How should AI support ecommerce trading decisions?

AI should help ecommerce leaders examine evidence, compare options and identify decisions that need attention. Its authority to change prices, promotions, advertising or inventory should be deliberately limited. The commercial owner remains responsible for the trade-offs and consequences.

A useful starting point is a trading decision: whether to discount ageing stock, redirect media spend or protect availability ahead of a promotion. Define the objective, constraints and evidence needed before deciding where AI belongs in the workflow.

Establish the numbers before interpreting them

Agree metric definitions with finance and trading leaders. Specify what revenue includes, how refunds and cancellations are treated, and which costs sit within the contribution measure used for decisions. Make the period, currency, tax treatment and comparison basis explicit.

Calculate those measures through tested queries, transformations or calculators outside the language model. The same inputs and rules should produce the same result. Give AI access to the resulting figures, their definitions and their source records. Ask it to explain movements, challenge assumptions and propose options, while keeping the arithmetic independently verifiable.

Missing product costs should produce an incomplete contribution calculation and an exception for its owner. They should never become plausible estimates silently supplied by the model. Where a decision needs an estimate, label it, record who approved it and show how it affects the recommendation.

Reconcile the commercial evidence

Treat the online store, ERP and paid media accounts as different evidence sources. Establish which system owns each measure and how records join across orders, products and campaigns. Check for mismatched identifiers, reporting periods, refunds, fulfilment status and currency before drawing conclusions.

Set expectations for data freshness and reconciliation. A trading view may need provisional numbers before finance closes a period; mark them accordingly. Keep unresolved differences visible, with a named owner and a threshold for pausing affected decisions.

AI can help investigate discrepancies and assemble supporting records. It should not write a convincing explanation that makes unmatched totals appear reconciled. Restrict access to the information the task needs; a product-level trading decision rarely justifies unrestricted access to customer records.

Judge contribution alongside revenue

Consider an illustrative retailer deciding whether to extend a discount. Orders and revenue have increased, but the decision also depends on product costs, fulfilment costs, refunds, media spend and the treatment of discounts within the agreed contribution calculation.

Compare the extension with alternatives: ending the offer, narrowing eligible products or clearing a smaller quantity. Calculate each scenario using explicit assumptions, including demand that might otherwise have arrived at full price. Label forecast uncertainty rather than presenting one precise answer as settled.

The commercial decision may favour clearing stock and releasing cash despite a lower margin. It may favour protecting availability or customer trust. AI should make those trade-offs easier to inspect. It should not choose the business objective because one metric is easiest to optimise.

Keep attribution separate from causality

Attribution allocates conversion credit to recorded interactions. Google’s attribution documentation describes that allocation; its Conversion Lift documentation describes comparing exposed and control groups to measure incremental effects. These answer different questions.

Require an AI recommendation to identify which evidence it uses. Do not accept attributed return on advertising spend as sufficient proof that increasing the budget will create equivalent additional sales. Avoid adding several platforms’ attributed conversions together and treating the total as unique orders.

For a material budget decision, consider an appropriately designed experiment with measurement support. Where that is impractical, state the uncertainty, limit exposure and agree what evidence would justify continuing or reversing the change. A persuasive explanation cannot supply a missing counterfactual.

Grant authority by action and consequence

Begin with read-only analysis and reviewed proposals. If an agent later receives permission to act, enforce its limits in the connected systems and execution code, independently of its prompt.

• Prices: apply approved floors, exclusions and maximum changes; escalate anything outside them.
• Promotions: check eligibility, timing and discount combinations before activation. Require approval for changes to the offer customers receive.
• Media: enforce daily and cumulative spend limits, with approval thresholds reflecting the budget and potential loss.
• Inventory: separate recommendations from purchase orders or stock reallocations that commit cash or affect availability.

Tie approval to the exact proposed values, affected records and current evidence. A changed proposal needs fresh approval. Recheck conditions immediately before execution so yesterday’s stock position cannot authorise today’s change.

Make ownership part of the trading workflow

Name the trading decision owner, the finance owner of metric definitions and the technical owner of access and execution. Agree who monitors changes, who can stop the agent and who coordinates an incident. An incorrect promotion needs a response that includes customer impact and remediation, not just a corrected prompt.

Record recommendations, approvals, executed changes and subsequent outcomes. Monitor realised contribution, spend, stock availability and data failures against the agreed expectations. Prepare rollback where feasible and a containment plan where a commitment cannot simply be undone.

Build this into the team’s existing decision process. Involve traders, analysts and operational colleagues in testing realistic exceptions; make it easy to challenge a recommendation and explain an override. Review whether the workflow improves decisions without creating excessive approvals or investigation work.

For a first implementation, choose one bounded decision and establish its evidence, authority and ownership through AI governance and implementation planning. Expand the remit only when the business can explain both the commercial benefit and how it will manage a failure.

Measurement sources reviewed on 25 September 2026. The retailer scenario is illustrative.