful on its own; the customer was refunded double.
What the refund check actually verifies.
When the agent's refund executes, Pruvz captures the claim and the decision-time policy snapshot, then a separate verification worker reads the billing and CRM systems of record inside a defined window with retries. The refund check requires exactly one succeeded refund matching the payment correlation, the amount, the currency, and the run; the CRM read is captured as a point-in-time snapshot of the customer state. A pending or unreadable source keeps the action verification-pending rather than misclassifying it as a mismatch, and the result becomes final on a terminal observation: verified, outcome mismatch, or verification failed. Only the verification worker assigns that result; the agent's claim never decides it.
Read-only access and the verification lifecycle, documentedWatch it catch a refund that never happened.
The recorded product demo runs this workflow twice with a live AI agent: one run where the refund is independently confirmed and classified as verified, and one run where the agent reports success while the billing system shows no refund, so Pruvz classifies an outcome mismatch and a human review decision closes the loop on the evidence trail. The public demonstration runs against deterministic sandbox source systems, a billing system and a CRM system with isolated state that resets per run. That is a deliberate strength of the demo: every verification path, including the failure paths, is reproducible end to end on demand, on the same verification engine and evidence model that will be applied to production systems of record with the founding design partners.
The verified run and the caught mismatch, with full transcriptThe evidence a refund leaves behind.
Every verified refund produces an evidence packet: the claim, the policy snapshot, the execution receipt, the independent billing and CRM read-backs with server-assigned trust levels, and the derived result, in an ordered, append-only sequence. The packet published on this site is a real one, captured from the product demo's refund workflow and explained field by field, with the public schema and a local validator. That is the artifact your operations, compliance, and finance teams read when someone asks, months later, whether a specific refund really happened.
Common questions.
How does Pruvz verify that an AI agent's refund really happened?
After the agent executes, an independent verification worker reads the billing system of record directly, read-only, inside a defined verification window with retries. The refund check requires exactly one succeeded refund matching the payment correlation, the amount, the currency, and the run; a successful CRM read containing all required fields is captured as an evaluable point-in-time snapshot alongside it. Only that read-back decides the result: the action is classified as verified, outcome mismatch, or verification failed, and the agent's own success report is recorded as a claim, never as proof.
Can Pruvz verify refunds in Stripe or my live billing system today?
Not off a shelf today, and this page will not pretend otherwise. The full refund verification flow runs end to end in the product demo against deterministic sandbox billing and CRM systems. Production connectors to live systems of record, such as Stripe and HubSpot, are exactly what the founding design-partner program exists to build: they will be selected and developed with the founding design partners, on the same verification engine and evidence model the demo runs.
What happens if a refund is reversed or disputed after verification?
The recorded result remains the record of what the billing system showed during the verification window, which is what an audit needs it to be. Later facts, such as a human review decision or an externally resolved dispute, are appended to the evidence trail as new entries and never rewrite the original result. Automatic re-verification of a source system after a final result is a planned, explicit capability on the roadmap; Pruvz does not claim continuous monitoring it does not do.
Does Pruvz issue, approve, or block refunds?
No. Pruvz never moves money, never executes corrections, and never approves or blocks an agent's action at runtime. It records what the agent claimed, verifies what the systems of record actually show, and routes mismatches to human review with the complete evidence attached, so your team resolves the discrepancy through normal business channels and the resolution becomes part of the record.
Run refunds you can prove.
If your agents execute refunds, credits, or billing changes in production, the founding design-partner program is open: 3-5 teams, production connectors developed together, and per-action evidence from day one. Start by watching the demo catch the refund that never happened.