Where AI and human judgement work together. A scalable way to categorise products without taking judgement away from the people who know them best.
With Marketplace approaching, Tesco's manual process could not keep up with the products it already had to categorise.
The launch would add millions of third-party listings. Products placed in the wrong category were harder to find through search and navigation. As Lead UX Designer in a five-person core team, I led the design of a system that categorised clear cases automatically and gave Online Operations colleagues the uncertain ones to review.
We shipped the MVP in six months, before Marketplace went live.
The manual process was already asking too much of the colleagues who kept product categorisation moving.
Even experienced Online Operations colleagues could categorise around 100 products a day alongside their other work, while more than 10,000 products arrived each week across Tesco's digital channels. Supplier descriptions were often incomplete or inconsistent, and some products could fit several categories. Colleagues had to work through that ambiguity as well as the volume.
To understand where the pressure sat, we followed products from their arrival in Tesco's systems through to Tesco.com and watched colleagues work. They described repetitive effort, difficult decisions and little feedback on their progress. With Marketplace approaching, hiring more reviewers would be costly and still would not keep pace with incoming products, either before launch or once Marketplace was live.
We traced the existing process and tested ideas with real product data, returning to earlier decisions as we learnt more from colleagues.
Following products from arrival to Tesco.com showed where they passed between teams and where colleagues had to make judgement calls.
Real product data showed which matches were clear and which still needed colleague judgement because descriptions were incomplete or several categories seemed plausible.
The handoff had to let colleagues accept a good match or correct a bad one with enough context to decide. I worked through those decisions with product and engineering before shaping the interface.
Realistic product volume exposed where review would become slow or repetitive, and colleague feedback guided each revision.
Filtering slipped from the MVP. We made the trade-off and follow-up delivery date explicit, then shipped the feature three weeks later.
AI categorised clear matches automatically. Colleagues reviewed products where the right category was uncertain.
The review workspace put that handoff into practice. Colleagues could approve, reject or remap AI suggestions, and their corrections fed back into the model.
A conceptual view of the system layers, with colleague judgement between AI categorisation and the customer experience.
The review workspace needed to help colleagues make informed decisions, handle volume and pass corrections back into the system.
Put product details, the AI suggestion and the taxonomy in one view, so colleagues could judge a match without searching across screens.
Give reviewers clear ways to approve, reject or remap a suggestion when the proposed category was wrong.
Use batch actions for repeated decisions, with a route into detailed review for products that needed closer judgement.
Capture corrected mappings and feed them back into the model to improve future suggestions.
The MVP let colleagues start with the workload, act together on routine products, or inspect an uncertain category in detail. The screens below span the MVP, later releases and tested concepts, not one before-and-after build.
The queue overview grouped work into Automatically mapped, Require approval and Unmapped buckets, with counters and trends showing where review was needed before opening a product.
In the categorisation table, colleagues selected products with checkboxes and used a floating action button to approve, remap or reject them together. The separate location browser pictured below shipped later.
The three-pane Inspector kept product details, the AI suggestion and a taxonomy tree together. Colleagues could drag locations into place and sort their priority. The scores shown below came later; variant views were designed and tested.
Filtering shipped three weeks later and the fuller dashboard eight weeks later. Other later-shipped tools included a separate taxonomy browser for scoped review, filters including proposition, location and confidence score, and quick lists. AI insights and richer review and performance metrics were designed and tested, but are not all confirmed as shipped.
The system went live across Tesco's online propositions before Marketplace launched, and the AI earned its place beside the people who reviewed its work.
The hardest part of this project was never the technology. It was keeping the AI clearly helpful without making the colleagues who ran the established manual process feel replaced by it. Everything we designed carried that balance: plain status cues instead of confident-sounding numbers, batch actions that took the repetitive work off their desks, and a review loop in which their corrections changed what the AI suggested next.
Colleagues started as sceptics. They had built a way of working that kept products categorised for years, and doubted a machine could take on the parts they knew best. We answered that with evidence rather than promises: they watched the system flag the products it was unsure about, they saw their corrections change its behaviour, and they kept the final say on every decision. Within weeks, the reviewers who doubted it most were the ones planning their day around it.