Meet COLVO: test, guard and verify AI agents that touch money
What we built, who it is for, and the three moments that matter: before launch, at the moment of action, and after the fact.
AI agents are moving from answering questions to doing things: refunding a charge, cancelling a subscription, changing a plan. The moment an agent can move money, “the reply looked right” stops being good enough.
COLVO checks what the agent did, not what it said.
Before launch: test in a copy of your world
COLVO Test runs your agent against realistic scenarios — a partial refund, a customer who wants to keep access until the end of the period, a duplicate request, a timeout on the write — in an isolated sandbox that behaves like Stripe. After every conversation, COLVO reads the account itself and compares it field by field with what should have happened.
| Field | Expected | Observed |
|---|---|---|
| cancel_at_period_end | true | false |
| access_enabled | true | false |
A green reply with a red table is a FAIL. That is the whole idea.
During: guard every real action
In production the agent does not hold your Stripe key. It proposes an action to COLVO Guard, which checks it against a mandate your backend registered: which customer, which actions, which limits. Guard answers ALLOW, REVIEW (a human decides), HOLD or DENY — deterministically, with no AI in the decision.
After: verify the outcome
When an action runs, COLVO reads the provider back and marks it VERIFIED only when the state matches. Retries never double an effect, and every step leaves evidence you can export.
Who it is for
Teams shipping support and billing assistants on n8n, Flowise, Voiceflow, Botpress, OpenAI or their own code. There is a free plan, and we are onboarding pilots by hand.
