AI product strategy · Essay
Shopping Assistants: Inheriting an Era of Higher Expectations
Over eleven years building commerce products I keep returning to the same problem, and the answer keeps moving. It is worth being clear about what has already happened before claiming something new is.
Customisation gave customers choices. Pick the colour or the configuration. The system created variants and the customer did the selecting. Personalisation inverted that relationship: the system started doing part of the job by remembering preferences, ranking products differently, adapting what it showed. Hyper-personalisation added context, on the understanding that the same customer is not solving the same problem every time she shops. Season, occasion, price sensitivity and immediate intent all change what "relevant" means.
Every one of those stages improved the same underlying capability. The system got better at deciding what to put in front of a customer.
AI assistants change something else. They change who the customer thinks they are talking to.
The assistant is no longer judged as a chatbot
A few years ago a retailer's chatbot was benchmarked against other retail chatbots, and the bar was low. Recognise a few intents, answer common questions, push the customer towards a product page, job done.
That reference point is disappearing. Someone opening a shopping assistant today compares it against the most capable AI they have used anywhere: the one that summarised a contract at work, the one that untangled a vague question and still worked out what they meant. Behaviours learned at work migrate into personal life, and AI is making that trip unusually fast.
So the assistant stops being read as the retailer's chatbot and inherits expectations set by the whole AI ecosystem. Customers expect it to resolve ambiguity and weigh trade-offs. Some expect it to push back. If I say I need something smart enough for a wedding, but not so formal I can't wear it again, I am not asking for "wedding" to become a filter. I am asking the assistant to see the tension in the sentence.
Fluent language also produces something awkward: people infer judgment from it. A mediocre recommendation carousel is a mediocre carousel. A conversational assistant that confidently recommends the wrong product reads as bad judgment by something that should have known better.
Competence is extended on credit. Trust is not.
Customers now arrive believing the technology should be able to help them before the retailer has shown that it can. Call it competence on credit. They assume the assistant can parse an incomplete sentence, recall something from earlier, reason rather than retrieve.
Customers lend the assistant competence. Trust still has to be earned, and it can be lost two different ways. The assistant can fail because it does not understand me. Or it can fail because I stop believing it is working in my interest.
Which is why I think a shopping assistant has three jobs: advocate for me, understand me, act for me.
1. Advocate for me
Of the three this is the least technical and the hardest to solve, which is why it takes up the most room here. Assistants live inside commercial systems. The customer wants a good outcome, the retailer wants a transaction, brands want visibility, advertising creates its own pull. None of those incentives is illegitimate on its own. The problem starts when one is presented as another.
Customers understand the visual grammar of ecommerce. A result labelled Sponsored is obviously an advertisement, and nobody feels deceived by it. But when an assistant writes I think this is your best option, the sentence carries the weight of judgment. If a commercial incentive influenced that judgment invisibly, the system has borrowed the language of advocacy without supplying any.
The principle I would hold to: commercial influence can exist inside an assistant, but it should never be disguised as independent judgment.
A real advocate has to be willing to say none of these is a good match, or the cheaper one is enough for what you described, or I don't think you need to buy anything yet.
That raises an uncomfortable question for anyone building one inside a retailer. Can a retailer's assistant fully advocate? It is bounded by its own assortment and its own economics. A general-purpose assistant can compare across merchants, which makes it better placed to answer what should I buy rather than what should I buy here.
Retailers are not without a case. They know their assortment in depth, they know live availability, they can connect advice to fulfilment, and they usually know a customer's purchase and return history better than an outsider does. Their advantage is context and execution rather than neutrality. The honest move is to name the boundary. There is a real difference between this is the best product for you and of what we carry, this is the one I would pick. The second claim is narrower and more believable.
2. Understand me
Advocacy is worth nothing if the assistant has misread the problem.
Traditional personalisation is good at remembering inputs. You buy this brand, you wear this size, you clicked these things, people like you bought that. AI raises the bar from remembering to understanding, and history becomes misleading the moment it is treated as a permanent description of a person. That I usually buy cheap does not make me price-sensitive when I am buying an anniversary gift. A return tells you nothing about why. Someone who normally shops for herself might be shopping for her son today.
So the assistant has to build and keep updating a model of the current problem. In plain terms: have I understood what you are trying to do?
Sometimes that means asking. One of the easiest mistakes in conversational design is assuming fewer questions is better experience. The target is not zero clarification. It is asking only where the answer changes the decision.
I ran into this building a skincare advisor that answers a deceptively simple question on a product page: is this suitable for me? The easy version always produces a recommendation. That version is wrong. Suitability depends on skin type, on what the customer is trying to fix, on ingredients and on product data that is often incomplete. The correct experience sometimes has to say no, explain why, and only then offer alternatives.
That sounds like a small interaction choice. It changes the objective of the product. A recommendation engine is rewarded for surfacing relevant items. An advisor should be rewarded for improving a decision. The same split applies to a laptop, a mattress or a bike. As AI moves from retrieval towards judgment, the difference stops being academic.
3. Act for me
The last job is turning understanding into action. A recommendation system says here are five running shoes you might like. An assistant says given the cushioning you preferred last time and the distance you run, I would take the second pair, and your size is in stock. Then: shall I add it?
This is where assistants separate from recommendation systems. One optimises what you see. The other participates in what you do. Extended out, that covers assembling an outfit to a budget, rebuilding a routine, replenishing what runs out, coordinating a purchase with several constraints attached.
Action brings its own tension. Removing friction is treated in commerce as an unqualified good, and I don't think it is. For a repeat purchase, immediate execution helps. For an expensive or sensitive decision, a moment of confirmation protects the customer. What you want is effortless execution with deliberate confirmation where it counts. The assistant needs to know when to stop and ask.
Trust needs a dashboard
If trust sits at the centre of the product it cannot stay a principle. Commerce teams are already sophisticated at measuring conversion and will have to get equally sophisticated at measuring whether the assistant helped.
One measure per job is enough to start. For understanding, how often customers have to correct the assistant. For advocacy, returns on the products it recommended most strongly. For action, reversals and cancellations after an assistant-led step.
One measure deserves separating out. Compare how sponsored recommendations perform against organic ones after the transaction, on returns and on repeat use. If the paid ones do worse, the assistant is spending trust to move inventory, and the number will say so long before customers do. It is an uncomfortable number to keep in front of you, which is the argument for putting it there.
Would we still call this interaction a success if the customer chose not to buy?
There is a reason conversion cannot be the final measure. A persuasive assistant can lift conversion while making customer outcomes worse. That is the test I would give a team. Sometimes the right answer from an assistant is wait, or buy the cheaper one, or you already own something that does this. Those conversations cost a transaction today and are the reason the customer comes back tomorrow.
Every category has a code
Good advice does not look the same everywhere. Skincare turns on suitability, safety and honest uncertainty. Fashion runs on taste, identity and occasion. In luxury, stripping out friction can damage the experience, because the browsing is part of what is being bought. In electronics, specifications and explicit trade-offs dominate.
An assistant therefore has to learn two things: the customer, and the code of the category. One that behaves identically while recommending a serum, a laptop and a handbag is general-purpose in the technical sense. It will not feel intelligent.
Anticipation sits on top of all three
Memory is not only about sparing the customer from repeating herself. It eventually lets the assistant move first. A swimwear purchase makes sun protection relevant. A foundation bought every few months is approaching replenishment. A saved jacket becomes interesting when the forecast changes.
Technically that is a prediction problem. As a product problem it is about permission. Have we earned the right to interrupt? The same message reads as thoughtful or invasive depending on timing, sensitivity and whether the customer can see why it appeared.
The distance between "that's useful" and "why does this company know that" is very short.
So anticipation needs visible customer control, sensible frequency, a confidence threshold and a plain explanation attached. Prediction is getting easier. Permission is not.
Where I think this goes
Competence on credit has a strategic consequence. It may decide who owns the first conversation.
A prediction, so I can be wrong in public. Within a few years the opening move in most considered purchases will happen outside the retailer, inside a general-purpose assistant, because that is where what should I buy gets a credible answer. Retailers will not lose the transaction. They will lose the first conversation.
If that is right, the retailer's assistant is a specialist rather than a rival. It knows the assortment, the stock and the customer's own history, and it owns execution and everything after the sale. The build implication is unglamorous: retailers should be investing as much in being queried well by someone else's assistant as in being visited by their own customers. Most are doing the opposite.
Spend the credit carefully
For most of the history of personalisation the question was what should we show this customer. Assistants replace it with what should we help this customer decide, and eventually what are we trusted to do on this customer's behalf. That is a more valuable relationship and a more demanding one.
Customers arrive expecting intelligence the product has not yet earned. That is the gift. Every failure is now measured against the frontier. That is the debt. Every confident misunderstanding draws the credit down. Every hidden commercial incentive draws it down faster. Every useful piece of memory tops it up, as does every admission of uncertainty instead of manufactured confidence, as does every recommendation against a purchase that was not worth making.
The assistant that wins will not be the one with the biggest model or the richest customer profile. It will be the one that keeps proving three things: I understand what you are trying to achieve, you know whose interest I am representing, and when you are ready, I can help you act.
Only then has it earned the right to anticipate what comes next.