Who Does Your AI Shopping Agent Serve?

A preprint submitted on 23 September by researchers at Carnegie Mellon University and Microsoft Research examines shopping agents in controlled marketplaces. Its matched comparison reports buyer-optimal choices falling from 78.6% to 17.3% when commercial steering is enabled. The research covers nine environments; the headline comparison concerns specified model configurations, not consumers or live retail transactions. ¹

These are preliminary research results, not a verdict on any named payment network or retailer. The tasks have clear objectives and fixed steering mechanisms. Real buyers may have changing preferences. The authors report improvements from additional verification, but also greater running cost and limited evidence of transfer. ¹

For leaders, the useful question is whether the partnership tests buyer benefit separately from task completion. A checkout demonstration alone cannot settle that question. Nor should an early laboratory result justify a blanket ban on useful assistance.

On 9 September, Mastercard announced Agent Connect and expanded merchant tools, including work with Anthropic’s commerce blueprint. Its stated aims include making products discoverable and supporting authorised transactions. This is a supplier announcement about capabilities and partnerships, not independent evidence that buyers receive better recommendations. ²

Visa’s 2 July announcement described live agent-led purchases at participating European merchants, with consumer authorisation and bank involvement. That is evidence of a working transaction route within the programme, not a measure of how often agents selected the best available offer. ³

Anthropic’s September commerce blueprint distinguishes agents helping shoppers from agents helping merchants. That distinction should survive the contract. A seller’s assistant can legitimately explain the seller’s catalogue; it should not be presented as an unrestricted comparison service if it cannot examine alternatives. ⁴

Ask each partner to state how it earns revenue, which products the agent can see, whether placement is paid and who controls the final recommendation. Commercial incentives are not proof of misconduct. They are facts the buyer needs when deciding how much authority to delegate.

The review needs more than the platform’s demonstration. Include a commercial owner who can explain the revenue model, a payment specialist who can explain authorisation and refunds, and an evaluator who is paid to find failures rather than increase sales. Bring in customers or customer representatives who can challenge what the business calls a good outcome.

An independent university group could contribute controlled tests; a merchant could supply accurate product and service information; a consumer organisation could examine whether the promise is understandable. This is a proposed partnership design, not an announced Institute relationship or endorsement.

The UK Competition and Markets Authority’s evidence review explains how digital presentation can affect consumer choice. US Federal Trade Commission guidance addresses the recognisability of advertising. Neither establishes that a particular AI shopping service has broken the law. They provide established questions for a local review of how commercial influence is presented. ⁵⁶

Keep the tests and reporting rights separate from the sales target. Agree who can publish an unfavourable result, who must correct misleading product information and who may halt the pilot. Without those rights, an impressive advisory panel may have little practical influence.

Start with one product category and a spending limit. Use test transactions or a controlled environment, not unsuspecting customers. Write down the buyer’s requirements before browsing: the total budget, delivery deadline, essential features and acceptable trade-offs. Include cases where the right answer is to ask a question or buy nothing.

Change the order and prominence of offers while keeping the buyer’s requirements fixed. Compare the chosen item with the alternatives the evaluator can verify. Check the final payable price, delivery terms and return conditions. Record whether the agent asks before changing the brief.

Measure cost as well as quality. Additional checking can consume time and money; a low-value purchase may not justify an elaborate process. Compare the agent with a simpler search-and-confirm journey. Expand only when the result is worth the extra operating effort.

Do not make staff responsible for correcting choices they cannot inspect. Customer-service teams need enough of the decision record to resolve a complaint, while access and retention should be limited to what the service genuinely needs.

A European card-payment pilot does not establish coverage for every payment method or jurisdiction. The US and UK sources above address different forms of commercial presentation. Local advice is needed on the applicable consumer, advertising, privacy and payment duties; this brief does not propose one worldwide legal checklist.

For a cross-border launch, test the intended language, currency, delivery area, payment method and route for refunds. In markets where the agent sees only a small share of sellers, describe that limit. A restricted catalogue can still be useful if the customer understands what is excluded.

Approve the partnership against a clear promise: whose interests the agent serves, what it can compare and when it must ask. Then assign a business owner to the failures that cross organisational boundaries. The next useful conversation is with the people who can demonstrate those conditions, not just the people announcing the partnership.