AI Consulting for Retail

On some lines, comfortably. On others, no — and knowing which is which is the entire value of the exercise. Forecasting models beat human planners where there is genuine history, real seasonality and enough volume for a pattern to exist: your top-selling, consistently ranged lines. They lose badly on the long tail, on new products with no history, and on anything driven by a one-off event the data has never seen. The failure mode we see most is a retailer buying a forecasting platform, applying it across the whole range, watching it produce nonsense on the tail, and concluding AI does not work. It worked fine on the 200 SKUs that fund the business, and was always going to fail on the 3,000 that sell twice a year. The right question is not whether to forecast with AI, but which slice of your range has enough history to justify it — and that is answerable from your own data before you spend anything.

Tighter than the technology vendors imply, and the enforcement has been real. The Privacy Act 1988 applies to personal information you collect, and biometric information — including facial images used for identification — falls into the sensitive information category, which attracts the strictest requirements in the Act. The OAIC has made determinations about facial recognition use in Australian retail settings, and the reputational consequences landed harder than the regulatory ones. The practical guidance we give retailers: personalisation from your own transaction data, with the customer’s knowledge and a real opt-out, sits comfortably in normal territory. Anything involving biometrics, cross-site tracking, inferred sensitive attributes, or data the customer would be startled to learn you held is a decision for your privacy adviser before it is a decision for your technologist. The startle test is not a legal standard, but it predicts the news story quite well.

Yes, and it is one of the fastest paybacks in retail — with one hard constraint that people underestimate. The Australian Consumer Law prohibits misleading or deceptive conduct, and it does not care that the claim was generated. If a model writes "machine washable" about a dry-clean-only garment, or invents a specification, or embellishes a warranty, that is your representation to your customer and your exposure. The architecture that works: generate from your actual product data — the supplier’s specification sheet, your attribute fields — rather than from the model’s general knowledge, and treat any claim not present in the source data as a defect rather than a flourish. Then review before publish. Retailers who wire the model to their PIM and constrain it to the source data get the speed without the exposure. Retailers who ask a chatbot to "write a description for a wool jumper" get invented facts at scale, which is a worse problem than the one they solved.

For routine, high-volume, verifiable questions, yes — where is my order, what is your returns policy, do you have this in a size 10 at Chadstone. Those questions have factual answers sitting in a system, and answering them instantly at 9pm is a genuine service improvement, not a cost-cutting exercise dressed up as one. The design rules that matter: the agent should answer from live system data rather than memory, it should say plainly when it does not know instead of improvising, it should hand over to a person without making the customer repeat themselves, and it should never handle a complaint, a dispute or a distressed customer. Consumer guarantees under the Australian Consumer Law are not something to have an AI improvise about — a wrong answer on a refund right is both a legal and a reputational problem. The retailers who get this wrong are the ones who route everything to the bot to reduce headcount, then discover the escalation path was the product all along.

The gap is narrower than it was, and it narrows in a specific place: you no longer need a data science team to do the things that used to require one. What you do need is data discipline, and that is where mid-sized retailers genuinely struggle — product data scattered across a POS, an e-commerce platform, a supplier spreadsheet and someone’s head; inventory that is accurate in theory; customer records duplicated across three systems. Those problems are not solved by buying an AI platform, and every AI platform sold to a mid-sized retailer quietly assumes they are already solved. So the honest sequence for most retailers is: fix the product and inventory data, then automate the customer-facing and forecasting work on top of it. That is less exciting than what the vendor promised, and it is the reason their pilot did not work.

A free initial consultation first — a conversation about margin, range, service load and where the actual constraint sits, with no demo. If there is a real opportunity, the next step is generally the AI Opportunity Audit at around $3,000: a written map of your systems and data, an honest assessment of whether the product, inventory and customer data can support what you want, the privacy constraints attached, and a ranked shortlist with realistic effort and payback. Yours to keep either way, including the finding that your first project should be a data clean-up. Where it stacks up, we build and ship the first workflow into production rather than handing over a report. The audit generally earns its fee at around 20 or more staff — and we would strongly recommend not deploying anything customer-facing in the eight weeks before Christmas.