Case 01 · Logistics

Northbank Logistics

A growing freight operator was losing hours every day to manual load planning and exception handling.

Client

Northbank Logistics

Sector

Logistics & freight

Headline

11 hrs  ·  Planner time saved each day

Time to ship

Seven weeks

The problem

Northbank Logistics is a mid-sized freight operator running daily multi-modal shipments through the Port of Liverpool and across the North West. Their planning team was losing roughly two-thirds of every working day to manual data entry — reading customs declarations, bills of lading, commercial invoices and gate-in paperwork, and pushing the structured fields into their TMS by hand. Volume was growing, the team was at capacity, and overtime spend was creeping up. Hiring more planners was the obvious move and the wrong one — the work was clerical, not judgement.

What we built

A document intelligence pipeline that reads inbound shipping paperwork — bills of lading, customs declarations, commercial invoices, packing lists, gate-in/gate-out photographs — and produces structured records pushed directly into the existing TMS. Confidence-scored extractions; anything below threshold is routed to a human reviewer with the source document and the model's reasoning shown side by side.

The build sequence:

  • Week 1: scoped the seven highest-volume document types with the planning team, agreed the fields per type, built an eval set of 200 real documents with the correct field extractions.
  • Weeks 2–3: built the extraction pipeline (vision-and-language model, retrieval grounding on Northbank's customer rules, structured output validation), achieved 96% first-pass accuracy on the eval set.
  • Weeks 4–5: integrated with the TMS via the existing CSV import endpoint, ran in shadow mode against live document flow, surfaced edge cases that needed refinement.
  • Weeks 6–7: live deployment with a single planner as the first reviewer, expanded to the full team in week 7 with monitoring dashboards in place.

What changed

Measured at the end of the third month in production:

  • 11 hours of planner time saved per day across the team — the headline number Northbank committed to before we started.
  • Throughput up 3.2× on document processing without adding headcount.
  • First-pass field accuracy stable at 96.5% after the second week of tuning; below-threshold documents (around 7% of volume) route to a human reviewer with the model's reasoning visible.
  • Two planning roles redeployed from data entry to exception handling and customer-facing work — higher-value, harder to automate.
  • Overtime spend down 22% in the planning team in the first quarter post-launch.

What we did not do

We did not replace the TMS or migrate it to an AI-native platform. Northbank's TMS is a working system that has to last another decade; the AI pipeline augments it via the existing import endpoints rather than rebuilding the core.

We also did not fully automate the customs declaration step. The volume runs through a human reviewer; the AI does the extraction and the suggested classification; the reviewer signs off. The regulatory risk on full automation is not worth the marginal time saving.

What is next

The same pipeline is now being extended to two adjacent document types — delivery notes and supplier statements — and the planning team is starting to redirect the saved time into proactive customer service. Northbank's planning lead is now the named owner of the system; we are on call for the next quarter, then on a retainer for ongoing tuning.

For the broader pattern: AI for logistics in Liverpool: from the Port to the warehouse floor.