The cooperative's 55 stores had adopted AI across three distinct capabilities, smart-cart checkout, online order picking and substitution suggestions, and each had arrived through a different vendor relationship. That history left three separate systems, each holding its own copy of product and inventory data, with no shared source of truth between them.

Integration work paid the price for that fragmentation. Every new AI use case meant a bespoke, months-long project to connect it to whichever vendor's data copy it needed, because there was no standard way for a new capability to plug into what already existed. Teams essentially rebuilt the same plumbing every time.

Oversight suffered just as much as engineering speed. With three vendor silos running independently, no one had a shared view of model behaviour, data access or spend across the estate, so it was hard to say whether the cooperative's AI investment across smart carts, picking and substitution was actually paying off in a comparable way.

We architected a unified enterprise AI platform rather than another point integration bolted onto the existing three-vendor sprawl. At its centre sits a shared product and inventory feature store, giving the smart-cart, picking and substitution systems one live source of data instead of three separate copies drifting apart from each other over time.

Around that feature store we built a common model-serving and monitoring layer, so every model, regardless of which capability it supports, runs and is watched the same way. Standard integration APIs let the smart-cart, picking and substitution systems all consume the platform through one consistent interface rather than three bespoke ones.

Central dashboards now track model performance and data access across the whole estate, replacing the three-way blind spot with one governed foundation. A new AI use case connects through the standard APIs rather than starting its own bespoke, months-long integration from scratch, the way every earlier use case had to.