When Your Next Customer Sends an AI Agent to Buy

Today, a customer may ask an AI assistant which product to buy. The more consequential possibility is that the assistant eventually carries out the purchase.

That changes the marketing question. A brand would need to persuade a person, provide accurate information a system can interpret and support a transaction made through an intermediary it does not fully control.

The future is easy to describe extravagantly. The current evidence deserves more restraint. Protocol announcements and early product features demonstrate work on the infrastructure. They do not establish that most Indian customers are delegating purchases or that every merchant should rebuild its store immediately.

For an Indian business, the useful preparation is to make the product and transaction intelligible before paying for a speculative integration.

What has actually been announced?

On 11 January 2026, Google announced the Universal Commerce Protocol, or UCP, as an open standard spanning discovery, buying and post-purchase support. The announcement named co-developers including Shopify, Etsy, Wayfair, Target and Walmart. Its initial checkout description concerned eligible US retailers, with further expansion described as planned.

That is a dated launch statement. It is not evidence that the feature is generally available to Indian merchants as of this article, nor that every planned capability has shipped.

Google's technical explanation provides a closer view of the proposed interoperability. In a separate April 2026 announcement, Google said it was donating the Agent Payments Protocol to the FIDO Alliance.

These developments concern related but distinct parts of the system. A commerce protocol and a payment-authorisation protocol should not be treated as interchangeable names for a shopping chatbot.

More recent Indian research also needs careful interpretation. NIQ's September 2026 report landing page describes an August study across eight major metros and 16 categories. It publishes headline findings about AI-assisted shopping, but the public page does not provide enough sampling detail to assess national representativeness. We have not purchased the full report. Using AI for a comparison is also different from authorising an agent to spend money. The research is a useful signal to investigate, not a basis for claiming that India has broadly delegated checkout.

Separate advice, action and authority

An assistant that compares three products is providing advice. An assistant that adds one to a basket is taking an action. An assistant that pays has exercised some form of authority over a transaction.

Those stages need different safeguards and business information. A persuasive comparison does not establish permission to spend money. A valid payment does not establish that the selected product fits the buyer's intention.

Imagine a hypothetical request to buy a replacement charger under ₹2,000. The assistant needs to understand compatibility, total price, delivery and the customer's approval conditions. A product named “universal charger” is not enough to settle those questions.

The example illustrates the information problem. It is not a claim that a particular assistant can reliably complete that task today.

For merchants, the distinction matters because a sale through an agent still creates expectations about the actual goods and service. The intermediary does not remove the need for accurate product records and a workable support route.

The first preparation is surprisingly ordinary

Make product names, specifications, variants, prices and availability consistent across the places you control. Explain delivery restrictions and return conditions in understandable language.

These tasks are useful for human shoppers too. That makes them less dependent on a particular forecast about agent adoption.

A product catalogue containing contradictory dimensions is already a problem. An automated intermediary can carry that contradiction into a recommendation without noticing what a knowledgeable employee would question.

Give important records an owner and update process. Know which system is the source of truth for stock and which record governs the offer. A new interface cannot compensate for unresolved internal disagreement.

Our article on AI productivity examines the same principle inside a workplace: reliable assistance depends on the information and checking process around it.

Machine-readable does not mean human-unreadable

Structured product information can help systems interpret a catalogue, but the visible page still needs to help a person understand the purchase. Treat them as two views of the same underlying facts.

Follow current platform documentation for feeds, structured data and integrations. Do not buy a package of invented “AI schema” claims without asking which actual system consumes it and what requirement it fulfils.

Google's search documentation for AI features explicitly says there is no special schema or extra technical requirement for appearing in its AI search features. Commerce integrations are a separate subject and may have their own requirements.

That separation prevents a common sales confusion: a genuine new commerce protocol does not prove that an unrelated website needs a secret SEO file.

Ask the supplier to distinguish search eligibility, catalogue quality, transaction integration and payment authority. Each should have a defined purpose and a testable result.

Product truth may become part of the competitive advantage

If a customer describes a need in detail, an intermediary needs evidence to match products to it. Vague promotional language contributes less than accurate attributes and clear limitations.

Consider a hypothetical office-chair seller. “Premium comfort” is a claim the customer cannot easily evaluate. Dimensions, adjustment range, materials, weight limits and applicable warranty details provide more concrete information, provided they are accurate.

This does not mean brand meaning disappears. People may still care about reputation, design and previous experience. Our interpretation is that operational product information can become more important alongside those factors.

The same opportunity exists in service businesses, although automated purchasing may be less suitable when scope requires inspection or negotiation. A clear description of eligibility and process can still improve research and qualification.

Do not pretend every service can be reduced to an instant checkout. Some decisions should begin with a conversation.

Trust will need several layers

The customer needs confidence that the agent understood the request. The merchant needs confidence that an action was authorised. Payment providers need evidence relevant to their own responsibilities. Everyone needs a route to resolve an error.

A protocol can address a defined part of that problem. Its existence does not prove that the full commercial and customer experience is solved.

For a merchant evaluating an integration, ask who is the seller of record, which party receives the customer's approval and how refunds, disputes and support work. Verify the answers in current documentation and the actual agreement.

Also ask what happens when the agent selects an unsuitable variant or the stock changes during the transaction. A demonstration that succeeds on an ideal path says little about those failures.

The business case should include the work required to handle them. An automated purchase still needs accountable operations.

Three plausible adoption scenarios

These are editorial scenarios, not market-share forecasts. They can coexist across categories and customer groups.

Routine replenishment may raise different questions from an expensive, subjective or safety-sensitive purchase. A business should examine its own category rather than assume that one shopping demonstration represents the whole market.

The most durable preparation across these scenarios is information quality and clear accountability. Costly integration work should follow evidence of customer use and supported availability.

Brand marketing will still have a human audience

It is tempting to imagine a future in which machines choose entirely on specifications and price. That is one possible simplification, not an established description of customer behaviour.

People can instruct an assistant to prefer a brand, avoid a company or choose an attribute that reflects their values. Prior experience and trust may influence those instructions.

The question becomes how the business earns a place in the person's preferences while making the product easy to assess. Advertising that builds familiarity and product information that supports evaluation can serve different parts of the journey.

Our Patagonia Worn Wear analysis examines how an operational commitment can make a brand promise tangible. A machine-readable catalogue would not replace that relationship; it would need to represent the offer accurately.

Avoid declaring that conventional brand work is obsolete on the basis of a new protocol announcement.

Measurement needs to follow the actual transaction

If an assistant becomes an intermediary, existing referral and attribution reports may not describe the whole path. A customer could research in one tool, buy elsewhere and return directly later.

Do not treat unclassified traffic as proof of agent activity. Ask what evidence the integration or platform actually provides, which events are recorded and what remains unknown.

For a pilot, measure completed suitable orders, errors, support effort and contribution. Record how the customer authorised the purchase and whether the business received the information needed to fulfil it.

Attributed revenue can still differ from incremental revenue. Our retail-media analysis explains why proximity to a transaction does not establish that an intermediary created the demand.

The aim is to know whether the channel adds useful business at an acceptable cost. A count of agent interactions alone does not answer that question.

A readiness review before an integration proposal

Start with a sample of real products or services. Compare the website, internal records and marketplace listings. Resolve contradictions in specification, price basis and availability.

Then inspect the purchase and support journey. Can the customer identify the total cost and relevant terms? Does the business know who owns a failed transaction or incorrect item?

Ask an integration vendor to demonstrate the actual supported environment, including geography and eligibility. Require a clear account of costs, dependencies, data handling and failure recovery.

Choose a bounded pilot only when there is a credible customer use case. Define success and the conditions for stopping. Preserve a conventional route so customers can still reach the business when the intermediary fails.

This is preparation for a possible channel, not a recommendation that every small firm launch an agent now.

What could slow the future down?

Customer trust, platform competition, inconsistent product data, commercial agreements and unresolved responsibility can all affect adoption. Their relative importance will vary.

A standard may make integration easier without making the customer willing to delegate. A useful comparison assistant may succeed while automated payment remains limited. A merchant may have technically compatible systems but an unattractive offer.

These possibilities are why an adoption deadline would be speculative. We do not assign one here.

Watch actual supported launches and customer behaviour in your market. Update the strategy when evidence changes, rather than treating a 2026 announcement as a permanent map of the future.

Prepare the truth before the technology

The next shopping interface may ask merchants for more structured and current information. It may also expose inconsistencies that a human sales team previously repaired in conversation.

A business can make useful progress now by documenting what it sells, who it suits and what happens when the purchase goes wrong. Those improvements help today's customers while preserving options for tomorrow.

The Umbrella connects strategy and digital experience. If you are being asked to invest in agentic commerce, bring the proposal and the customer task it claims to improve. The first decision is whether the business has a real problem this technology is ready to solve.

Sources & Further Reading

• Google: UCP and retail tools, 11 January 2026. Launch announcement; initial US checkout scope and planned expansion are distinguished.
• Google Developers: Under the hood of UCP. Technical background.
• Google: AP2 donation to FIDO Alliance, April 2026. Payments-protocol governance update.
• Google Search Central: AI features. Search eligibility guidance, distinct from commerce-integration requirements.

- NIQ: Agentic Commerce India 2026, September 2026. Public report summary; eight-metro scope, limited publicly available methodology, no paid report accessed.