Rolling Out Generative AI at a Logistics HQ Without Breaking the Field

Context

A logistics and transport group operating across several continents employs more than ten thousand people. Its business combines a headquarters concentrating the steering functions and a dense network of branches and warehouses where daily operations actually happen.

Management had launched an ambitious program to integrate generative AI: writing copilots for support functions, route optimization tools, document-processing assistants for customs operations and compliance. The program had been designed and run from headquarters, by a transformation team reporting to senior management.

The technical rollout went without major incident. The tools worked. HQ teams used them. The field, however, stayed largely on the sidelines.

Problem

Six months after launch, a yawning gap had opened up between headquarters and operations. At HQ, adoption exceeded 70%. In branches and warehouses, it did not break 15%. The route optimization tools, designed for the field, were the least used of all.

The stakes were strategic. The program’s expected value sat mostly in operations, where the optimization margins were largest. A rollout successful at HQ but absent from the field captured only a marginal fraction of the expected return.

The transformation team initially blamed ergonomics or training. The field’s own feedback told a different story. The tools were seen as built by people who did not know what daily life on routes and in warehouses actually looks like.

Intervention

Management changed tack. Instead of pushing harder from headquarters, it went looking on the ground for the profiles who could carry adoption from the inside. To do so, it applied the Potential Stack — not to HQ managers, but to branch staff.

The exercise surfaced a population that had been invisible until then: team leads, warehouse managers, and route planners who appeared on no talent pool but who showed a rare combination — an intimate knowledge of the field, genuine curiosity about the new tools, and the ability to explain things simply to their colleagues. Neither the most credentialed nor the best-rated, but the best placed to bridge the gap.

The Fortress, Front Line, Laboratory doctrine structured what followed. Critical operations that engaged safety or customs compliance were kept as Fortress, out of reach of any push for speed. Optimization and planning tasks — where AI could produce a real gain — were treated as Front Line. And a few volunteer pilot branches were set up as Laboratory to test new uses before generalization.

The rollout had failed not because the field rejected the tool, but because no one on the field had been recognized as legitimate to carry it. The value was sitting dormant in profiles HQ could not even name.

The profiles flagged by the Potential Stack were formed into a network of in-field referents, on a volunteer basis, with a clear mandate and formal recognition. Their role was not to train but to translate — to show, using real cases from their own branch, what the tool actually changed in a route or a document workflow.

Headquarters had to give up some of its control. Field referents earned the right to adapt certain uses to local realities, in cases where the original rollout had imposed a uniform framework that was a poor fit for the diversity of branches.

Outcome

Within five months, field adoption on Front Line tasks rose from 15% to 53%. The optimization gain measured on the routes of branches with an active referent reached 18%, against negligible gains in branches without one.

Two unforeseen effects emerged. First, several of the field referents brought to light by the system were identified as serious candidates for steering roles they would never have reached through the classic path. The adoption program had surfaced an internal mobility pool. Second, the local adaptations designed by the referents fed back into the tools themselves — the field became a source of design rather than a mere recipient.

Lessons

1. An AI rollout steered solely from headquarters captures value where it is weakest. The value often sits in operations, which requires legitimate relays from within.
2. Applying the Potential Stack to frontline staff, not just managers, surfaces profiles able to carry adoption that the classic talent pools ignore.
3. Separating the zones where speed is dangerous from those where it creates value avoids imposing AI where caution has to remain the rule, and concentrates the effort where the gain is real.
4. Formally recognizing field relays and giving them room for local adaptation turns the recipients of a rollout into co-designers of it. The tool improves through the field that uses it.