Who watches an AI shopper after approval? DataDome and Experian add checks during the visit

When you ask an AI agent to shop for you, is knowing whose agent it is enough?

An identity check at the entrance does not, by itself, explain whether every subsequent action stays within your instructions.

The DataDome and Experian partnership brings together the connection between a person and an agent, and checks on the agent's behavior during the visit.

Experian announced that DataDome had joined Agent Trust on October 1, 2026 (Announcement).

This article explains the division of responsibilities and suggests practical checks for an online store.

The partnership checks what happens during AI shopping

Experian establishes the relationship between the consumer and the agent, while DataDome assesses activity during the visit.

The idea is to connect who the agent represents with what it is doing now (DataDome's partnership overview).

Experian checks whom the agent represents

Experian links a verified consumer to the AI agent acting for them.

Its explanation calls this Human-to-Agent Binding and describes a token issued at authorization.

Identity, however, is not permission to make every possible purchase.

An agent that legitimately represents you may still have a budget, a specific delivery address, or permission only to compare products.

A store should keep those purchase conditions separate rather than infer them from identity alone.

DataDome checks what the agent does next

DataDome evaluates requests and behavior throughout the visit.

The partnership connects those observations to the authority delegated to the agent (Official division of responsibilities).

A request is an individual interaction with the site, such as opening a product page or submitting information.

There is still something to check after an identity check succeeds.

Showing a membership card at a store does not grant permission to do anything inside it.

That analogy explains the distinction; it is not a description of a retail security implementation.

A verified identity does not settle every action

Stores need to distinguish evidence of identity from evidence that a particular operation is authorized.

A single label saying “safe” can hide which checks have actually happened.

Identity and authorized actions answer different questions

An agent linked to the correct person can still perform an action outside the intended shopping conditions.

Consider a request to compare chairs within a budget.

Looking up prices, adding a chair to a cart, and paying for it are different actions.

The store needs more than evidence that the same visitor is continuing a session.

It also needs to know whether the current action fits the user's instructions.

The following is an editorial worksheet, not a description of either vendor's configuration fields.

Question	Example evidence	What it does not establish
Whom does the agent represent?	A person–agent relationship	Approval for this purchase
What happened?	Viewing, changes, and order records	Whether a change was wanted
What was delegated?	Product scope, budget, and expiry	Whether later actions complied

A trust score is not consent to purchase

DataDome documents an existing feature called Agent Trust Score.

Its dynamic 0–100 score evaluates factors including identity and behavior; it is not a buyer's instruction to place an order (Agent Trust Score documentation).

For example, a high score cannot establish that a person has seen the final total.

Write separate rules for admitting an agent and for completing a purchase.

The score documentation describes an existing DataDome feature, not a newly announced Experian setup screen.

The partnership announcement does not establish that choosing one score threshold completes an integration.

Three checks across a shopping journey

Review the agent's authority, its intervening actions, and permission for the final order as separate checkpoints.

The chair purchase below is a fictional example for building your own worksheet.

Follow the actions while the agent searches

During product selection, record the products viewed and conditions changed.

Start with one agent, one test product, and one shopping instruction.

Use an available test environment rather than a real customer's order.

1. Save the instruction given to the agent, such as finding a chair with armrests.
2. Check the sequence and times of the product listing, product page, and cart visits.
3. Compare changes to color or quantity with the original instruction.
4. If the journey stops, save the error and the preceding action.

A blocked agent and an unavailable product require different fixes.

Record the stage reached instead of relying only on the number of completed orders.

Match order conditions to the buyer's permission

Before completion, compare the product, quantity, total, delivery conditions, and permission given.

Changing one chair to two creates a different order from the original quotation.

Shipping costs or the destination can change even if the unit price stays the same.

Keep order details and authorization evidence distinct in the worksheet.

Item	Example check
Order details	Product, quantity, total, destination, delivery conditions
Scope of approval	Which version of the order was approved
Timing	Whether approval preceded or followed a change
Outcome	Order reference, incomplete status, or cancellation

This does not require copying every piece of personal or payment information into a new log.

Decide what evidence is needed and where authorized staff can access it.

What a Japanese online store should verify before adopting it

A Japanese retailer should first establish which parts of its contracts and checkout infrastructure are supported.

The partnership page reviewed on October 4, 2026 does not establish universal Japanese availability or pricing.

Confirm scope, costs, and required contracts

Provide the vendor with both your site architecture and the operations you want agents to perform.

Specify product search, login, cart management, or checkout rather than asking only whether AI is supported.

You could use the following inquiry.

We operate an online store for customers in Japan.

Which stages from product browsing to order completion does this partnership support for our setup?

Please also explain supported agents, required contracts, additional fees, implementation work, and test environment requirements.

Record answers as available, subject to consultation, or unsupported.

Do not present a roadmap commitment internally as a feature already available today.

Check that ordinary shoppers can still buy

After adding agent controls, verify that human shoppers can still complete purchases.

Fewer agent requests are not necessarily an improvement if ordinary orders also fail.

Run separate human-operated and agent-operated test journeys.

Compare product display, cart changes, confirmation, and completion in the same worksheet.

When an error occurs, capture the time, page, action, and message for the responsible team.

Adjust controls within a scope where their effect can be observed.

Check your own readiness for AI commerce

The practical lesson is to assess trust throughout an agent's visit, not only at the entrance.

The announcement does not mean every retailer must immediately adopt the same products.

Ask your team three questions.

• Where do we receive evidence of whom an agent represents?
• Which records let us follow its actions from selection to ordering?
• Which version of the order is connected to the buyer's final approval?

An unanswered question identifies the next area to investigate.

Helping AI discover products and designing a reliable purchase journey are connected tasks.

For the checkout side, see our explanation of Shopify's WebMCP checkout.