What retail media AI actually does today, separated from what it might do later

Retail media AI applications need a distinction between documented functions and proposed uses. For England retailer-controlled advertising, a reporting assistant does not automatically imply autonomous campaign management, accurate outputs or lawful processing. This article separates a current product record from conditional applications that have not been tested.

Research was completed on 8 September 2026 using public sources. No AI account, retailer integration, output sample or shopper dataset was examined. Publisher commercial relationships and named qualified reviewers remain unverified or unassigned, so no product recommendation is made.

What the product record actually describes

Evidence class: current global first-party documentation. Criteo's In-Platform Agent guide describes reporting assistance inside Commerce Max, including sponsored-product campaign information. A publication date or version was not established; the access date identifies this inspection.

The context is media-buying reports about retailer inventory, not control of the retailer's entire service. The record says the agent does not forecast or activate campaigns. It therefore cannot support a claim that this documented function predicts England sales or independently changes campaign settings.

No advertised speed, accuracy or productivity claim is adopted as an observed result. The record establishes a described function; it supplies neither an independent test nor an adoption forecast for 2027.

Conditional retailer applications

Publisher scenario: if an authorised workflow uses AI to draft a summary of a retail-media export, a reviewer could compare each statement with the source. Evidence would need to preserve the report period, cohort, units, attribution settings and output being assessed.

A separate hypothetical use is drafting questions about missing definitions. That would be preparatory work, not an automated finding that a campaign performed poorly. Neither scenario is asserted to be a capability of the named agent beyond what its own record establishes.

Keep ad requests, returned responses, rendering and viewability distinct from attention. A summary of attributed sales cannot establish causal incrementality. Modelled statements need their assumptions and uncertainty exposed rather than being presented as observed outcomes.

Accuracy needs the right evidence

The ICO's accuracy and statistical accuracy guidance distinguishes data-protection accuracy from statistical performance. Evidence class: UK regulator guidance, currently marked under review following legislative changes. Refresh its interpretation before deployment decisions.

Do not assume that plausible language is a validated answer. The proposed house review should retain someone able to question the source, reject unsupported statements and record corrections. Human involvement alone is not presented here as automatic legal compliance or proof of accuracy.

Apply privacy and advertising boundaries

The ICO's online advertising guidance covers consent for advertising storage and access technologies. It does not permit additional collection simply because a proposed AI application could use the information.

Refer actual data purposes, recipients and permissions to qualified privacy reviewers. An aggregated report or a first-party label does not, on its own, settle every processing question. No lawful basis has been assessed for a retailer in this research.

CAP's substantiation rules concern objective marketing claims. If an AI draft becomes such a claim, the relevant evidence still needs review. The advertising code does not certify the output or prescribe this particular editorial workflow.

Preserve an answer that can be checked

Criteo's linked guide says conversations disappear when the chat closes and describes copying data or exporting a CSV file. Save any response needed for later review before closing the session.

For a proposed assessment, choose a question whose answer can be reconciled with an authorised report. Save the question, export and filters. If a reviewer cannot reproduce the comparison, describe that unresolved condition before relying on the summary in a retailer presentation.

Set a bounded next step

Define the proposed application and authorised evidence before testing it. Record observed errors and untested conditions without inventing an accuracy rate or saving. Future functionality, England adoption and retailer outcomes remain UNKNOWN. Publication requires real reviewers, confirmed conflicts and refreshed sources; this desk article supports an evidence request rather than automated release or campaign decisions.