The tracking copilot

Logistics · AI inside

The tracking copilot

Forty people, one playbook, and the calls that stopped

  • 2xteam productivity, as assessed by the business
  • ~50%less tracking effort
  • 40–50people on the team
The tracking copilotscreenshots and video coming

The problem

A central team of forty to fifty people tracked shipments by phone. The system raised an alert, a person called the driver, took the update, relayed it to whoever was waiting. Only alerted shipments were tracked; there was no proactive tracking because there was no time for it. Task allocation was manual and the playbooks lived in people’s heads.

What I decided

Treat the playbooks as the product, and treat the copilot as phase one of a longer roadmap. The AI fleet tracker was designed in phases: first a copilot that works alongside the team and, in doing so, trains a custom model on the problems trackers actually deal with and how they resolve them; then proactive tracking; then automation of the routine path. If every branch of “what do we do when a driver says X” could be written down and learned, the people could work the exceptions.

The process I designed

  1. Map every branch. The tracker flow and the exception-handling playbook were drawn out as a mind map: alert types, the questions to ask, the answers that end a call, the answers that escalate.
  2. Translate playbooks into AI-assisted workflows. Centralised shipment visibility and historical tracking intelligence, so the assistant knew what happened last time on this lane with this driver.
  3. Human-assisted deployment. People stayed in the loop for continuity. The copilot took the routine update path. Adoption came from the deployment model, not from the model.
  4. Aim at the exceptions. The design target was to remove about 60% of manual tracking so the team could work critical escalations and, later, track proactively.

What it changed

Team productivity roughly doubled and tracking effort fell by about half, as assessed by the business during the rollout. Dependency on manual task allocation went away. Proactive tracking was scoped as the next phase.

Decisions I would defend

  • Write the playbook before the prompt. The mind map was the real deliverable. Everything the assistant did well came from a branch somebody had argued over first.
  • Keep humans on the critical path. A copilot that handles routine updates and hands off exceptions is adopted. One that tries to handle everything is switched off.

Visual

A clean flow diagram redrawn from the Miro map: alert in, assistant questions, resolved or escalated, human picks up. Anonymised; no shipment or customer data.