Where business logic ends and AI begins in commercial selling

Retail has a habit of putting new labels on old machinery. A banner that says “AI recommended” may still be driven by a simple rule: the shopper bought something red, so show more red; the shopper looked at stripes, so show more stripes. That can be useful. It is not yet a commercial selling assistant.

The distinction matters because conversational and agentic shopping interfaces are moving from experiment to product. Amazon has positioned Rufus as a generative AI shopping assistant, and OpenAI's Instant Checkout work shows how product discovery, decision support and transaction can start to live inside conversational surfaces rather than only inside a retailer's own catalogue page.[1][2]

AI should not replace retail business logic. It should make that logic usable in messy commercial conversations.

Editorial conclusion

Business logic is the hard floor

A selling system cannot improvise its way around the basics. Price, tax, margin, inventory, delivery promise, promotion rules, customer agreements, return policy and legal restrictions are not creative suggestions. They are constraints. If a product is not available in the required size, cannot be delivered in time, breaches a channel rule or destroys the commercial target, the assistant needs to know that before it sounds persuasive.

That is where business logic belongs: in deterministic systems that can be tested, audited and owned. A promotion engine should decide whether a basket qualifies. An inventory service should say what can be fulfilled. A pricing service should calculate the valid price. AI can ask those systems better questions, but it should not silently invent their answers.[4]

Recommendations are not the same as selling

Classic recommendation systems often optimise for similarity, popularity, previous behaviour or likely conversion. That is valuable, especially at scale. Retail recommendation platforms also let teams apply controls such as filtering, boosting and business rules around what should be shown. But those mechanisms still do not automatically understand the selling job in front of them.[3]

A customer who bought a red jacket may not want more red products. They may be completing an outfit, replacing a damaged item, shopping for a gift, working within a budget, trying to avoid returns, preparing for an event or asking whether the cheaper option is good enough. A commercial seller has to reason about that goal, not merely replay a behavioural echo.

AI belongs in the uncertain middle

The useful space for AI is the uncertain middle between a customer's messy intent and the retailer's hard operating rules. A shopper says they need “something smart enough for a client dinner but comfortable for travel”. A buyer asks for alternatives that protect margin because a supplier is late. A store associate needs to recover a sale when the preferred size is missing. These are not only search problems. They are interpretation problems.

In a commercial selling app, AI should translate natural language and incomplete signals into structured intent, then call the right retail systems: catalogue, stock, pricing, promotions, customer history, product attributes and fulfilment options. The answer should be grounded in those systems. The AI layer earns its keep by framing options, explaining trade-offs and asking for missing context.[4]

• Intent: what is the customer or seller actually trying to achieve?
• Constraints: what is allowed, available, profitable and compliant?
• Options: which products, bundles or next actions fit the situation?
• Trade-offs: what changes when price, speed, margin, fit or availability matters more?
• Explanation: why is this recommendation commercially sensible now?

The interface should show its reasoning

Good commercial AI should feel less like a magic banner and more like a capable colleague with access to the right systems. It might say: “This alternative protects margin, is available in two nearby stores and matches the customer's stated delivery window. The cheaper option is lower margin and has a higher historical return rate in this size range.” That is a different product experience from placing four sponsored tiles below a basket.

This is also where governance becomes part of product design. The NIST AI Risk Management Framework emphasises mapping, measuring and managing AI risks. The Federal Trade Commission has separately warned businesses not to exaggerate what AI can do. For retailers, that means the assistant should expose its assumptions, avoid unsupported claims and hand off to deterministic systems when precision matters.[4][5]

A practical architecture

The safest architecture separates responsibilities. Business logic remains the source of truth. AI handles language, intent, summarisation and option framing. Retrieval gives the model current product and policy context. Tool calls fetch prices, stock, eligibility and fulfilment. The final answer is assembled from verified data, with the model explaining rather than guessing.[4]

• Do not let the model calculate final prices when a pricing service exists.
• Do not let the model promise stock without a fulfilment check.
• Do not let the model override promotion rules to make an answer sound better.
• Do let the model ask better questions and compare valid options.
• Do let the model explain which business rules shaped the recommendation.

The real product question

The question for retailers is not whether an app has AI in it. The question is whether the AI is doing the work that rules and recommendations cannot do well: interpreting intent, handling ambiguity, connecting systems and explaining trade-offs. If it only decorates a rules engine, customers and sales teams will learn to ignore it.

A commercial selling assistant should make the retailer's operating model easier to use, not less accountable. The best version will not be the one that sounds most human. It will be the one that understands when to reason, when to retrieve, when to call a system of record and when to stop talking because the business rule has already answered the question.[4]