Don't Let an AI Assistant Advise and Sell in the Same Voice

The next important design problem in AI is not whether an assistant can complete a transaction. It is whether the customer can tell when the assistant has stopped helping and started selling.

That distinction is becoming less obvious. Shopping assistants can recommend an item and take the customer directly to checkout. Ads are appearing next to generated output. AI systems are entering interviews, group conversations and other settings where people may treat their responses as neutral guidance. At the same time, watermarking requirements and AI writing policies show how much concern there already is about provenance: who produced something, why it is there and whose interests it serves.

Product teams should treat that as one connected problem. An assistant cannot advise, advertise and close a sale in the same voice without creating an incentive conflict. The interface may look consistent, but the product has changed roles.

In a conventional ecommerce journey, those roles are easier to recognise. Editorial content helps someone understand a category. Search and filters help narrow the options. Sponsored placements are labelled. The checkout asks for a commitment. There are still plenty of questionable patterns, but the customer can usually see that different parts of the journey have different purposes.

A conversational interface compresses all of that into a single answer. That is convenient, but removing steps also removes signals. If the assistant recommends one option, gives a persuasive reason and immediately offers one-click checkout, the customer has little basis for judging whether it selected the best answer, the most profitable answer or the answer an advertiser paid to place.

I would not solve this with a disclaimer buried below the conversation. I would make the product change modes visibly. Exploration, recommendation and purchase are different customer jobs. Moving between them should be clear in the interface, recorded in the product data and reflected in the success criteria.

For example, an assistant in exploration mode might compare options and explain trade-offs. In recommendation mode, it should state which criteria drove the ranking and make alternatives available. If sponsored inventory is introduced, it should not inherit the visual treatment or implied neutrality of the assistant’s own recommendation. Before purchase, the product should make the transition from advice to transaction explicit rather than presenting checkout as the natural completion of a supposedly neutral conversation.

This also changes measurement. Product and growth teams are used to removing friction and watching conversion improve. But a shorter path is not automatically a better decision. If we only track recommendation acceptance and completed purchases, commercial pressure will gradually shape the assistant’s behaviour. The dashboard can look healthy while the product becomes less useful.

I would therefore separate decision-quality measures from transaction measures. Are customers inspecting alternatives? Are they overriding recommendations? Do they return to the assistant for another decision? What reasons do they give for rejecting a suggestion? These signals do not replace revenue metrics, but they reveal whether the assistant is earning influence or merely exploiting it.

This is also a product management responsibility, not something to hand to legal after the experience has been designed. Product leaders define the purpose, roadmap, roles and metrics of the system. If those artifacts do not explain when the product serves the customer and when it serves a commercial objective, the team has left its most important strategic conflict unresolved.

The current product conversation puts a lot of attention on AI maturity, model choice, faster experimentation and cleaner metrics. Those matter. But maturity also means refusing to hide incompatible goals inside a smooth interface. A good assistant can help someone decide, and a good commerce product can help someone buy. Combining them can work, but only when the customer can see exactly where one ends and the other begins.