Promod wants its merchandisers to question the company’s data in their own words. “There is no magic,” Claire Thelliez, its director of transformation and IT, told Républik Retail: projections and other advanced uses first need a better “context layer.” Opening the data to the business moves the bottleneck from writing queries to writing down what the numbers mean.

The 400-store chain wrote its own ERP and e-commerce platform, yet a dashboard usually meant a request to the data team. After a three-month trial of Myriade, a Lille start-up founded in 2025, about fifty people use the platform, roughly half the IT department. Thelliez says it lets Promod “work faster and more simply.” Of the first agents to venture beyond IT, one flags order-logistics problems that once required a SQL query; another checks that commercial policy is being applied. Both read the record; a projection has to interpret it. The split follows the budget as much as the glossary: Myriade’s license carries a per-user token allowance, and Thelliez keeps pure dashboarding out of the AI’s hands because other tools already do that work without spending any.

Google, whose cloud hosts Promod’s BigQuery data platform, has already made room for someone to write that interpretation down. Its documentation for data agents built on BigQuery includes a field called “glossaries,” for definitions of business terms. Such “authored context” is optional, it says, though it makes the agent’s answers more accurate. The glossary is optional, then, in the sense that accuracy is.

Myriade sells the context layer as something a machine can mostly build. Its homepage says agents “document 90% of your tables without supervision,” leaving a catalogue “complete, correct, and live in hours, not months.” A customer quoted there praises its semantic “magic.” The FAQ is more patient: business users get in “eventually,” once the data is “cataloged and quality-verified.” Meanwhile, “when something is unclear, Myriade asks your team,” and the data team “reviews, corrects, approves.”

The unclear things are the ones the data team used to settle without being asked. A merchandiser who wanted last month’s sales got a number with decisions already inside it. Returns came off, or they did not, depending on who wanted the number. About a third of Promod’s stores are franchised, and a franchised store’s “sales” can mean what Promod sold the franchisee or what the franchisee sold a customer. The analyst made those calls, and the number did not mention them.

A warehouse keeps every sale and none of the arguments about what counts as one.

Promod’s own timetable puts the glossary first. Thelliez gives the context problem as one reason the rollout began inside IT, where the machine’s answers are still checked. Within twelve months, the platform should also be enriching the data catalogue and analysing commercial events. More autonomy for commerce, merchandising and finance comes afterwards. Replenishment, already on the list, is a projection with a delivery date.

Data teams, Myriade’s homepage promises, “get their time back.” The method described on the same page spends that time answering the machine, which asks what the analysts once decided on their own. This time the answers will be written down, where the merchandisers can finally argue with them.