The New Customer Journey: From Assistant to Agent
An AI customer is defined less by a product label than by where decision authority moves. This chapter follows that movement across discovery, evaluation, purchase, and post-purchase service, then examines the preference model an agent must carry between those stages. The practical value of the distinction is simple: the more authority a person delegates, the more important it becomes to know what the system is representing, what it may do, and how either party can correct it.
The shift is gradual. A person might ask an assistant to collect options and remain involved in every comparison. The same person might later authorize an agent to search several retailers, choose within a budget, and place an order. During a return, they might delegate the unpleasant conversation back to the system. “Assistant” and “agent” describe points on this continuum, not permanently separate species of software. The operative question is what the system is allowed to decide and execute at each moment.
Autonomy is a change in authority, not a product label
At the discovery stage, an AI system can already influence demand without buying anything. It chooses which sources to inspect, which product pages to retrieve, and which alternatives to show first. That selection creates a consideration set: the small collection of options a person will seriously evaluate. If an agent omits a retailer, discounts a product because its attributes are hard to parse, or gives prominence to an easily retrieved brand, the human's eventual choice is shaped before any transaction occurs.
An assistant can do this while remaining collaborative. It may ask follow-up questions, explain why an option appears relevant, and wait for the person to decide which direction to pursue. The human supplies frequent corrections, so an imperfect interpretation can be repaired quickly. An agent, by contrast, is expected to pursue a user-defined goal with less intervention. It may decompose “find a suitable laptop battery” into compatibility checks, price comparisons, warranty searches, and seller verification. The person need not specify every intermediate action, but the authorization remains bounded by the goal and its constraints.
That boundary matters because autonomous execution changes the cost of an error. A weak recommendation wastes attention. A wrong purchase can spend money, expose personal information, create a return obligation, or lock the person into a poor service interaction. The relevant control is therefore not only a final confirmation screen. It includes the ability to set limits in advance, inspect the basis for a choice, override an action, and recover when the agent encounters a condition outside its authority.
The customer journey makes these differences visible:
- Discovery is about finding information and constructing possibilities. The system may search brand sites, retailers, reviews, communities, or other sources, then narrow them into a shortlist.
- Evaluation is about comparing those possibilities against the person's needs. The agent may summarize specifications, prices, ratings, shipping terms, and apparent trade-offs. It is now making judgments about what counts as relevant evidence.
- Purchase is an action stage. The agent may add an item to a cart, select a seller, or complete a transaction if the person has granted the necessary authority. Limits such as budget, delivery date, or acceptable retailer become operational constraints rather than preferences mentioned in conversation.
- Post-purchase service extends the role beyond the checkout. An agent may request a refund, negotiate a remedy, submit feedback, or track a delivery. It is acting in a relationship that can contain frustration, ambiguity, and competing interpretations of what the customer is owed.
These stages do not carry the same risk, and they need not have the same autonomy setting. A person may delegate broad exploration but require approval for payment. They may allow an agent to return a routine item but want to handle a health-related complaint themselves. The marketing research on AI customers emphasizes that task complexity, perceived control, and the way delegation is initiated affect whether people accept autonomous choice. A system that silently expands its authority can therefore undermine trust even when its recommendations are often useful.
The firm and the platform participate in every stage. A firm decides how clearly it describes an offer, what policies it exposes, and whether its service channels can be used by software. A platform or web environment determines what an agent can retrieve, how results are ranked, what interaction history is visible, and what information survives for later decisions. The agent is not simply carrying a human message from one side to the other. It is operating inside an environment that can favor some sources, make some actions easy, and make other evidence effectively invisible.
Each transition in the journey changes what must persist. Discovery produces a set of candidate options, but evaluation must preserve the reasons those options were considered. Evaluation produces a recommendation, but purchase requires the agent to carry forward the constraints that make the recommendation permissible. After purchase, the outcome becomes feedback only if the system records enough context to interpret it: a return might signal a defective product, a wrong size, a late delivery, or a change of mind. Treating every outcome as a simple approval or rejection throws away the distinction needed to improve later decisions.
The transitions also have different failure and authorization boundaries. A search agent can usually recover from an incomplete result by consulting another source. A transaction agent may not be able to recover from an incorrect quantity, seller, or payment choice without incurring cost. A service agent may have permission to ask for a replacement but not to accept a settlement or disclose sensitive information. Good delegation therefore specifies not only the desired outcome but the permitted transitions: what the agent may infer, which actions require confirmation, what evidence must be shown, and when the process must stop and return to the human.
The interaction can be understood as a sequence of authority transfers and feedback loops:
The diagram also shows why “human in the loop” is not a sufficient description. A human can remain nominally present while the agent has already chosen the sources, framed the alternatives, and defined what evidence reaches them. Conversely, a person can delegate an action while retaining meaningful control if the boundary is explicit and the system exposes its reasons, current facts, and reversal path. Autonomy is best understood as a property of the relationship and the transition rules, not as a binary attribute of the model.
The agent carries a partial model of the person
Delegation works only if the agent can form a usable representation of the person it serves. That representation may include an explicit request, remembered preferences, past choices, context, budgets, and constraints that were never stated in one sentence. It must also distinguish hard requirements from soft preferences. Exact compatibility for a replacement part may be non-negotiable; a preferred color may be negotiable if the price or durability is compelling. If the agent treats every signal as equally rigid, it will either reject useful options or recommend products that violate the actual purpose of the purchase.
AgenticShop illustrates how this hidden model can be made inspectable. Its benchmark derives user profiles from purchase histories and review text, then organizes the result into a narrative context, a query suited to the shopping situation, and a checklist. The checklist separates brand preference, price sensitivity, review sensitivity, functional requirements, aesthetic preferences, and purchase preferences. That separation is valuable because shopping decisions often combine unlike kinds of evidence. A buyer of a laptop battery may care about exact part-number compatibility and safety protections, while a buyer of trail shoes may weigh fit, traction, cushioning, and appearance. A single relevance score cannot show which of these conditions was met or ignored.
But a profile is still a model, not the person. It is assembled from behavior and language, both of which can be incomplete or misleading. A purchase may have been a gift. A positive review may reflect one unusually good experience. A repeated purchase may indicate convenience rather than strong brand preference. The benchmark uses checklists to make these assumptions testable, but its own evidence treats the checklist as an operationalization validated against ratings and human judgments—not as a complete account of utility. The distinction is essential: making a preference explicit improves evaluation without making the inference infallible.
The difficulty also depends on what kind of product is being evaluated. Search goods have properties that can be compared relatively directly before purchase. A contract with a stated price and coverage, for example, can be screened against explicit criteria. Experience goods are harder because quality emerges through use: comfort, taste, reliability, or the feel of a garment may be only imperfectly described in advance. Credence goods are harder still when the buyer cannot easily verify quality even after using the service and must rely on external signals. Reviews, certifications, warranties, and explanations can help, but they are proxies rather than direct access to the underlying experience or quality.
The evidence needed at each transition follows from this difference. For a search good, a structured specification may settle much of the comparison. For an experience good, the agent may need to represent whose experience is relevant, how comparable the review context is, and whether a general rating hides a recurring fit problem. For a credence claim, the agent may need to preserve the source and qualification rather than repeat the seller's assertion as fact. This is one reason a polished summary can be less useful than a shorter answer that shows which claims are directly documented and which remain uncertain.
A working preference model must also preserve conflicts instead of smoothing them away. A person can prefer a familiar brand but accept an unfamiliar one for a large price difference; want fast shipping but tolerate delay for a scarce item; or prioritize appearance until a safety or compatibility requirement is threatened. These are conditional rules, not a flat list of adjectives. They may be inferred from behavior, but the inference should remain revisable because circumstances change. A model that records only the final choice loses the trade-off that produced it and may repeat the wrong lesson in a different situation.
This is why an agent must represent more than keywords. It has to infer which facts matter, which sources are credible for that fact, and how much uncertainty the person will tolerate. It may need to preserve a distinction between “must have,” “prefer,” and “acceptable compromise.” It may also need to ask a question rather than guess. A system that fills every ambiguity with a confident default can appear efficient while quietly rewriting the customer's objective.
Repeated interaction makes the relationship dynamic. The agent can update its recommendations after the person accepts or rejects an option, comments on a purchase, or reports that a product failed in practice. The person, in turn, sees a stream of options selected by the agent and may adjust their own expectations or preferences around what is presented. Over time, the two sides can influence one another. The marketing research describes this as possible co-evolution of human and AI preferences: the system learns from feedback while the human adapts to the system's outputs.
That possibility brings both value and risk. Learning can reduce repetitive explanation and make routine decisions easier. It can also narrow the consideration set around an early mistake, amplify a proxy that was convenient to measure, or make the person dependent on recommendations whose origin they no longer inspect. Trust should therefore not be reduced to whether the agent sounds personable. It depends on whether the system is predictably aligned with the person's goals, communicates uncertainty, respects the authorized boundary, and offers a meaningful way to correct the model.
The commercial side of this relationship is changing as well. If an AI customer chooses what to read and what to recommend, firms must make their products understandable to a machine as well as attractive to a person. The marketing literature calls the broader effort machine marketing and links it to practices such as generative engine optimization: structuring and distributing information so that AI systems can retrieve and incorporate it. That does not mean that a machine-facing description should replace human communication. It means that firms increasingly have two audiences in one exchange—the human whose needs matter and the intermediary that interprets the offer.
The next question is how to test whether this partial representation actually produces a fitting choice. Once preferences are made explicit enough to inspect, a recommendation can be evaluated against concrete criteria rather than judged by its fluency. That shift—from trusting the agent's apparent understanding to checking the evidence behind its choice—is where the customer journey becomes a measurable curation problem.