AI assistants are moving from answering product questions to narrowing choices and, in selected journeys, initiating transactions. Brands must now influence human preference while providing the structured truth, evidence, services and controls that intelligent intermediaries require.
A customer needs a new washing machine. Instead of opening several retailer websites, they tell an AI assistant: “Find a quiet model that fits this space, costs less than £650, can arrive before Saturday and includes removal of the old appliance.”
The assistant converts the request into constraints. It compares dimensions, energy use, reviews, delivery slots and removal policies. It explains the compromises between three suitable models and prepares the preferred option for purchase. The customer checks the recommendation and confirms.
This scenario combines capabilities that are already available with integrations that remain selective by platform, retailer and market. Its significance is not that every purchase can now be delegated. It is that the customer no longer needs to construct every stage of the journey themselves.
The agent can assemble much of it on their behalf.
The disappearing journey
The customer journey is becoming a decision service
Digital marketing was built around a familiar sequence. A customer recognised a need, searched for options, encountered a brand, visited its website and moved through a journey designed by that brand.
AI-mediated discovery compresses that sequence. The customer can begin with a problem rather than a product name. The assistant can interpret context, run several searches, compare attributes and synthesise reviews before the customer encounters a brand-controlled page.
McKinsey's December 2025 survey of 749 consumers in France, Germany and the UK found that 38% already used AI to research products or decide what to buy. Among shopping-related uses, comparing brands, models, prices and reviews was the most common. Consumers were more willing to let AI support judgement than to give it unrestricted authority to act.
The traditional funnel has not disappeared. It is being partially absorbed into a decision service operated by an intermediary.
That changes the basis of competition. Brands still need to be remembered, desired and trusted by people. They must now also enter a shortlist that may be assembled before the customer reaches an owned channel.
The question is no longer only whether a message attracts attention. It is whether the brand can be retrieved, understood, compared, justified and acted upon for a particular customer intent.
The emerging model
B2A names one part of a wider market shift
The language around this development remains unsettled.
Business-to-agent marketing, or B2A, describes the work required to communicate with AI systems that influence or act for customers. Forrester applies the term to information architecture, content governance and consistency across owned and third-party sources.
Agentic commerce describes the broader environment through which agents discover products, assemble baskets, connect identities, initiate payments and support orders.
Machine customer is a wider category for non-human economic actors, potentially including personal assistants, connected equipment and automated procurement systems.
These terms are related, but they are not interchangeable or universally accepted. AI-mediated customer journey is the more neutral description of the change now taking place.
This is not a successor to B2C or B2B. The customer still creates the demand, defines their preferences and decides which intermediary to trust. Brand meaning, emotion, creativity and loyalty continue to influence those choices.
The new requirement is operational: the brand must be understandable and actionable by the system helping the customer.
The distinction
Search is the foundation, not the whole discipline
It is tempting to treat this as another optimisation contest. That would reduce a structural change to a set of publishing tactics.
Google states that established search engine optimisation principles remain relevant to its generative search features. It recommends useful original content, crawlable websites and sound technical structure, while rejecting the need for special AI files, artificial content fragments or pages rewritten solely for machines. It also cautions that structured data does not guarantee inclusion.
Those are important boundaries. A brand does not become agent-ready by publishing hundreds of formulaic answers or adding a new acronym to its SEO plan.
Search helps the agent find information. Agent-mediated commerce requires the brand to do more with that discovery. It may need to provide:
- Dependable product and service facts
- Current prices and availability
- Delivery, support and returns terms
- Evidence supporting important claims
- Interfaces for baskets, booking or checkout
- Customer identity and loyalty connections
- Permissions governing what the agent may do
B2A is therefore less a replacement for SEO than an extension from being found to being usable.
The capability boundary
Influence is current; execution is uneven
The transition is taking place in stages.
Available today, AI systems can interpret needs, research categories, synthesise information, compare products and recommend options. ChatGPT supports conversational refinement and side-by-side product comparison. Selected US merchants can transact through Microsoft Copilot, while other agent-mediated journeys transfer the customer into a merchant-controlled checkout.
Emerging now, platforms are connecting product catalogues, real-time prices, inventory, baskets, customer accounts and loyalty programmes. Google's Universal Commerce Protocol includes optional catalogue, cart and identity-linking capabilities, while Universal Cart is being introduced across Search and Gemini before wider geographic and channel expansion.
Reasonable near-term applications include repeat purchases, approved substitutions, appointment booking, basket building and transactions within a customer-defined budget.
More speculative scenarios include widespread autonomous purchasing of expensive, emotionally significant or highly regulated products without human confirmation. Adoption will differ by category, consequence and reversibility.
The likely future is not an abrupt transfer of all purchasing authority. It is a variable spectrum of delegation.
What agents select for
Agents select under uncertainty
A customer can compensate for poor information. They may recognise that two slightly different product names refer to the same item, call a branch to confirm availability or decide that one of two conflicting delivery pages is outdated.
An agent must resolve that uncertainty from the information and interfaces available to it.
The practical signals will vary by platform, task and customer context. They may include relevance to the request, explicit product attributes, current price, stock, location eligibility, service terms, credible reviews and the merchant's ability to fulfil the transaction.
OpenAI's merchant feed specification illustrates the breadth of information involved. It includes stable identifiers, titles, descriptions, variants, pricing, availability, shipping, returns, seller details, warnings, reviews and geographic conditions.
This does not reveal a universal recommendation algorithm. Platform logic remains only partly visible and is likely to change.
The ACES academic research environment reinforces that limitation. In controlled simulated marketplaces, different shopping agents made different choices from the same product sets and showed varying sensitivity to position, ratings, reviews, endorsements and price. The research does not disclose how a live commerce platform ranks products, but it demonstrates why no single optimisation formula should be trusted.
Brands should not attempt to manipulate an imagined machine preference. They should reduce avoidable uncertainty.
The new brand infrastructure
An agent-ready brand makes four operational contracts
The information an agent consumes is only one part of the system. A brand becomes genuinely agent-ready when it can fulfil four connected operating promises.
The meaning contract: this is what we offer
The first contract establishes a consistent account of the brand, product or service.
It requires canonical product names, stable identifiers, explicit attributes, service definitions and policy terms. Those facts should remain coherent across websites, retailer listings, marketplaces, support content and product feeds.
A product information management system may govern the attributes. A digital asset management system may control approved imagery and usage rights. Structured content can separate reusable facts from the different creative expressions used across campaigns and channels.
The customer consequence is simple: fewer incorrect comparisons and fewer promises built from outdated information.
If the meaning contract fails, the agent cannot confidently identify what is being offered.
The evidence contract: this is why the claim is credible
The second contract provides the basis for belief.
A brand's own description matters, but it is not the only available account. Reviews, product tests, expert references, regulatory records, independent reporting and customer commentary may all shape the evidence available to an AI system.
This does not mean every negative review becomes a deterministic penalty. It means contradictions and unsupported claims create uncertainty that the system may need to disclose or resolve.
Marketing, public relations, product governance and customer service consequently become more tightly connected. A campaign claim that cannot be traced to approved evidence is no longer only a compliance risk. It may also become a discoverability and representation problem.
If the evidence contract fails, the agent can describe the product but cannot confidently justify recommending it.
The service contract: this is what we can reliably do
The third contract turns information into an executable offer.
It includes the price that applies now, inventory, delivery locations, appointment availability, installation, accepted payments, fulfilment, returns and post-purchase support. These capabilities may be supplied through product feeds, merchant platforms or application programming interfaces rather than inferred from a web page.
This is where customer experience becomes inseparable from operational systems. An assistant may produce a compelling recommendation, but the experience fails if the product is unavailable, the installation promise is wrong or the return conditions were misrepresented.
If the service contract fails, the agent can recommend the offer but cannot dependably complete it.
The authority contract: this is what the agent may do
The fourth contract defines permission and accountability.
It covers the information the agent may access, the identity it is acting for, the payment methods it may use, its spending limits, when it must request confirmation and how actions are logged or reversed. It also governs representation: which prices, substitutions, claims and offers the intermediary is authorised to present.
The UK Information Commissioner's Office has identified purchasing, sale monitoring, finance sourcing and price negotiation as plausible agentic commerce activities, while warning that development cannot come at the expense of privacy.
If the authority contract fails, convenience becomes an uncontrolled commercial and data risk.
The four contracts are interdependent. Meaning without evidence creates unsupported claims. Evidence without serviceability creates recommendations that cannot be fulfilled. Serviceability without authority creates transactions nobody can adequately govern.
The commercial tension
Agent accessibility does not require commercial surrender
Brands will need to decide how deeply to participate in agent-mediated channels.
That decision is often framed as a binary choice between accessibility and customer ownership. Current platform models show more variation.
Google and Microsoft state that retailers can remain the seller or merchant of record in selected agent-mediated checkouts. OpenAI has placed renewed emphasis on merchant-controlled checkout after finding that its initial Instant Checkout implementation did not offer the flexibility it wanted to provide.
Merchant-of-record status is important, but it is not equivalent to owning the journey. The intermediary may still determine which products are compared, how trade-offs are explained and when the brand enters the conversation.
A useful participation ladder has five levels:
- Discovery access: The agent can retrieve public content and product facts
- Qualified referral: It can recommend an offer and transfer the customer to an owned experience
- Connected experience: The customer can link their account, loyalty status or preferences
- Embedded transaction: The purchase takes place inside an intermediary while the brand fulfils it
- Delegated service: The agent can reorder, book, amend or resolve issues within defined permissions
Brands do not have to use the same level everywhere.
A low-margin replenishment product may benefit from delegated ordering. A luxury purchase may depend on controlled storytelling and human consultation. Financial services, healthcare and regulated products require stronger identity, consent and approval boundaries.
The decision should account for customer value, margin, service complexity, reversibility, regulatory exposure and the importance of the owned relationship.
The objective is not maximum agent access. It is the right access for the category and customer.
An illustrative journey
A blended journey preserves human judgement
Consider an illustrative customer looking for an electric cargo bike for a school run.
They ask an assistant to identify models suitable for two children, a steep route, limited storage and a total budget that includes servicing. The assistant translates those needs into criteria, then compares dimensions, load capacity, motor performance, warranties, independent reviews and access to approved maintenance.
It excludes models that will not fit the available storage space. It explains the difference between two remaining options and identifies where its evidence is incomplete.
The customer does not delegate the final decision. The assistant books test rides at a nearby retailer. A specialist discusses handling, safety and finance, while the customer tests the products.
Once the customer chooses, the assistant retrieves an approved offer and prepares the transaction using the customer's linked account. The customer checks the finance terms and confirms.
The retailer fulfils the order, registers the warranty and retains the service relationship. The assistant can later remind the customer about maintenance or help arrange a booking within the permissions they have granted.
The agent reduces research and coordination. The brand provides product confidence, physical experience, fulfilment and continuing care. The person remains responsible for the consequential decision.
That combination is likely to be more common than either a completely manual journey or an entirely autonomous one.
What brands should do now
Preparation should follow the dependencies
Brands do not need to rebuild the customer journey around an uncertain future. They do need to address foundations that already affect discoverability, customer experience and operational quality.
Immediate actions: establish the current truth
Audit how the brand appears across high-intent prompts and leading AI interfaces. Check whether answers are accurate, which sources are used, which competitors enter the shortlist and where the system expresses uncertainty.
Reconcile product, policy and service information across owned websites, feeds, marketplaces and support channels. Focus first on commercially important products and customer questions.
Review technical discoverability, product markup and merchant feeds, but reject tactics that create low-value content merely to influence an answer.
Set a current capability baseline. Record which agent journeys are live in each market rather than building strategy around global product announcements.
Capability-building actions: connect information to execution
Appoint an accountable owner for product and service truth. Marketing should not have to correct facts independently across every channel.
Define how product information management, digital asset management, content, ecommerce, inventory, customer service and order systems maintain shared entities and update rules.
Assess interface readiness for the capabilities an agent may need: product retrieval, real-time price and availability, delivery, booking, baskets, checkout, order status, returns and support.
Build the authority model before increasing autonomy. Identity, authentication, consent, payment permissions, transaction limits, audit records, escalation and dispute resolution should be designed into the workflow.
Develop measurement that follows the new journey. This may include retrieval accuracy, recommendation presence, agent referrals, assisted conversion, account linking, margin, return rate and post-purchase retention. A mention without a commercially useful outcome is not success.
Longer-term actions: decide where control should sit
Segment agent participation by product category, customer need and risk. Do not expose every service merely because a protocol permits it.
Pilot bounded tasks with reliable data and reversible consequences. Compare the customer outcome with the established journey rather than measuring novelty or usage alone.
Design continuity into the relationship. Account linking, membership, warranties, service history and post-purchase support can preserve customer value even when an external agent introduced the product.
Treat platform dependency as a channel strategy issue. Maintain portable product data, clear interfaces and the ability to change participation levels as economics, customer expectations and platform terms evolve.
The takeaway
Marketing must supply a decision environment
The next phase of customer experience will not be wholly directed by people, designed by brands or executed by autonomous systems.
People will express goals, preferences and boundaries. Intelligent intermediaries will interpret those needs, assemble evidence and coordinate parts of the journey. Brands will compete to provide the meaning, proof, service and authority that make a recommendation safe to act upon.
Marketing still has to create preference. Ecommerce still has to convert demand. Technology still has to integrate systems. Governance still has to protect the customer and the organisation.
The difference is that these responsibilities can no longer operate as separate layers.
When an AI agent mediates the journey, the brand is not only publishing a message. It is supplying a decision environment.
The brands that matter will be the ones people desire, agents can justify and both are permitted to choose.
