AI and LLMs in car transport

how and how not to use AI

Tobias Carlén · Axess Logistics · ECG

Introduction

this talk is on where LLMs and machine learning work well — and where they don't - it will cover

• real examples from Axess Logistics
• why it went like this
• a working approach for applying AI to real-world problems
  

subject: car dimensions

we have a major data quality problem with car dimensions and weight:

• many orders have incomplete or incorrect data
• OEM data is often generic, or based on older versions of the car
• manual entry errors, wrong unit of measure, data in the wrong fields…. etc etc etc

Approach: let AI fix it

train a model on a large set of correct vehicles,  let AI check and correct orders for us.

miserable results

• data quality improved slightly — but AI made lots of errors
• AI overwrote good data with bad
• AI results were far worse than an experienced human guesstimating from brand and model

why did this fail?

the models created false correlations. a human knows that color has no real correlation to weight — AI does not.
also, minor errors in the training creates strange decisions.

instead we built a tree model

at the top sits a model — say Volvo XC60

with SQL queries that catches even badly misnamed XC60s,

and fallback average values for all dimensions, if we have nothing better

transport models & inheritance

• each rule defines a transport model — a range of variants close enough to make no difference from a transport perspective
• each rule inherits everything it can from above — only what differs is specified
• values must fall within ranges — outliers get replaced by the transport model average
• identification varies: a piece of the model code, part of the name, something in the VIN — whatever we could find to make a rule
• external sources as cross-checks: a plate number gives us registry data, which for most cars is fairly accurate

now LLMs started making sense

with the tree model in place, improving it — fine-tuning rules, extending the rule set — turned out to be something an LLM can do!

in hours, precision and coverage of our ruleset improved dramatically!

improvement loop

the LLM never outputs dimensions directly — it works to improve the correction rules. A human checks all changes before production - Not all suggested changes are good!

same loop a human expert would run. minutes, not a training cycle.

the result

• after the correction engine runs, every car has data for every dimension
• >99% of them are exactly right
• on average 14% of orders are corrected by the current rule set

numbers from shadow-prd snapshot, 2026-06-09

why it worked

• AI doesn't have common sense — you must establish clear boundaries and show where true correlation exists
• a strict system like the tree model is easily understood by an LLM — it can work within the rules and suggest changes to them
• a big test data set (cars with known real dimensions) gives quick, exact evaluation: "did it improve predictions? did it destroy any data?"

finding and learning from mistakes is far more effective than trying to force the LLM not to make mistakes.

use LLMs to maintain the structure,
not to be the structure

1. define the boundaries — with specific, hierarchical examples: a passenger car weighs 900–3500 kg; an average car ~1700 kg; a compact car 1200–1500 kg
2. point out valid correlations — drive train impacts weight significantly
3. point out false correlations — color has no correlation to weight
4. define a system for AI to work within — fine to let AI help build it, but nail the problem to a framework first
5. create a test data set — inputs and expected outputs, like any normal testing; fast discovery when something goes wrong
6. document all assumptions carefully — let AI do this for you! when something goes bad, check assumptions first

what LLMs do — and don't do — well

what this means for human skills

• Real experts become even more necessary - people who really understand how your business works are vital to implement AI
• System thinkers are also vital - people who can extract knowledge from experts and codify it, turn it into useable rules that systems, including AI, can understand
• Some people MAY be able to do both, but it's rare, and this is still something that needs to be done collaboratively; it turns out experts do things very differently, and it takes discussion and analysis to determine best practise 

leverage from this - an example

• Your best planner will only be able to plan a limited number of trucks and orders every day
• if all your best planners distill their best thinking and decision rules for doing things right, you can have a model that combines them and applies them tirelessly and rigorously, to ALL trucks and ALL orders
• Planners can shift their focus from planning trucks to tuning the rules, handling special cases as they come up and thinking about how to handle similar cases automatically

AI is a great tool — but it's just part of the right toolkit

LLMs are trained on large volumes of text. they are very good at interpreting language. they are not solution engines — design the structure of the solution first, then AI can feed and adjust it.

combine AI with deterministic models, SQL, optimization engines and human common sense to produce the best results.