Watch Pruvz catch what the agent trace missed.
A real AI agent executes a refund and reports success. Pruvz independently reads the billing and CRM systems back, verifies the real outcome, and catches the run where the refund never happened. Watch the complete flow of AI agent outcome verification against systems of record: agent execution, policy snapshot, evidence, classification, and human review.
The recorded end-to-end demo, 4 minutes 27 seconds: a real AI agent works a refund case, Pruvz independently verifies it, catches a mismatch the agent reported as success, and a human review decision closes the loop. The billing and CRM systems shown are deterministic sandbox source systems, so every path in the video is reproducible on demand. Also on YouTube; full transcript below.
Watch it live
We run the full demo in a 30-minute session: verified outcome, caught mismatch, human review decision, and the business overview built from verified results. Bring one of your own agent workflows and we'll map it to the same flow.
Four verification paths available in the live walkthrough.
Most tools demo the happy path. Pruvz demos what makes verification trustworthy: what happens when the outcome is late, wrong, or unreadable. The recorded demo above focuses on the verified and mismatch paths; the live walkthrough also demonstrates delayed outcomes and unreachable source systems.
Verified outcome
The agent approves and executes a refund. Pruvz independently reads the billing and CRM systems of record, confirms the money moved and the case closed, and marks the action Pruvz Verified with a complete evidence packet.
VerifiedOutcome mismatch, caught
The agent claims it refunded the customer, but the system of record shows no refund. Pruvz flags the mismatch, routes it to the review queue, and a human decision is recorded on the evidence trail, without rewriting the original result.
Outcome mismatchDelayed outcome, no false positives
The refund lands late. Pruvz keeps the action pending and retries the system of record until the outcome is terminal, then verifies it. A slow system is never misclassified as a business mismatch.
Pending, then verifiedSource system unreachable
The system of record cannot be read within the verification window. Pruvz records a verification failure: a technical outcome, clearly separated from a business mismatch, with the attempts on record.
Verification failedReproducible by design
The demo runs against deterministic sandbox billing and CRM source systems, so every verification path above, including the failure paths a live-system demo can rarely show on demand, runs end to end, deterministically, in every session. The same verification engine, evidence model, and review flow will be applied to selected design-partner workflows; production connectors for live systems of record, such as Stripe and HubSpot, will be selected and developed with founding design partners.
What a verification result means
A Pruvz result reflects what the systems of record showed during a defined verification window: Pruvz retries until it observes a terminal outcome or the verification window expires, then records the result as final together with its evidence. If a source system changes later, the record still preserves what Pruvz observed, when it was observed, and the evidence behind the result. Pruvz does not silently re-verify in the background; re-checking a final result is an explicit, planned capability.
Ready to run it against your own workflow? Become a founding design partner.
Demo video transcript
The full narration of the recorded demo above. Timestamps start the player at the matching chapter.
0:00 The question: did the agent's action really happen?
AI agents no longer just answer questions. They issue refunds, update CRM records, and act inside real business systems. Every number on this screen comes from an action an AI agent claimed to complete. The question that matters is simple: did those actions actually happen? Pruvz answers it, not from what the agent says, but from what the systems of record show. This is a working, end-to-end product demo. The billing and CRM systems you'll see are deterministic sandbox source systems, built to demonstrate the verification workflow, which means every path in this video, including the failures, is reproducible on demand.
0:44 A real AI agent takes a refund case
This is a real AI agent, backed by a live language model, receiving a customer refund request. It works the case through business tools: it reads the case from billing and CRM, checks the refund policy (purchase age, amount limit, payment state, an open ticket), then executes the refund, updates the support ticket, and closes the case. Notice the label on the agent's report: Claimed. In Pruvz, nothing an agent says about its own work counts as proof. The agent is done, and confident it succeeded. Now a separate question begins: did it really happen?
1:20 Independent verification and the evidence record
Pruvz now independently reads the systems of record back. Not the agent's logs, not its trace: the billing system itself. While that read-back runs, the action is Verification pending. Pending is eventual consistency, not a verdict. Within seconds, the ruling arrives: Verified. And here is the full record. Two separate state machines: execution (did the agent finish its job) and verification (did the outcome really land in the source system). A policy snapshot, frozen at decision time, showing exactly which rules this refund was decided under. And an ordered, append-only evidence timeline: the agent's claim, then the execution receipt, each with a trust level assigned by the server, never by the caller. This is what a verified business action looks like.
2:11 Same agent, and a mismatch only read-back can catch
Now the same agent, with the same request, one more time. Everything the agent sees is identical. It executes the refund, receives a receipt, updates the ticket, closes the case, and reports success. But this time, the billing system quietly recorded no refund. If your source of truth is the agent's report, or its logs, this failure is invisible. The trace looks perfect. Pruvz reads the billing system back and finds no refund for that payment. The ruling: Outcome mismatch. What was expected, what was claimed, and what was independently observed, recorded as evidence, side by side. No fabricated outcome. Nothing averaged away.
3:00 A human decision closes the loop
Every mismatch enters a review queue for a human decision. The reviewer sees the evidence, not a summary of it, records a decision, and documents the reason. That decision is appended to the evidence ledger as a new item. The original mismatch is never rewritten, and never deleted. The loop is closed, and the full history stays intact.
3:29 The verified business record at scale
At scale, this becomes the trusted business record behind your most important agent actions. Verified value counts only what was independently confirmed in the systems of record. Exceptions and their resolutions stay visible. And every aggregate drills down to the evidence behind it. That is the difference between a metric and a record.
3:51 What this demo is, and the design-partner program
Everything in this video is the working Pruvz product demo, end to end: action capture, independent read-back, classification, ordered evidence, and human review. The billing and CRM systems here are deterministic sandbox source systems; that is what makes every path reproducible. Applying the same verification engine and evidence model to real systems of record is exactly what our founding design-partner program is for, and the program is open. If your team runs AI agents that act in real business systems, see what verified looks like at pruvz.ai.