Agent verification & audit · Built in Europe

The flight recorder for AI agents.

Relay Verify sits in front of any AI agent, blocks the wrong answer before it reaches your customer — invented prices, medical or legal overreach, PII leaks, off-policy promises — and writes a tamper-evident record of every action. So you can ship agents you can actually prove behaved.

Get early access See how it works
Model-agnostic · Works with any agent, any channel · EU AI Act ready
Relay Verifychecking agent reply · 41 ms
BlockedInvented price “€49/mo”Not found in your approved pricing — held before send.
BlockedMedical advice detectedReply gave a dosage — escalated to a human.
PassedBooking confirmed for Tue 10:30Matches the calendar; policy-compliant.
LoggedSigned audit record #A7F3Who acted, what was allowed, proof it happened.
The problem

You can't ship an agent you can't prove.

Every team racing to deploy AI agents hits the same wall: one hallucinated price, one piece of medical or legal advice, one leaked customer record — and it's your liability, with no record of what happened. Guardrails are an afterthought bolted on by hand, and nobody can answer “prove the agent behaved.”

1 in 5
replies drift off-policyAgents invent prices, promise refunds, or answer outside their remit — quietly, at scale, in every conversation.
€0
proof you have todayWhen a regulator, a client, or a lawyer asks what your agent said and whether it was allowed, most teams have no auditable answer.
Aug 2
the deadline is hereEU AI Act Article 50 transparency duties apply from 2 August 2026. Agents talking to people must disclose they're AI — and you must be able to prove it.
How it works

Intercept. Verify. Prove.

Relay Verify sits between your agent and your customer. Every draft reply passes through it in milliseconds — checked against your policy, blocked or allowed, and logged forever. No model change, no rebuild.

1 · InterceptEvery reply, before it's sentYour agent drafts an answer — over phone, chat, WhatsApp, or email. Relay Verify catches it before the customer ever sees it.
→
2 · VerifyChecked against your policyDeterministic rules plus your own knowledge base test the reply: real prices, no medical/legal overreach, no PII, promises within policy. Fail → block or route to a human.
→
3 · ProveA tamper-evident recordEvery action is written to a signed, immutable log: what was said, whether policy allowed it, and proof it happened — ready for a regulator, a client, or a court.
What it catches

The failures that become your liability.

€
Invented prices & termsBlocks any figure or commitment not found in your approved pricing and policy.
✚
Medical adviceFlags diagnoses, dosages, and treatment guidance — escalates to a human instead.
§
Legal adviceCatches definitive legal claims the agent isn't authorised to make.
🔒
PII leaksStops the agent from exposing another customer's data or sensitive records.
✎
Off-policy promisesRefunds, guarantees, and commitments outside what you actually offer.
⇄
Inconsistent actionsA booking or order that doesn't match your calendar, inventory, or the conversation.
🌐
Wrong language & toneReplies drifting off the customer's language or your brand's register.
🤖
Missing AI disclosureEnsures the “you're speaking with AI” notice required by the EU AI Act is present.

Checks are configurable per agent and per policy. Start in observe-only mode to see what your agent would have been blocked on — then switch on enforcement when you're ready.

Built for regulated teams

Where a wrong answer is a lawsuit.

Relay Verify is built first for teams deploying agents in health, legal, and finance — where “the AI said something wrong” isn't an embarrassment, it's exposure. Every conversation becomes defensible evidence instead of a liability.

Healthcare & clinicsLegal & professional servicesFinance & insuranceRegulated marketplaces
2 Aug 2026 EU AI Act Article 50 transparency obligations become enforceable. Relay Verify handles AI disclosure and keeps the audit trail that proves compliance.
Integrate

Drop it in front of any agent.

Relay Verify is model- and vendor-agnostic. Point your agent's output through it with a single call, or run it as a proxy — no rebuild, no lock-in.

Any model, any stackWorks with GPT, Claude, Gemini, open models, or your own — plus voice platforms like Retell and Vapi. Verify wraps the output, not the model.
Any channelPhone, chat, WhatsApp, email, in-app. One policy, enforced everywhere your agents talk to customers.
One API callSend the draft reply to POST /verify; get back allow, block, or escalate plus a signed log id. Sub-100ms.
Observe, then enforceStart read-only to measure your risk, share the report with compliance, then turn on blocking with confidence.
Early access

Deploy agents you can prove behaved.

We're onboarding a first cohort of design partners in regulated verticals. Tell us about your agent and we'll set up a verification pilot on your own traffic.