AI in fleet software: what has actually shipped

Every fleet software vendor now has an AI story. That is not a criticism, it is an observation about a market responding to what buyers are asking about. But it does make the announcements hard to read, because the word covers everything from a working anomaly detector to a rebranded rules engine.

We have covered six of these since July. Here is the sort.

Category one: shipped features

Chevin's FleetWave dashboard, covered in issue 2, is the clearest example of something that exists and does something. A dashboard is a dashboard - the interesting question is whether the analysis inside it changes a decision that would otherwise have been made differently, and that is answerable by a buyer during a demo rather than by us from a release.

Chevin adding Lightfoot driver scoring, from issue 11, is the more instructive one and we will come back to it.

Category two: integrations wearing a new label

A great deal of what is being announced as AI is one vendor connecting to another vendor's existing capability. This is genuinely useful - most fleet operators are running four or five systems that do not talk to each other, and closing those gaps saves real time. It is just not a new capability, and pricing it as one is where buyers get caught.

The Chevin and Lightfoot pairing sits here. Lightfoot's behavioural scoring engine is an established product. Surfacing it inside FleetWave means an operator sees driving behaviour next to maintenance and cost data instead of in a separate login. That is worth paying for. It is not a new intelligence.

Real-time driver-to-dispatch communication improving fleet operations, from issue 7, is the same shape. Real-time messaging is not new. The value is in the routing and the record, not the model.

Category three: positioning

"The UK's first agentic AI expert for fleet insurance", reported in issue 4, is the announcement that most needs a buyer's scepticism. Agentic is currently the most-stretched word in enterprise software. It can mean a system that independently takes actions with real consequences, or it can mean a chat interface over a document set. The gap between those two is the whole purchase decision.

Central Dispatch enhancing AI-powered pricing intelligence (issue 5) and field service fleets using AI to solve their top three challenges (issue 8) are both vendor-authored framings of vendor products. That does not make the underlying tools bad. It makes the claims unverifiable from the outside.

The Chevin arc is the story

Look at the three Chevin items in sequence and something more interesting appears than any one of them.

Issue 2: launch of a new dashboard. Issue 7: the company publicly telling fleet operators to look beyond implementation - that buying the system is not the same as getting value from it. Issue 11: bolting a partner's driver scoring engine into the platform.

That middle step is unusual and, frankly, more honest than most vendor communication. A software company telling its market that implementation is where the value is won or lost is a company that has watched customers fail to get value and decided to say so. It is also a company positioning itself for a services conversation, which is the commercially sensible read.

Both things are true at once. The useful signal for a buyer is the same either way: whatever the AI does, the failure mode is not the model, it is adoption.

Five questions for the demo

If you are being sold AI in a fleet context this year, these separate the categories quickly.

1. What decision does this change? Not what does it show. What would a manager do differently on a Tuesday morning that they would not have done without it.
2. Is this your model or someone else's? Neither answer is wrong. Knowing which tells you who you are dependent on and what happens if that relationship ends.
3. What does it do when it is wrong? Every predictive system is wrong sometimes. Ask what the failure looks like, how you find out, and who is accountable.
4. What data does it need from us, and do we have it? Most fleet AI underperforms because the maintenance history is patchy and the mileage data is manual. The model is rarely the constraint.
5. What did your last three customers actually change? Ask for the operational change, not the ROI figure.

The honest position

There is real capability arriving in fleet software, particularly around predictive maintenance and cost anomaly detection, where the data is structured enough for it to work. There is also a substantial amount of relabelling.

Telling them apart from a press release is not possible, which is why we report what was announced and let the demo do the rest. But eleven weeks of coverage does suggest one pattern worth naming: the vendors making the most specific claims are making the smallest ones, and the vendors making the largest claims are the hardest to pin down on what the software does.

That is usually a reliable signal.