Skip to content
Agentic checkout control

Decide which buying actions an agent may execute, and avoid claims and disputes.

Trusteed checks who is acting, what they are allowed to do, your spend limits and the real price and stock. It allows, blocks or asks for confirmation according to your rules, and keeps the proof of every decision.

Your rules, not ours · Measure first, block later · Human approval when it is needed · Signed proof of every decision

Scope: Trusteed does not charge or process payments, does not sell on your behalf and does not guarantee the outcome of a claim. See what it does not do

The problem

Your checkout was designed for people. An agent does not behave like one.

A person hesitates at an unexpected total, stops if the price has changed and does not repeat the same attempt forty times. An agent does: it executes the instruction it received with the data it had, in milliseconds. If the instruction or the data is wrong, the mistake does not stay in an abandoned cart. It becomes an order, a charge and a call from your customer.

cost 01Orders you should never have accepted

Amounts above what was authorised, multiplied quantities, products the buyer never approved.

cost 02Sales you lose without knowing

Blanket blocking, because the only alternative to accepting everything is rejecting everything.

cost 03Support hours with no answer

Nobody can explain which rule applied, with what data, and what the buyer had authorised.

Five scenarios

Recognise yours. Each one shows its proof and its limit.

These are the five moments where a decision changes the outcome. Every scenario has a direct link: you can send it to the team that has that specific problem.

LIMITS

The agent asked for twelve units. You authorised two.

An agent can exceed the authorised amount, multiply quantities or add products the buyer never approved. The mistake reaches payment in under a second; by then it is already an order you have to handle.

Trusteed checks identity, permissions, amount, quantities, categories and speed before completing checkout. Your store decides whether to allow, block or send the operation to review.

See the rule library
PROOF · AMOUNT OUTSIDE WHAT WAS AUTHORISEDpilot
What the buyer authorisedmaximum amount100.00 EUR
Cart prepared by the agent3 lines · 12 units250.00 EUR
Deciding rulespend limit per operationR030.simple-controls
Decisioncheckout does not continueBLOCK

Limit: available on agreed integrations. Which rules can be applied depends on your platform and the data it exposes.

HUMAN APPROVAL

Not every purchase should be autonomous.

A high-value order, a first purchase or a sensitive product may need someone to confirm it. Without that step, the alternative is blocking everything or accepting everything.

Your rules send only those operations to confirmation. Without a valid approval checkout does not continue; every other order goes through with no extra steps.

How Control works
PROOF · TIMELINEpilot
01 · Agent request420.00 EUR · above the autonomous thresholdrequest
02 · The order pausesa single-use approval link is sentHITL_REQUIRED
03 · A person confirmsthe link expires once usedapproval
04 · Checkout continuesa record of who approved remainscheckout resumed

Limit: human approval is switched on during the pilot. You cannot configure it on your own today.

REAL STATE

The price it prepared the purchase with no longer exists.

The agent closed the cart with a price, a stock level or a discount that has since changed. A correct authorisation does not prevent an incorrect order if the data changed along the way.

Trusteed reads the real price and stock from your system again before deciding. If the agent's data has expired or cannot be refreshed, it asks for confirmation instead of buying with stale data.

Which signals your platform needs
PROOF · REQUESTED VERSUS REALdepends on the integration
Price the agent hadcaptured 41 minutes ago64.00 EUR
Real price in your storeread at decision time71.50 EUR
Available stock2 units requested1 unit
Decisionno buying with expired dataRECONFIRM

Limit: in development and different on each integration. It does not work on every platform; the example is a prototype, not a real response.

STOP THE CHANNEL

Something is wrong and you need to close the channel now.

Facing fraud, a misconfigured rule or an incident, waiting for a cache to expire or editing every integration separately leaves a window of risk open.

The kill switch takes precedence over any cached rule: it stops agent purchases instantly and does not touch your customers' checkout.

See platform status
PROOF · AGENTIC CHANNELpilot
Agentic channel statebefore the incidentON
Kill switch activatedtakes precedence over cached rulesOFF
Next attempt by an agentwithout waiting for anything to expireBLOCK
Human checkoutunchangedavailable

Limit: it only stops channels integrated with Trusteed. A channel that is not cannot be stopped from here.

DECISION EVIDENCE

Someone asks why it was blocked and nobody can answer.

A block with no explanation costs sales and support hours. An allowed purchase with no proof leaves you without an answer when someone asks which rule applied, with what data, and what the buyer had authorised.

Trusteed keeps the decision together with the rules in force at that moment, the one that decided, the outcome and the operation. Anyone can verify it without depending on Trusteed.

What a verifiable receipt looks like
PROOF · RECEIPT FIELDSsigned on ALLOW
Version of your rulesfingerprint of the applied rulessha256: 4b8d…21f0
Deciding rulethe one that determined the outcomeR030.simple-controls
Outcomewhat was done with the operationALLOW
Signatureverifiable without depending on TrusteedJWS · Ed25519

Limit: today the proof of allowed operations (ALLOW) is signed. Blocks, approvals and reconfirmations are in pilot. The receipt proves the record is authentic and unaltered, not how a claim ends.

How it works

Four steps, always in the same order.

No model scores or guesses. These are fixed checks over the data you provide: with the same inputs, always the same decision.

01Identify

Who is acting, what the buyer authorised them to do and with which signature. An identifier is not enough.

02Check the data

Real price, stock and cart from your system, not the ones the agent brought.

03Decide

Allow, block or send to review according to your rules, with the reason stated.

04Prove

A signed record of what was evaluated and what was decided remains, verifiable outside Trusteed.

You do not have to take our word

Change the scenario and watch it decide.

Three operations: one allowed with confirmation, one that pauses and one that is blocked. The outcome tells you which rule applied and what has to happen next.

WHAT PROOF YOU KEEP
Operation
Identifier of the evaluated operation
Rules
Exact version of the applied rules
Deciding rule
The one that determined the outcome
Outcome
Allowed · blocked · pending approval

Not shared: your limits, your lists, or the detail of rules that never applied.

POST /demo-store/mcpsimulation
{  "tool": "complete_checkout",  "amount": 6400,  "agent_status": "verified"}
policy evaluation
DECISIONALLOW WITH CONFIRMATION

The operation may continue. The user confirms before checkout.

rule R030.simple-controlsreceipt queued
What you control

The rules are yours. Trusteed only executes them.

You start in observation mode: everything is evaluated and nothing is blocked. You see what would have happened with your real orders before a rule affects a sale.

rulesWhat gets checked

Who is acting, what they are allowed to do, amount, quantities, categories and order rate.

modesMeasure first, block later

You see the effect of each rule before switching it on. Starting does not cost you sales.

limitsWhere your line is

Maximum amount without approval, first purchase, sensitive categories, acceptable rate.

reviewWho approves and when

Only the operations that need it reach a person.

Real coverage

Where it works today, and with what status.

A control is only useful where Trusteed can see the operation. This is the status per integration channel; not every rule works on every one.

Integration channelWhat it allows todayPublic status
MCPAgent connection with rule evaluationavailable
UCPPublished commerce capabilitiesavailable
A2AStore discovery by other agentsavailable
x402Payment over HTTP 402in testing
WebMCPBridge available; native mode not validatedbuilt
AP2Agent payments (Agent Payments Protocol 2)experimental

Source: the statuses are resolved by the same matrix that powers the protocol coverage page, anchored to the capability registry. See full coverage

Before you ask

What this control does not do.

What it does
  • Evaluates your rules before the order reaches payment.
  • Asks for human confirmation on the operations that need it.
  • Stops the agentic channel without affecting human checkout.
  • Leaves signed evidence of the decision on supported operations.
What it does not do
  • It does not charge or replace your payment provider.
  • It does not sell on your behalf: you remain the seller.
  • It does not guarantee you win a claim: it only proves the record is authentic and unaltered.
  • It does not cover every channel equally: that depends on your platform and the data it exposes.
  • It does not replace your records or your legal obligations.
Reasonable objections

What a merchant asks before switching this on.

What if it blocks legitimate sales?

That is why you start in observation mode: everything is evaluated and nothing is blocked. You see what would have happened with your real orders before switching on any rule.

Do I have to change my platform or my checkout?

No. Trusteed is a layer over the systems you already have. What changes is what gets checked before a delegated action is executed.

Is this available today?

The control is offered through guided pilots on agreed deployments. Policy evidence is signed on ALLOW; extending it to blocks and reviews is in pilot. Checking real price and stock is in development and depends on each integration channel.

How much does it add to checkout latency?

The checks are deterministic and run before your backend. The concrete figure depends on your deployment and is measured during the pilot; we do not publish a number we cannot stand behind in your case.

Is the receipt useful with my payment provider?

It proves the record is authentic and that nobody has modified it. Whether it is accepted in a specific process depends on the format your provider asks for and on the country. We are designing that package with partners; we do not treat it as solved.

Control pilot

Tell us which operation worries you.

No card and no account needed. Knowing which operation worries you is enough to tell whether there is a fit; if there is not, we say so in the first conversation.

IT FITS IF
  • You already receive agent orders, or you are opening the channel this quarter.
  • You can name one concrete action you want to allow, review or block.
  • Someone is going to use the proof: risk, support, finance or your payment provider.
  • You accept starting in observation mode before blocking.
Request a pilot
STRUCTURED SUMMARY · AGENT READABLE
Problem
An agent can buy above what was authorised, with expired prices or stock, or leaving no explanation of the decision.
Decision moment
Before completing checkout, after identifying the agent and reading the merchant's real price and stock.
Data it needs
Agent identity and signature, what the buyer authorised, cart lines, amount, categories, and the merchant's real price and stock.
Possible outcomes
ALLOW · BLOCK · HITL_REQUIRED · RECONFIRM
Proof it generates
Signed record with the version of the applied rules, the deciding rule, the outcome and a verifiable signature (JWS Ed25519).
Availability today
Guided pilot on agreed integrations. Proof signed today on allowed operations (ALLOW).
Known limits
It does not charge or process payments and does not sell on the merchant's behalf; it does not guarantee the outcome of a claim; checking real price and stock is in development and depends on each integration channel; signed proof on blocks, approvals and reconfirmations is in pilot.
Agentic checkout control — decide what an agent may execute