Agentic commerce
Every transaction scored. Every decision explained.
Risk decisions for payments initiated by AI agents. One API call returns PROCEED, REVIEW or BLOCK—with reasons—before authorization.
- What it is
- A pre-authorization risk layer
- One call in, one decision out: PROCEED, REVIEW or BLOCK, with a score and its reasons. Non-custodial, network-agnostic, and it never touches card data.
- Who it's for
- Whoever carries the risk
- Acceptance-side platforms taking payments initiated by AI agents: merchants, PSPs and micropayment platforms.
- What it solves
- Fraud tools go blind when an agent pays
- No device, no biometrics, no session, and a hijacked agent moves at machine speed. We verify the agent, check the transaction against what the cardholder actually authorized, and score it in real time. Good payments go through; off-mandate spend and impostor agents stop.
Integration
One call in, one decision out
You send the transaction, we return the decision. Four fields are required; everything else only makes the answer sharper. You never send a card number, a CVV or any payment credential.
Your request
POST /v1/assessmentsAuthorization: Bearer shai_live_...{ "platform_transaction_ref": "ord_88213", "buyer_agent": { "external_ref": "shopper-agent.acme.ai" }, "amount": { "value": "1340.00", "currency": "EUR" }, "merchant_meta": { "name": "nordic-outfitters.eu" }, "intent_summary": "Order the winter jacket"}Our response
200 OK{ "decision": "REVIEW", "trust_score": 0.41, "confidence": 0.86, "reason_codes": [ "MANDATE_AMOUNT_EXCEEDED", "VELOCITY_SPIKE_1H" ], "assessment_id": "as_9f21c8a4", "audit_ref": "sha256:4e9b...c721"}Three words out. You decide when each one fires.
Without a graded decision the acceptance side has two options: take all agent traffic, or refuse it. The output stays small on purpose, because a decision you have to interpret is not one. The work is in being right, and in being able to show why.
- PROCEED
Consistent and safe
Continue to authorization
- REVIEW
Ambiguous or conflicting
Hold, step up, or continue with caution
- BLOCK
Off-mandate or anomalous
Stop the transaction
- BLOCKED_BY_SCHEME
The network declined it
Stop. Only returned if you opt in
- You draw the lines, not us
- A platform taking EUR 5 micropayments and one taking EUR 5,000 orders should not put the boundary in the same place. Thresholds are set per platform and versioned, so you can move them and still explain what happened last month.
- You decide whether we can stop anything
- Advisory means we return the decision and you act on it. Hard block means we can halt the transaction outright. Most pilots start advisory and tighten once the scores have earned enough trust.
Every decision carries
- A score from 0 to 1
- A confidence level
- Machine-readable reason codes
- An immutable audit record
What we never receive
- Card numbers
- CVV or track data
- Usable cryptograms
The score
Six factors determine the score, and you see all of them
Two are gates: fail them and the transaction blocks regardless of everything else. The other four are weighted signals that grow more accurate as history accrues. Every one of them returns a reason code, so a decision is never a number you have to take on trust.
- F1Gate
Identity and authenticity
Signature validity, replay window and domain binding. A forged or replayed signature blocks outright.
- F2Gate
Mandate consistency
Does the transaction match what the cardholder actually authorized: amount, counterparty, scope. A mismatch is the strongest early fraud signal there is.
- F3Signal
Behaviour and velocity
Frequency, burst detection and amount drift measured against the agent’s own baseline, not a generic one.
- F4Signal
Cross-network reputation
Disputes, refunds and completions aggregated across networks. This is the part no single network can see on its own.
- F5Signal
Transaction context
Does the purchase match the instruction the user gave: the right amount, the right merchant, the right items.
- F6Signal
Counterparty risk
The seller side. New or unverified sellers, dispute history, and prices that do not match the catalogue.
0.41 / 1.00
REVIEW- Identity & authenticityF1signature_valid · agent_enrolled+0.18
- Mandate consistencyF2mandate_amount_exceeded−0.31
- Behaviour & velocityF3velocity_spike_1h−0.12
- Cross-network reputationF4counterparty_first_seen−0.04
118 ms
Median decision latency, returned before the transaction reaches authorization.
Zero
Card numbers received, stored or transmitted. Token references and metadata only.
Any network
Scheme-agnostic by design. One integration, whatever rails the transaction arrives on.
Agentic commerce
Already seeing agent traffic?
We are running early pilots with acceptance-side platforms. Tell us what you are seeing and we will tell you whether we can help.