What an AI commerce watch found

Are people ready to buy through AI assistants, and are companies building ways for agents to pay? Those are different questions.

A Grex research agent followed both. Grex coordinates AI work and records its output for review. This assignment watched agentic commerce: shopping and payments involving AI agents.

The report below draws on the twenty-fourth recorded research pass, dated July 29, 2026. It describes that historical snapshot rather than claiming to be a current market survey.

How the research loop worked

The agent was scheduled four times daily. It began by gathering background information about the market.

A ledger tracked coverage and whether each pass found something new. Once background coverage was sufficient and new information became scarce, the assignment switched to lighter monitoring for changes.

Each factual claim needed a source. A pass without adequate grounding withheld its brief. Findings entered a review queue; the agent did not publish them directly.

Twenty-four recorded passes should not be read as proof of uninterrupted four-times-daily execution for a month. The useful evidence here is the dated brief and its cited material.

Shopping adoption and payment tools told different stories

The brief cited Forrester’s report of weak uptake for some checkout experiences inside AI assistants. Forrester discussed changes to OpenAI’s checkout approach and Walmart conversion results that compared unfavorably with buying through its website. Those were reported examples, not a measurement of all AI commerce. Forrester’s analysis

At the same time, Stripe announced tools for merchant catalogs, agent access, and an agent wallet with spending approvals. Its Sessions 2026 announcements also described payment infrastructure for agent transactions. These announcements established product investment; they did not establish widespread consumer adoption. Stripe’s announcement

The research team’s interpretation was that consumer shopping behavior and payment infrastructure were developing at different rates. That remained an interpretation of the sources.

Approval controls were part of the story

Stripe’s agent wallet included approval and purchase-visibility features. Their presence was relevant to Grex’s own question: what limits and records are needed when software can take consequential actions?

An announced control is not proof that every implementation is safe. It is a concrete feature a later research pass can inspect and compare.

What the next pass needed to resolve

The brief left open questions about checkout migrations, merchant integrations, and the availability of payment-protocol specifications. Those questions became follow-up research targets.

It also collected material on chatbot regulation. That requires a separate, scoped legal review; companion-chatbot rules cannot simply be treated as requirements for every commerce agent.

The loop’s output was a source map and a set of questions. It could still propagate an error from a cited source. Its value depended on keeping the sources visible and distinguishing their claims from the team’s conclusions.