Customer-facing AI can move from experiment to recurring expense faster than retailers can determine whether it creates value. A shopping assistant, for example, may require model usage, cloud infrastructure, data pipelines, systems integration, product-catalog maintenance, security controls and ongoing evaluation. If those costs sit across several departments, the pilot can survive without anyone seeing its full economics.

That is the operational risk highlighted by reporting on AI spending and the absence of dedicated cost ownership. Kohl’s shopping assistant offers a useful deployment context: retailers are applying AI to product discovery, but a launch should not be mistaken for proof of return. The right question is not whether an assistant works. It is whether it improves customer and commercial outcomes enough to justify its complete, continuing cost.

Give one owner authority over the full cost

Every pilot needs a named business owner who is accountable for its budget, performance and final disposition. This person should be able to challenge technical choices, obtain cost data from participating teams and recommend scaling, redesigning or terminating the tool. Shared participation is sensible; shared accountability is not.

The owner should maintain a single cost ledger covering model or API charges, cloud compute and storage, integration work, internal engineering time, vendor fees, testing, security and privacy reviews, observability, content or catalog preparation, customer-support impact and post-launch maintenance. Costs absorbed by existing teams still belong in the ledger because they represent capacity that cannot be spent elsewhere.

Finance should validate the accounting method, while technology teams provide usage and infrastructure data. Merchandising, digital commerce and customer-experience teams should agree on outcomes. The owner then has both sides of the equation: the resources consumed and the business value produced.

Set a budget envelope before opening traffic

A pilot budget should be a hard operating envelope, not an initial estimate that automatically expands. Define the test period, eligible user population, maximum interactions, total approved spend and acceptable cost per completed customer task. Add alert thresholds below the cap so the team can intervene before exhausting it.

Separate one-time costs from recurring costs. Initial integration may make a short pilot look expensive, while subsidized vendor pricing may make future operation look deceptively cheap. Build three views: pilot cash cost, normalized monthly run rate and projected cost at realistic production traffic. Include higher-usage scenarios because conversational tools can generate multiple model calls from one customer session.

The pilot agreement should also state what happens at the limit. Traffic pauses, the feature is restricted or executives approve a documented exception. Without that rule, “temporary” overruns become the normal operating model.

What most people miss

Usage is not value. A customer can open a shopping assistant, ask several questions and leave without finding a suitable product or buying anything. High engagement can therefore increase variable AI costs while producing no incremental revenue. Retailers must connect each interaction to a defined customer task and a measurable commercial result rather than treating conversation volume as success.

Measure discovery, conversion and incremental value

For a shopping assistant such as the one deployed by Kohl’s, measurement should follow the customer journey. Discovery indicators can include successful query resolution, product-detail-page visits after an interaction, relevant result engagement, add-to-cart behavior and reduction in searches that return no useful result. Experience measures can include task completion, abandonment, response quality and escalation to human support.

Commercial measures should include conversion rate, revenue per session, gross margin per session, basket composition, returns and cancellations. Gross margin matters more than topline revenue when recommendations steer shoppers toward heavily discounted or costly-to-fulfil products.

Use a control group or another credible comparison wherever possible. The assistant should receive credit only for incremental outcomes beyond what comparable customers would have achieved through conventional search, navigation or merchandising. Segment the results by device, traffic source, customer type and shopping mission; an overall average may conceal a valuable use case alongside several wasteful ones.

Before launch, define the minimum improvement required to proceed and a quality floor that commercial gains cannot override. A tool that lifts conversion while giving unreliable product information, mishandling customer data or increasing returns has not passed.

Track unit economics, not only the invoice

A monthly AI invoice shows expenditure but does not explain whether expenditure is productive. The owner needs an operating dashboard linking technical usage to retail outcomes. Track cost per session, cost per resolved task, cost per product discovery, cost per add-to-cart and cost per incremental order. Compare incremental gross margin with fully loaded variable cost and the recurring fixed cost required to operate the experience.

Monitor the drivers underneath those measures: prompts or requests per session, model calls per request, token or compute consumption, latency, fallback rates, error rates and traffic by use case. Sudden increases can reveal loops, abusive traffic, inefficient prompts, unnecessary context or routing to an expensive model when a simpler method would suffice.

Assign thresholds to each driver and review them frequently during the pilot. Cost controls may include rate limits, session caps, caching, retrieval improvements, smaller models for routine tasks and conventional search for queries that do not require generation. Optimization must preserve the pre-agreed quality floor; a cheaper assistant that gives poorer answers can damage conversion and trust.

Run a formal scale-or-stop review

Schedule decision gates before the pilot begins. An early review should test technical stability, safety and instrumentation. A midpoint review should examine budget burn, customer behavior and emerging unit economics. The final review should compare results with the predefined commercial, quality and cost thresholds.

Allow four explicit outcomes: scale, continue within a tightly defined extension, redesign or stop. An extension needs a specific unresolved question, additional budget and a new deadline. “More data” is not enough. Redesign is appropriate when a narrower use case appears valuable; stopping is appropriate when incremental value remains below total cost or quality risks cannot be controlled.

Scaling should trigger a fresh production forecast, procurement review and operating plan. It should not happen by leaving the pilot switched on. The accountable owner must document the decision, evidence, assumptions, future budget and conditions that would cause the retailer to reconsider.

Retail AI cost-governance checklist

• Name one business owner with budget and stop authority.
• Create a ledger covering infrastructure, models, integration, labor, vendors, governance and maintenance.
• Separate one-time pilot costs from normalized production costs.
• Set limits for duration, traffic, interactions and total spend.
• Define discovery, conversion, gross-margin and customer-experience metrics before launch.
• Use a control group or credible baseline to measure incremental value.
• Track cost per session, resolved task and incremental order.
• Monitor model calls, consumption, latency, failures and fallback behavior.
• Set commercial thresholds and non-negotiable quality, privacy and safety floors.
• Schedule early, midpoint and final decision gates.
• Require a bounded budget and deadline for any extension.
• Record an explicit scale, redesign or stop decision rather than allowing passive continuation.