How Xpol grew order capacity without hiring, and kept a retiring specialist's knowledge in the business

When volume grows faster than the team can

Fresh flower supply chains run on tight windows and tighter margins. Xpol — a Dutch wholesaler moving flowers from African farms to retailers across Europe — sits in the middle of that chain. Order volume is climbing. Product and customer complexity is climbing even faster. And in May 2026, one of the order-entry specialists is retiring, adding even more pressure on a small team.

The default response would have been to hire a replacement and hope the knowledge transfer goes smoothly. Xpol chose a different path. Rather than adding headcount to absorb growth, they built an AI agent that could carry the workload and preserve the team's accumulated rules — turning years of customer-by-customer learning into a system the whole desk could work from.

An agent that reads anything, writes one thing

Lleverage built an AI agent around a dedicated order intake inbox. Customers email orders to the inbox in whatever format they already use. The agent reads PDF, Excel, CSV, and portal exports, applies the customer-specific ruleset, and pre-fills a webshop order in Business Central for a human to do a final review.

The design principle was deliberate: the agent never posts a final sales order on its own. It prepares a draft that a person signs off with a single click. Ambiguous items are left blank rather than guessed. That is the design, not a workaround.

"It's a matter of building trust in the organisation with these kinds of initiatives. You can't just throw something like this over the fence."

— Cees Maaskant, General Manager, Xpol

Leaving ambiguous fields blank does two things at once. It keeps the human in the loop on the exact cases where judgment is required, and it forces the source data to improve over time. Every blank field is a signal that a framework contract has overlapping validity dates or a customer code needs cleaning up. The agent turns every ambiguous order into a cleanup prompt.

Turning tribal knowledge into codified rules

The build ran through working sessions with Xpol's domain experts and the business information coordinator. Each session pulled out the rules that lived in the team's heads — the conversion factors, the weekday label logic, the depot-splitting conventions — and turned them into configuration inside the agent.

Rollout was phased rather than big-bang. The largest retailer went live first, and over the following weeks, the supplier set covering roughly 60 to 70 percent of weekly volume. Long-tail customers — the ones sending one order a week — are being added in parallel as the agent's rule library grows.

Live with multi-national retailers

The agent went live in production on 20 April 2026. It is now processing orders for some of the largest French, Swiss and Dutch retailers in Xpol's customer book. The rules the supply chain team had accumulated over years of customer-by-customer learning are codified in the agent rather than depending on any one person being in the building — which matters more when a retirement is weeks away.

One week into production, the team reports roughly 20 minutes saved per large order. Across the 150 orders the desk processes in a typical week, that adds up to more than a full team member's time returned every week. The figure will firm up as more suppliers come online, but the direction is clear: the manual data-entry step that used to define the order desk's day is gone, replaced by a review step that takes a fraction of the time.

What made this work was not only the agent itself, but three decisions around it. Start with the priority supplier set that covers the majority of volume, and expand from there. Build human review into the design rather than bolting it on later — blank fields are a feature that drives data quality, not a limitation to apologise for. And treat a retiring specialist's knowledge as an asset worth codifying, not a handover problem to manage.

Looking ahead

Xpol is scoping production and/or supply chain planning as the next use case for AI. The logic that drives weekly forecasting and supply allocation — what Cees describes as calculated risk-taking built on years of operational experience — is the kind of rules-and-judgment problem the order-entry agent has already proven a fit for.

A reporting layer is also in scope. Dashboards tracking orders by supplier, ambiguity frequency, and processing trends will feed directly back into the source-data cleanup loop the order agent already drives.

The long-tail suppliers currently running in parallel to the priority set will move to production as their rulesets stabilise, extending automated coverage to the full customer book.