For merchants with AI agent traffic

Stop wrongly declining your AI agent customers.

Every decision, provable.

AI agents from ChatGPT, Perplexity, and Operator are buying from your store. Your fraud stack is wrongly declining 1 in 10. We help you recover that revenue, with a tamper-evident audit trail for every decision.

No credit card · 20 min · See if there's recoverable revenue in your declines.

11%

of legitimate AI agent orders are wrongly declined by today's fraud stacks

$1.5T

in projected agent commerce volume by 2030 (Juniper Research)

<100ms

decision latency — never the reason your checkout slows down

The problem

Your fraud stack was built for humans.

AI agents behave like attackers: new device, no browsing history, programmatic cadence. They're your best customers, and your stack is rejecting them.

When an AI shopping agent completes a purchase, your fraud model sees no session cookie, no prior browsing history, an unusual device fingerprint, and a checkout flow that doesn't click around. Every signal points to "fraud."

But the agent is acting on behalf of a real customer who authorized it. The order is legitimate. The decline is a false positive — and the customer may never know why their agent failed, so they never retry with you.

Retrieva gives that agent a cryptographically verifiable identity before the authorization request reaches your stack. Trust the agents your customers actually use.

Sample day · $20M GMV merchant

DEMO live

Declined today $0

What we read

The signal stack.

WBA_VERIFIED Cryptographic identity verified via RFC 9421

ASN_KNOWN_AGENT Network origin matches a known AI agent provider

VELOCITY_LOW Transaction velocity below merchant baseline

PRINCIPAL_HASH_MATCH Buyer identity hash matches verified principal

FAST_RULE_V1 Deterministic rule overrides the classifier

How it works

An afternoon to integrate. A day to see results.

Drop Retrieva in alongside your existing fraud stack — it adds a trust layer specifically for agent traffic without replacing anything.

AI agents register a cryptographic keypair at deployment time. The public key is anchored to a tamper-evident ledger entry alongside the agent's declared capabilities and merchant authorization scope.

<50ms globally

Score in real time

At checkout, the agent signs the authorization request. Retrieva verifies the signature, evaluates behavioral and network signals, and returns a trust score before the order hits your fraud stack.

<100ms p99

Decide with evidence

Your policy rules run against the trust score. Approve, challenge, or decline. Every decision is written to the ledger with full signal provenance — auditable forever, exportable on demand for disputes.

Evidence-grade

Designed around trust

Six invariants we won't compromise on.

INVARIANT 01

Fail-open by design

If Retrieva is unreachable within the timeout window, the authorization flow proceeds without our signal. Your checkout never waits on us. We configure a safe default — approve and flag for async review.

INVARIANT 02

Sub-100ms decisions

Every evaluation is p99 under 100ms globally, measured from your webhook receipt to our response. We instrument this continuously and page on-call if it drifts above 80ms.

INVARIANT 03

Cross-PSP, never locked-in

Retrieva is payment-processor agnostic. It sits upstream of Stripe, Adyen, Braintree, or your custom gateway. Switching processors doesn't touch your Retrieva integration.

INVARIANT 04

Tamper-evident ledger

Each decision is chained to the previous one via cryptographic hash. Mutating any historical record breaks the chain at that point, detectable by any verifier. No central authority can revise history.

INVARIANT 05

PCI scope-free

Retrieva never touches cardholder data. The authorization request includes agent identity, order amount, and merchant ID — not card numbers, CVVs, or billing addresses.

INVARIANT 06

Shadow-mode safety

Every new policy deploys in shadow mode first — it scores traffic but doesn't affect decisions. You see the approval rate delta before any real orders are impacted. Promotion requires six real-outcome gates.

Integration

An afternoon to integrate.

Add the Retrieva ACP handler to your existing authorization webhook. No new infrastructure. No database migrations. No rip-and-replace of your fraud stack.

Works with Stripe, Shopify, BigCommerce, Adyen, Braintree, and custom stacks. TypeScript and Python SDKs. Full OpenAPI spec included.

Sits in your existing webhook flow — not a proxy
Fail-open by default; timeout window is configurable
Average integration time: 2 engineering hours
Full audit trail from day one, zero-config

checkout.ts — your ACP /complete handler

// checkout.ts — ACP /complete handler
import { Retrieva } from '@retrieva/sdk'

const retr = new Retrieva({
  apiKey: process.env.RETRIEVA_API_KEY
})

export async function POST(req: Request) {
  const { agentId, merchantId,
          amount, currency } = await req.json()

  try {
    const decision = await retr.decisions.evaluate({
      agentId,
      merchantId,
      transaction: { amount, currency }
    })

    if (decision.verdict === 'APPROVE') {
      return Response.json({
        status: 'approved',
        decisionId: decision.id
      })
    }

    return Response.json({
      status: 'declined',
      reason: decision.reason
    })

  } catch (err) {
    // Fail-open: approve + flag for async review
    return Response.json({
      status: 'approved',
      fallback: true
    })
  }
}

ROI Calculator

What are false declines costing you?

Adjust the inputs to your store and see your estimated recoverable revenue.

Monthly GMV ($)

Agent share of traffic (%)

False decline rate (%)

11%

Estimated recoverable revenue

per month, conservative estimate

Monthly agent GMV $0

False-declined revenue $0

Recovery rate (30%) $0 / mo

Annualized recovery $0

FAQ

Common questions.

What exactly is an "AI agent" in this context?

An AI agent is any software system — a shopping assistant, procurement bot, or subscription manager — that initiates purchase requests on behalf of a human user. They're already operating in production at companies like Perplexity, Anthropic, and OpenAI. By 2026 they'll represent a meaningful fraction of e-commerce GMV.

Does this replace my existing fraud stack?

No. Retrieva runs alongside your existing fraud tools (Stripe Radar, Kount, Signifyd, etc.). It adds a trust signal specifically for agent traffic. Your existing rules and models handle everything else. Think of it as an additional input to your authorization decision — not a replacement.

How does the tamper-evident ledger work?

Each agent registration and authorization decision is hashed alongside the previous entry's hash, creating a chain. Mutating any historical record breaks the chain at that point — immediately detectable by any verifier. This means every decision is permanently attributable to the state of the chain at the time it was made. The ledger is signed by AWS KMS and auditable by you at any time.

What happens if Retrieva is down?

Retrieva fails open. If the evaluation API doesn't return within a configurable timeout (default: 150ms), your authorization flow receives a null signal and proceeds normally. We maintain a 99.95% uptime SLA for production customers and page on-call for any p99 latency regression above 80ms.

Is customer PII shared with Retrieva?

No. Retrieva evaluates agent identity signals, not customer identity signals. The authorization request includes the agent ID, order amount, and merchant ID — not customer name, email, payment details, or address. Customer PII never leaves your systems.

What's the pricing model?

Retrieva is priced per authorization evaluation. Early access design partners receive flat-rate pricing during the beta period (Q3–Q4 2026). General availability pricing is per-evaluation with volume tiers. We publish full pricing at GA — no surprises.

When can I get access?

We're onboarding early access design partners now. We prioritize mid-market DTC, travel, ticketing, and digital services merchants who can give us real-outcome feedback during the beta period. Book a 20-minute call and we'll review your decline patterns together — no slides, no pitch deck.