When AI Agents Become the Buyer

A January 2026 Gartner survey found that willingness to let AI make a final purchase decision reached only 11 percent in the lower risk categories studied. Openness was higher when AI narrowed the choices, reaching 31 percent for household supplies and 28 percent for personal electronics. Gartner This gap matters more than any general forecast about agentic commerce. It shows that buying is a sequence of permissions, not one act that becomes automated all at once.

Early usage evidence also places commerce inside a wider set of agent tasks. Harvard Business School's Working Knowledge reported that shopping and commerce represented 10 percent of interactions in a study of hundreds of millions of anonymised requests to Perplexity's Comet assistant, behind productivity, learning and media use. The researchers describe this as early evidence from one product rather than a general consumer forecast, and retain human oversight for high consequence or irreversible tasks. Harvard Business School

How much of the decision should an agent control?

An agent can assist with research, reduce a long list, recommend an option, reserve it, place an order or manage what happens after purchase. Each step carries a different level of authority. Treating all of them as “AI buying” hides the point where convenience becomes commitment.

Harvard Business Review describes retailers preparing for journeys in which software may move from research to checkout and the buying conversation may occur between a customer's agent and a vendor's systems. Harvard Business Review That possibility creates a design requirement. It does not establish that customers have granted the same authority across every category or consequence.

We use four levels when reviewing that boundary. Assistance gathers information without excluding options. Narrowing produces a shortlist against stated criteria. Commitment reserves an option or prepares a transaction. Action completes the purchase or changes an existing service. The levels are Studio analysis, designed to help a business decide what the customer is actually authorising.

A buying agent becomes commercially important at the point where its recommendation can create a commitment. The business needs to know which facts were used, which conditions applied and whether the customer understood the action being proposed. A low cost repeat order may support a broad mandate. A professional service, financial product or long contract will usually require a more visible approval point because the downside is harder to reverse.

Where should buyer approval appear?

Current commerce systems are being designed around explicit authority. OpenAI's description of Instant Checkout states that users confirm each step, payment tokens are limited to a specific amount and merchant, and the merchant remains responsible for fulfilment, returns and support. OpenAI Mastercard describes verifiable intent as a record that connects an agent's action to the person's permissions and preferences. Mastercard

The principle extends beyond payment. Approval should become more specific as the action becomes more consequential. A customer may allow an agent to compare available plans while retaining the choice of contract term. A procurement team may permit recommendations from approved suppliers while requiring a person to confirm quantities, data terms or renewal conditions.

The design question is where that approval appears and what it contains. A confirmation that simply says “continue” offers little protection when the important choice concerns cancellation, substitution, recurring payment or a change in scope. The person needs to see the decision the agent is about to make, the material conditions and the party responsible after the action.

What happens when the agent makes the wrong choice?

An agent can complete a valid transaction that still produces the wrong outcome. The product may be unsuitable, a date may have been interpreted incorrectly, or a policy may have changed since the information was retrieved. Preparing for agent supported buying therefore includes the path back from an error.

Before allowing an agent to act, a business should be able to answer four questions.

• Which action did the customer authorise?

• Which information and conditions governed the decision?

• Can the action be cancelled, corrected or reversed?

• Who handles the problem when the result is disputed?

These answers should survive the move between the agent, payment provider, merchant and support team. Visa's Trusted Agent Protocol addresses one part of this problem by helping merchants recognise approved agents acting with commerce intent. Visa Recognition establishes who or what is acting. The business still needs clear commercial terms and a usable route for resolving the result.

What information must the offer expose?

The website and product information now have to support actions that may occur beyond the designed sales journey. An agent comparing options needs current prices, availability, scope, exclusions and fulfilment terms in forms it can retrieve accurately. The customer needs the same facts presented clearly at the point of approval. Our guide to structuring a website for AI search explains how those facts can be given a reliable public home.

For product commerce, structured feeds are becoming a direct route into agent based discovery. OpenAI's current merchant guidance asks participating partners for product data including titles, descriptions, images, price and availability so products can be represented accurately in ChatGPT. OpenAI Developers A service business will not use the same feed, yet the underlying responsibility is similar. The offer must state what can be bought, under which conditions, and what remains subject to review.

This is where brand standards become operational. Marketing cannot promise flexibility that the transaction flow removes. A helpful sales page cannot be followed by opaque renewal terms or an unclear cancellation route. When an agent acts on the available information, contradictions become transaction risk.

Where should a business begin?

Begin with one buying journey and mark the points where software may assist, narrow, commit or act. For every point, record the permission required, the facts shown to the customer, the available reversal and the owner of any dispute. Then compare those requirements with the website, product data, checkout, policies and support process.

McKinsey's work on agentic commerce reaches the same operational breadth from the merchant side. It identifies customer discovery, loyalty, commerce platforms, payments, physical service, fulfilment and returns as connected areas of preparation. McKinsey The practical implication is that an agent ready product page cannot compensate for an approval flow or support process that leaves responsibility unclear.

The lowest level of authority that still creates meaningful value is usually the sensible starting point. A company can improve comparison and recommendation before it allows autonomous action. Evidence from current consumers supports that order, and the emerging payment protocols place explicit permission and traceability near the centre of their design.

The arrival of buying agents does not make the customer disappear. It changes where the customer exercises judgment and how much of that judgment the business must preserve in the transaction. Companies that define the delegation boundary now will be better prepared to offer convenience without making control, accountability or recourse difficult to find.

If an offer is moving towards agent supported discovery or purchase, get in touch. We can review whether the public information, transaction points and customer experience carry the same decision and define where human approval must remain visible.