Payment outcome assurance for AI agents

Your agent refunded them? Prove it.

Lumniverse reads Stripe and verifies every money move your agent claims — flagging the silent failures before your customer does.

$0M
agent payment claims verified
0 in 12
claims wrong on first audit
0
write access, ever
lumniverse · session 882a
watching
Agent · support-bot14:02:11

idle
Stripe · source of truth14:02:14
FALSE_SUCCESS
confidence 96
CI gate · blocked · regression case generated
Read-only access
Never moves money
Stripe + your support tools
Verdicts + IDs only — no card data
SOC 2 on the roadmap
In-VPC option for finance
Read-only access
Never moves money
Stripe + your support tools
Verdicts + IDs only — no card data
SOC 2 on the roadmap
In-VPC option for finance
Read-only access
Never moves money
Stripe + your support tools
Verdicts + IDs only — no card data
SOC 2 on the roadmap
In-VPC option for finance
Your exposure

Put a dollar number on what's slipping through. Move the sliders.

On a typical first audit, a slice of "refund processed" messages have no matching Stripe record — silent, no error thrown. Here's what that gap is roughly worth in your business.

5,000
$120
8%
70%
Exposure report · estimate
annualised
Estimated annual exposure you could recover
$403,200
400
wrong claims / mo
$48,000
flowing through lies / mo
$6,000
support + chargeback drag / mo

Estimate only, from your inputs — not a guarantee. A free teardown gives you the real number from your own Stripe records.

Get your real number — free teardown →
Positioning

We are not reconciliation.

Reconciliation tools match your bank against your ledger for the finance close. Observability tools watch the trace. Neither reads what your agent told the customer. We verify whether your agent told the truth about money it moved — and help you stop the lie from shipping again.

Reconciliation tools
Ledge, BlackLine, HighRadius…
Sees

Bank vs ledger, for the finance close.

Gap

Never reads what your agent told the customer.

Observability / eval
LangSmith, Langfuse, Helicone…
Sees

The agent trace and the model output.

Gap

Never reads the payment system of record.

Lumniverse
Payment outcome assurance
Sees

Joins the claim to Stripe and verifies it was true.

And

Blocks the repeat — gate stays client-configured.

Verdict vs Diagnosis

Anyone can shout "something broke." We show exactly where.

The same wrong answer to a customer can come from five very different places. Watch each one trace itself.

The symptom (always the same)

Customer was told "Refund processed."
No refund exists.

Five different reasons
Diagnosis · Case A
Tool

Tool said no. Agent said yes anyway.

Claimed
Step 1
Call refund tool
Step 2
Read result
Step 3
Confirm
Actual
What the customer was told
AgentYour refund has been processed.
Actual ↓
POST/v1/refunds403
{
"error": "permission_error",
"refund_id": null,
"retries": 0
}
Diffagent: "Refund processed"actual: permission_error
Fix · Never claim success unless the tool actually succeeded.
Where it fits

Anywhere your agent talks about money.

Zendesk replies. Intercom chats. Voice calls. Outbound emails. Internal Slack ops. If your agent makes a promise about a refund, charge, or cancellation, Lumniverse listens in and checks the receipt.

Zendesk · live
01 / 05
Ticket #41209
"Hey — still waiting on that refund?"
Refunded $149.00 ✓
⚠ no Stripe refund found
Lumniverse checks
Stripe · source of truth
Reality

No matching refund in Stripe for cus_9P2L in last 24h.

FALSE_SUCCESS
How it landsYour support copilot replies inside a Zendesk macro. We read the agent's reply, check Stripe, and post a private note back on the same ticket before it sends.
The answer

Six clear answers. No grey area.

Every claim your agent makes about money gets one of six verdicts. The only question that matters: did the money actually move? Every wrong answer is a dollar amount and a customer.

VERIFIED

The agent told the truth. The refund actually happened.

FALSE_SUCCESS

The agent said done. Nothing actually happened. Blocked.

DUPLICATE

Refund went out twice. We catch the second one before it lands.

MISMATCH

Right action, wrong customer or wrong amount.

PENDING

Still settling. The agent reported done too early.

UNVERIFIABLE
?

Not enough info to check. Reported honestly, never guessed.

How it works

Listen. Check. Tell you. Block the repeat.

Four steps, no changes to your agent code. We watch what your agent says, look up what really happened, and close the loop so the same mistake can't ship twice.

Step 01

Listen

We listen for the moment your agent makes a money promise — in a trace, a ticket, an email, or a chat.

Step 02

Check the receipts

We look up what actually happened in the place that knows — Stripe — and match it to the promise.

Step 03

Tell you the truth

Verified, false success, duplicate, mismatch — a clear verdict you can act on, with the evidence attached.

Step 04

Stop the repeat

Every lie becomes a regression case. Bad agent versions get blocked before they ship to your customers.

Trust & security

Built so your finance team can sign off in an afternoon.

The whole architecture is read-only and advisory by default. We reduce your exposure with earlier detection and an evidence trail — we never take it on.

Read-only, always

Restricted Stripe API key, read scopes only. We can't issue, alter, or move a payment.

Structurally can't move money

No write scope is ever requested — impossible by design, not by policy.

In-VPC option

Run the verifier inside your own network. Credentials never leave; only the verdict comes back.

Verdicts + IDs only

We store the verdict and object IDs — not card numbers, not full payloads. SOC 2 on the roadmap.

Why it matters

One quiet miss costs more than the agent ever saved you.

A confidently wrong refund becomes a chargeback, a churned customer, and a long Slack thread. Catching it the moment the agent says it keeps the bill — and the trust — intact.

0 in 12
agent money claims we flag as wrong on first audit

On the typical first audit, roughly one in every twelve refunds, charges or cancellations the agent reported as done didn't actually happen — each one a real dollar amount on the line.

0 days
average time a silent miss sits before a human notices

Without reconciliation, a false success lives quietly in the inbox until the customer chases it — usually a week later, usually louder.

0
write access required to start

Lumniverse never moves money. Read-only keys, audit only — your finance team can sign off in an afternoon.

Trusted by teams shipping money-touching agents
DTC CO
FINTECH
MARKETPLACE
SAAS
VOICE AI

"We thought our refund agent was clean. The first teardown found $80k of refunds it said it sent that Stripe never saw."

Head of Support Ops, DTC brand (illustrative)

$0M
agent payment claims verified
0 / 100
false-successes caught per audited batch
< 0 min
to first verdict on a connected source
Getting started

Start at the level you trust.

You don't have to hand us live access on day one. Most teams start with a demo or a one-off audit, then turn on the always-on version once the verdicts speak for themselves.

Path A01/04

Demo on our data

Synthetic refunds and tickets with planted mismatches. See the verdicts in five minutes.

Access · None
Path B02/04

Audit on your data

Drop ~100 of your tickets plus a Stripe export. We hand back a real exposure report in dollars the same day — this is the free Diagnostic.

Access · File drop
Path C03/04

Always-on

Connect one claim source and Stripe with read-only keys. Verification turns on in about five minutes.

Access · Read-only keys
Path D04/04

Inside your network

Run the verifier in your own VPC. Credentials never leave. Only the verdict ships back.

Access · In-VPC

The idea

Agents talk a good game. Money leaves a paper trail. We compare the two — and only ship the agents that match.

Read-only. We never move your money — we just prove whether your agent did.