Calculating and explaining: why planning software can now explain how it reaches a plan

Anyone who works with planning software knows the feeling. The system makes a plan, and it is often a good one. But why is truck 12 not doing the afternoon round? It says nothing about that. For years, that was simply how it was. That is changing now, and it may well be the most important change AI brings to the planner.

The old problem: a good answer without an explanation

Optimisation models for planning have existed for decades. In a short time they search through huge numbers of combinations of trips, vehicles and drivers, and deliver a plan with a total score. What they did not deliver was a story.

If you wanted to know why an order was not planned, you had to reason it back yourself, or call the supplier. The result is something I often see in practice: planners do not fully trust the proposal, start moving things around by hand, and the software delivers less than promised.

The odd thing is that the information was there. The optimisation model "knows" exactly which rule is binding: a time window, a driving time, a payload. But that knowledge was hidden in numbers and log files, not in language a planner can use.

What has changed

Two things now come together.

1. Optimisation models record the why better. Modern solvers keep track, per rule, of how much it costs or how tight it is, and why an order was not planned.
2. Language models can read that record and retell it. A language model, the kind of AI behind ChatGPT and Claude, turns the numbers into plain language. It also works the other way round: a question such as "what if truck 7 leaves an hour later?" is translated into a new calculation. The optimisation model recalculates, the language model tells you what changed.

This is what a conversation with your planning system can look like (fictional example):

Planner: Why is truck 12 not doing the afternoon round?

System: After the morning round, truck 12 has 24% battery left. The afternoon round needs 31%, plus 10% reserve. Charging at the depot is only possible from 13:45, because the grid connection is full with two other trucks until then. Truck 12 would then miss the first customer's time window. Truck 7 can do the trip; that costs 38 euros more in kilometres.

Every number in that answer comes from the optimisation model. The language model only turns it into a story.

Who does what

Who	Does what
Optimisation model	Makes the plan, calculates costs, battery and driving times, checks the rules, runs scenarios
Language model	Understands the question, turns it into a calculation, explains the outcome, summarises for driver, customer or management
Planner	Decides, brings in knowledge that is not in the data, checks the explanation

Where it can go wrong

• The language model starts calculating itself. Then you get a convincing answer that is wrong. Insist that every number comes from the optimisation model.
• The explanation cannot be checked. You should be able to find a good explanation back on the screen: the binding rule, the time window, the battery level. If you cannot, it is a story, not an explanation.
• The explanation is only as good as the data. If time windows or payloads are wrong, the system will explain a wrong plan very convincingly.
• Your data. Ask which data goes to the language model and where that model runs. There are set-ups in which your company data never leaves your own system.

What this means for the planner

The planner no longer has to reason back how a proposal came about. But they do need to be able to judge whether the explanation is right, ask the right questions, and pass the story on to a driver or customer. Those are exactly the skills the profession is moving towards.

Read also: Which planning software can calculate and explain? and How to make your planning system explain itself: five kinds of questions. Earlier I wrote about what AI can and cannot do in transport planning and the ground rules for AI in planning.

On the AI day of the Certified E-Fleet Planner programme you work with an AI route planner yourself and practise assessing proposals and explanations.

Sources

• Li, B., Mellou, K., Zhang, B., Pathuri, J. & Menache, I. (2023). Large Language Models for Supply Chain Optimization. Microsoft Research.
• Timefold. Validating an optimized plan with Explainable AI.