Logistics: a control tower that can only watch

The shape of the problem

A mixed freight network runs thousands of movements a day across road, rail and air, with a control tower of dashboards and controllers. Operators typically invest heavily in visibility, and it works — when something goes wrong, somebody usually sees it within minutes.

Seeing it is not the constraint. Between a controller spotting a problem and the network actually changing sit several phone calls, two systems, and a driver who has already left. The gap between exception and resolution routinely exceeds the window in which a resolution would have helped.

Why the obvious fix does not work

The control tower is an observation deck attached to a manual process. Every dashboard improvement makes the gap more visible without making it smaller.

There is also a volume problem nobody says out loud. With hundreds of exceptions a day and a fixed number of controllers, the tower can give each exception only minutes of attention if it does nothing else. In practice controllers triage by instinct, and the exceptions that get attention are the loud ones rather than the expensive ones.

The design

Exception classification against actual cost

Every exception is scored on what it will cost if nothing changes — contractual penalty, downstream connection risk, customer tier, recovery options remaining. That score, not arrival order or volume, orders the queue.

Expect the first week to be uncomfortable. A class of exception the tower has always treated as urgent often turns out to be cheap, and a quiet category turns out to be the most expensive thing in the network.

Agents that resolve within a defined envelope

The re-planning agent holds tool access to the transport management system, the rostering platform and carrier booking APIs. Inside a defined envelope — cost delta under a threshold, no service-level breach, no driver-hours violation, no hazardous-goods implication — it resolves autonomously and logs what it did.

Outside the envelope it assembles the options and puts them in front of a controller with the trade-offs costed.

Compliance as hard constraints

Driver hours, vehicle restrictions and dangerous-goods rules are deterministic checks, not learned behaviour and not model judgement. An option that violates one is never presented, never mind executed. This is the boundary that makes transport compliance willing to sign off.

What to measure afterwards

• Share of exceptions resolved with no human touch, and median exception-to-resolution against the useful window.
• On-time delivery, network-wide.
• Empty running kilometres — continuous re-planning takes repositioning opportunities when they appear rather than at shift boundaries.
• Quarterly review of which exception classes should move in or out of autonomous handling, based on the traces.

Where it runs

The operator's cloud tenancy, with dedicated endpoints for the latency-sensitive path. Every autonomous resolution carries a full trace — the exception, the cost model output, the options considered, the constraint checks, the tool calls made and the outcome.

That trace store tends to be worth more than expected outside operations. Insurers ask for it.