Before You Pilot AI in Retail, Write the One-Page Decision Contract

By Kirill Naumov · 2026-10-25

Retail AI projects have a curious habit of beginning with a model and ending with a meeting about who was supposed to approve what. By then the demo looks lovely, but nobody wants to put it in a store. I would start with a less photogenic document: a one-page decision contract. It is not a legal agreement. It is a shared description of one decision the system will support, the information it may use, the actions it may take and the point where a person takes over.

Name one decision, not a department

“AI for retail” is not a pilot scope. Neither is “AI for sales”. Try a sentence a shop-floor colleague would recognise: When a customer asks whether a promotion applies to this product, the assistant finds the current approved rule, explains it plainly and shows the colleague its source. Now we can test something real. Was the rule correct and still in force? Was the product eligible? Could the colleague explain the answer without an awkward apology at the till?

A first pilot should be small enough to describe on one page. Choose the customer question, the colleague who handles it, the system that owns the answer and the moment the response must be ready. A broad promise to “increase engagement through AI” will fit nicely on a slide; it will be rather harder to debug on a Saturday afternoon.

Decide what counts as the source of truth

The hard part of the promotion example is not writing a sentence. It is knowing which offer is valid today. Write down the systems and fields that decide it: product identifier, market, channel, customer eligibility, dates, exclusions and approval status. Show the policy or offer version behind an answer. If no approved current source exists, the assistant should say so. Refusing to invent a discount is a successful outcome.

The same applies to staff learning. An answer drawn from last year’s training module may sound polished and still be wrong. Content needs an owner, a review date and a route for colleagues to flag a contradiction. The model cannot repair a policy nobody has maintained.

Give the system an action budget

Make a ladder of permitted actions. In an early pilot, retrieval and explanation may be allowed. A suggested next step can be shown with its reason and source. A customer-facing message can be drafted for a colleague to approve. Changing a price, granting a loyalty benefit or sending a message should remain outside scope until the business has designed the approval, logging and recovery steps. The point is to learn safely, not to make commercial promises by accident.

Define an abstention rule as carefully as an answer rule. Conflicting offer versions, missing consent, uncertain product match, unavailable stock or a question outside the pilot should lead to “I do not know” or a handoff. In loyalty, “do nothing” is also a legitimate recommendation. A customer who has just complained does not always need another personalised offer.

Measure the work, not the applause

Record a baseline before the assistant arrives. For the promotion question, measure correct answers against approved rules, time to resolution, escalations, colleague confidence, customer complaints and the work needed to repair errors. Compare with the existing process, including a simpler searchable reference sheet. If the new tool saves ten seconds at the till but creates an hour of correction elsewhere, the pilot has not earned a rollout.

For training, test a realistic situation later, not just whether people liked a lesson on the day. Agree in advance who reviews failures each week and who can pause the test. A rollback plan is not pessimism; it makes the first experiment easier to approve.

Put a name next to every boundary

Commercial teams own offer truth. Operations own the store workflow. Learning teams own training content. Technology owns access and logs. Someone must own customer communication. “The AI team” cannot be the answer to every question. When those responsibilities are clear, choosing a model becomes easier because the required data, response time, test cases and approval points are finally visible.

A good pilot contract ends with three possible decisions: expand, revise or stop. A dazzling demo is easy to applaud. A store-ready workflow is harder—and worth considerably more. Start with one customer question and one accountable owner. That is enough work for a first page.

Explore the English AI for Retail workshop → · Discuss a retail AI consulting brief

Start a conversation →