Growth outran the fax machine, basically
Eataly spent most of its history opening a new store roughly once every six years. Then it opened ten in a single year, and order and transaction volume climbed more than 40% over two years. The Italian grocery and restaurant chain's supply chain communications had not scaled with it. Purchase orders, invoices, and shipment confirmations moved through email, PDFs, paper, and WhatsApp, a stack that was adequate for a slower-growing company and was, in the company's own description, no longer keeping pace once growth accelerated.
That gap is familiar to any enterprise technology leader who has watched a business unit's growth outrun its back-office systems. The symptoms are rarely dramatic at first: a few more exception emails, a few more invoice disputes, a few more days of delay on purchase order confirmation. They compound quietly until the manual process becomes the actual constraint on how fast the business can grow, which is roughly where Eataly found itself before it acted.
The fix was boring infrastructure, not AI
Eataly's response was not an AI pilot. It implemented Cleo's electronic data interchange platform to standardize and automate supply chain communications, migrated its ERP backbone to Microsoft Dynamics 365, and added Cleo's Transaction Monitoring and Management service, later upgraded to the Plus tier, for ongoing support. Purchase orders now transmit automatically, with buyers receiving confirmation and exception details in about 30 minutes instead of days. Thousands of vendor invoices are processed monthly through EDI feeds into Dynamics 365, which matches invoices against purchase orders and receipts so that matching invoices can be posted and paid without manual review.
None of that is glamorous, and none of it will show up in a vendor's AI product marketing. But it is the layer that makes everything built on top of it trustworthy. Cleo's transaction management service adds a one-hour response commitment when something fails, with tickets generated automatically that include the document, a job ID, and the required action, backed by a dedicated project manager. That is infrastructure discipline applied to a problem that most retailers still handle with a shared inbox and institutional memory.
Only then did the agentic AI project start
With EDI and ERP in place, Eataly has begun building an AI agent that reviews active warehouse inventory and flags products nearing expiration. That is a narrow, specific use case, and it is notable precisely because of how narrow it is. An expiration-flagging agent needs structured, timely, trustworthy inventory data to function at all, and Eataly did not have that data in usable form until the EDI migration gave it one. Building the agent first, against the old patchwork of email and PDF records, would have produced an agent confidently flagging stale or wrong information.
This is the part of the agentic AI conversation most vendor pitches skip. An agent is only as good as the data pipeline feeding it, and most retailers evaluating agentic AI today have not done the equivalent of Eataly's EDI work. They are closer to where Eataly was before this project started: real business pressure, real growth, and a back-office data layer that cannot yet support the kind of automated judgment an agent is being asked to make.
The sequencing lesson is worth stealing
Enterprise AI budgets in 2026 have skewed heavily toward agent pilots and copilots, often funded ahead of the data infrastructure work that would make those agents reliable. Eataly's approach reverses that order: fix the plumbing, prove it with a boring but measurable win like a 30-minute purchase order turnaround, and only then point an agent at the resulting clean data. The expiration-flagging agent is a comparatively low-risk first agentic use case, since a wrong flag costs an inventory review, not a financial write-off or a compliance violation.
That risk profile is itself a lesson in sequencing agentic AI rollouts generally. Retailers under pressure to show an AI initiative to the board often reach for a customer-facing use case first, because it is visible, rather than a back-office use case like this one, where failure is cheap and recoverable. Eataly's path suggests the better order is internal, low-stakes, and data-dependent use cases first, with customer-facing agentic bets following only once the organization has evidence its data foundation can support them.
What to check before funding the next agent pilot
Any CTO or CIO evaluating a new agentic AI proposal should be able to answer a version of the question Eataly answered for itself: does the data this agent needs already exist in structured, current, and trustworthy form, or does the project quietly depend on a data cleanup that nobody has budgeted for? If the honest answer is the latter, the EDI or ERP work belongs in the project plan and the budget, not treated as somebody else's problem to solve later.
The commercial case for sequencing the work this way is straightforward. Eataly's 30-minute purchase order turnaround and automated invoice matching are already paying for themselves in labor hours and reduced exceptions, independent of whatever the expiration agent eventually delivers. The infrastructure investment functions as a separate win that also happens to satisfy the AI project's prerequisite. Retailers that fund the data work as its own initiative, with its own measurable return, will find the subsequent agentic AI pilot both cheaper to build and far more likely to survive production.
