AI-enabled work in retail: reconcile a promotion before it reaches customers

Define the sellable offer first

Start with a promotion ID and write the offer as structured fields: product or SKU set, selling price, reference price basis, mandatory charges, customer eligibility, channel, geography, start and end time, stock scope, exclusions, substitution rule, and rollback owner. Treat copy as an output of those fields, not as the source of truth.

The UK Competition and Markets Authority's current price transparency summary says pricing information should be clear, complete, and accurate, and explains the need to show the total price when customers are invited to purchase. The UK government's Price Marking Order guidance describes selling and unit-price information, including its treatment for promotional items. These sources are jurisdiction-specific. A real promotion needs review against the rules and policies that apply where it is offered.

Sources: [1] [2]

Build the promotion readiness sheet

Give each row one testable condition rather than a broad approval. Useful rows include:

Add a status of ready, blocked, exception under review, or not applicable to every row. One overall green marker is too coarse. It can hide a correct web price beside a stale shelf label or a valid offer beside an unsupported availability claim.

• the current and reference price by channel, with source system and extraction time;
• unavoidable fees and taxes, with the customer journey where they first appear;
• promotion window and time zone for each channel;
• eligible SKU, pack size, location, customer group, and redemption limit;
• available-to-promise inventory and the age of the feed;
• unit-price treatment and label rendering;
• every customer-facing claim and the exact field that supports it;
• email, web, app, shelf, paid media, and support-script versions;
• stop conditions, monitoring owner, and rollback route.

Worked artifact: a synthetic weekend campaign

A fictional retailer prepares a weekend promotion for a countertop mixer. The campaign sheet says the item will be 20 percent off in all channels from Friday morning through Sunday night. An AI assistant drafts: “Save 20 percent this weekend, available everywhere while stocks last.”

The readiness sheet finds four conflicts:

| Check | Evidence | Result | Action | | --- | --- | --- | --- | | Web price | commerce export at 09:10 | correct for Friday start | ready | | Store price | point-of-sale schedule | starts six hours later | block all-channel copy | | Stock | inventory feed at 02:00 | two regions below reserve level | limit geography and refresh | | Reference price | price-history extract | one channel used another price during the comparison window | send basis for qualified review |

The team changes the offer to channel-specific language, removes the unsupported “available everywhere” phrase, refreshes stock before send, and holds the savings comparison for the affected channel. The record links the revised copy to the price export, inventory timestamp, approval, and deployment version.

The example is synthetic. It illustrates evidence reconciliation, not a legal conclusion about a particular promotion or market.

Run a channel replay before launch

Verification should replay the offer as a customer would encounter it. Use representative products, locations, customer types, devices, and basket states. Confirm the displayed selling price, unit price where relevant, unavoidable charges, eligibility message, stock treatment, start and end behavior, and support response.

Include boundary tests. Check one minute before and after the promotion window, an excluded region, a low-stock location, a logged-out customer, a product variant outside the SKU set, and a basket containing a conflicting discount. Capture screenshots or structured outputs with time, environment, input, expected result, actual result, and reviewer.

AI can help generate test cases or compare channel outputs, but the launch owner should inspect material discrepancies and decide whether to fix, narrow, delay, or stop the offer. A model's assertion that two prices match is not the same as evidence from the live or pre-production channel.

Monitor the offer as an operating decision

After launch, watch for price mismatches, feed age, stock exceptions, expired copy, fee differences, redemption errors, and customer-contact themes. Set explicit stop triggers, such as a reference-price mismatch or a promotion continuing past its approved end time. Link each incident back to the promotion ID so the team can see which source, transformation, or deployment failed.

Do not interpret high conversion or low complaint volume as proof that the offer was represented accurately. Operational metrics and evidence quality answer different questions.

Two Prune can assess how a participant coordinates a bounded retail promotion, reviews source evidence, uses AI, detects conflicts, and communicates a launch decision. It does not run retail operations, determine legal conformity, promise commercial outcomes, or make employment decisions.