Skip to content
Case study · Logistics

The hardest part
wasn't the AI.

We implemented AI at a logistics company. The hardest part wasn't the AI — it was understanding how the business actually ran.

71%

of exceptions resolve with no human touching them

29%

land in a queue with the context already attached

$4m+

saved

What we found

The company's real knowledge wasn't in the TMS.

It is almost never in the TMS. This is what we find at most operators, and it is the reason AI projects stall after the demo.

Inboxes

Threads where the real negotiation happened, invisible to every system of record.

Rate confirmation PDFs

Terms agreed per load, filed as documents nothing could read.

Check-call notes

Where a load actually was, written down in a form only a person could interpret.

One dispatcher's head

Which carriers actually show up. The single most valuable dataset in the company, stored in a person.

Years of operational data. Almost none of it usable by a machine.

How we did it

Structure first. Model second.

  1. 01

    We sat with dispatch.

    Before writing anything. Watching the work is the only way to find the interventions people have stopped noticing they make — those are invisible in any process document, because whoever wrote it had also stopped noticing.

  2. 02

    We traced one load, end to end.

    Tender to POD, marking every point a human had to step in. One load followed properly tells you more than a quarter of aggregate metrics, because it shows you the sequence rather than the totals.

  3. 03

    We built the unglamorous layer.

    A structured map of their entities — loads, carriers, lanes, exceptions — and how each relates to the others. No model involved. This is the part nobody wants to fund and the part everything else rests on.

  4. 04

    Only then did we put AI on top.

    By that point the agents had something to reason over: not a pile of documents, but a model of the business where a load knows its carrier, and a carrier knows its history.

AI doesn't create leverage. It compounds the structure you already have.

If your data is a mess, AI gives you faster mess.

This is why Deskwise is an ontology with agents on top, rather than a chatbot with an integration list. The model is the easy part — understand the business first.

Want the same trace of your operation?

A 30-minute audit: where agents would act, what they would be allowed to do without asking, and what the record of it looks like.

Figures are from a single engagement with one operator and are not presented as typical results. The client is not named; details are shared with their agreement.