Question 1

How do you keep PCI-DSS scope contained when AI services touch POS data?

Accepted Answer

PANs are tokenised at the perimeter (P2PE-validated terminals where available, network tokenisation otherwise) so the analytics lake and ML feature store only ever see tokens. The cardholder data environment (CDE) is a thin enclave: POS, payment gateway, key management. Inference services that score transactions for fraud or basket anomalies run inside the CDE on the token-and-metadata view; everything else — forecasting, personalisation, returns scoring — runs outside CDE on tokens. Reference deploys on AWS PrivateLink / Azure Private Link with customer-managed KMS keys, segmented VPCs, and quarterly ASV scans on the CDE perimeter only.

Question 2

How do you forecast demand for new items with no historical sales?

Accepted Answer

New-item forecasts use a hierarchical Bayesian model that borrows strength from lookalike SKUs. We embed every existing SKU on attributes (category, subcategory, price band, fabric / material, colour family, vendor, lifecycle stage, store-cluster fit) and pull the posterior of the k nearest lookalikes as a prior on the new SKU. The first 2–4 weeks of actuals get Bayesian-updated against that prior. Typical MAPE is 18–25% for the first month — meaningfully better than naive "category average" — and tightens to sub-15% by week six as real signal accumulates.

Question 3

How do you stop a conversational shopping agent from hallucinating inventory or prices?

Accepted Answer

Three layers. (1) The agent has no implicit knowledge of stock, price, or promo eligibility — those are tool calls against Shopify Plus Storefront API, Salesforce Commerce Cloud OCAPI, or the merchant's OMS at message time, with a 30-second TTL cache for read-heavy reference data only. (2) Output is rendered against a strict catalog schema; any product mentioned that does not resolve to a live SKU is stripped before send. (3) Price, availability, and promo claims pass through a deterministic policy engine that re-checks against the source of truth and overrides the LLM if it drifted. Net result: the agent can be wrong about taste but cannot be wrong about whether the size 8 is in stock or whether the 20% off code applies to that brand.

Question 4

What does post-cookie personalisation actually look like for a mid-market DTC brand?

Accepted Answer

Step 1: get to 60%+ first-party identity coverage — email at checkout (mandatory), SMS opt-in at landing (incentivised), account creation soft-walls on the second visit. Step 2: stitch identity in a CDP (Segment, mParticle, Lytics, or a custom Iceberg-based one if scale justifies). Step 3: for off-platform reach, use clean rooms (LiveRamp, AWS Clean Rooms, Habu) to join with retail media networks and CTV partners on hashed PII rather than cookies. Step 4: use Google Privacy Sandbox Topics / Protected Audiences for the long-tail open web. On-site, the recommender runs on logged-in + cookied-anonymous-stitched-to-logged-in sessions; for true first-time anonymous visitors, you fall back to content-based and bestseller signals — and that's fine, conversion lift from personalisation on a cold visitor was always overstated.

Question 5

How does markdown optimisation interact with IFRS / GAAP retail inventory accounting?

Accepted Answer

Two valuation methods care. Retail Inventory Method (RIM) cost ratios get distorted if markdowns are taken aggressively at end-of-season — the cost complement drops and inventory valuation falls, which can flip a quarter. So the markdown optimiser is constrained: it can recommend a markdown cadence, but the GL impact is pre-computed under both RIM and weighted-average-cost so finance sees the P&L tradeoff before approval. For SKUs subject to lower-of-cost-or-net-realisable-value (IAS 2 / ASC 330), the optimiser flags any markdown that would push NRV below carrying cost so the writedown is recognised in the right period rather than smeared. The model output is a recommendation set; finance owns the cadence call.

Question 6

How do you run BOPIS-aware replenishment on Manhattan Active Omni?

Accepted Answer

The forecast splits store-level demand into ship-to-home, BOPIS pickup, BORIS returns inflow, and walk-in. Manhattan Active Omni is the OMS source of truth for available-to-promise and the inventory pool model (single pool vs reserved BOPIS pool vs ship-from-store). Our replenishment recommender writes back to MAO's allocation engine via its REST integration, respecting the existing inventory segmentation rules. The model accounts for the fact that BOPIS demand looks like store demand at the store but is initiated online — so promo-driven BOPIS spikes get attributed to the right elasticity bucket and don't poison the walk-in baseline forecast.

Question 7

How do you tune return-fraud detection so you do not punish good customers?

Accepted Answer

Return fraud is a precision-not-recall problem — false positives cost loyalty, true positives recover dollars. The model scores returns on serial-returner behaviour, wardrobing patterns (high-AOV returned within tag-attached window), receipt-mismatch, store-vs-online channel arbitrage, and identity-graph signals (same payment method across multiple accounts). We tune the operating point against a calibrated cost matrix: in apparel a typical balance is to flag the top 2–3% of returns for manual review where the expected recovery exceeds 8x the labour cost of review, and to soft-decline (require receipt, no cash refund) the top 0.3%. Hard declines are reserved for confirmed organised-retail-crime patterns with corroborating evidence, never on model score alone.

Question 8

How do you measure promo lift in a way that survives a test-vs-control critique?

Accepted Answer

Naive year-over-year promo lift is useless — it confounds promo with seasonality, weather, competitor activity, and category trend. We run geo-based or store-cluster-based holdout tests where matched control stores or DMAs do not get the promo. The estimator is a synthetic-control or difference-in-differences design with parallel-trends pretests; for digital, we use intent-to-treat with randomised audience holdout. Where holdouts are politically impossible (the merchant won't starve their best stores), we fall back to Bayesian structural time-series (CausalImpact-style) with covariates for category baseline and competitive activity. Every reported lift comes with a confidence interval and a parallel-trends diagnostic — not a single point estimate.

Retail AI: sub-10% MAPE forecasts, BOPIS-aware replenishment, return fraud caught at the POS

Platforms we integrate with

Six AI systems we ship to omnichannel retail & DTC

Demand forecasting at SKU × store × week

Markdown, promo & price optimisation

Post-cookie personalisation & recommendations

Conversational shopping grounded in inventory

Loss prevention & return-fraud detection

Visual search & catalog enrichment

Retail-grade data infrastructure

Commerce & POS event ingestion

Customer Data Platform & identity stitch

Order, inventory & OMS truth

Clickstream, search & session analytics

Product catalog, PIM & enrichment

PCI-DSS scope containment & deployment

Retail AI questions we get from retailers