What AI actually changes in fleet software

Every category is being told it needs AI and most of them are being shown the same demo, which is a chat box bolted to the corner of something that already exists. In fleet and telematics software, where I have spent the last two years, that demo does almost nothing. The version that would actually help is dull enough that nobody puts it in a launch video.

A chat box is the wrong shape for this job

The operator watching a live map of thousands of vehicles is not short of ways to ask a question. She is short of attention.

Everything worth building on those products took a step away from her. Rotating the map to the vehicle's heading so nobody has to translate north into left. Fitting a whole route into one view instead of hunting for it. Replaying a trip rather than reading a table of coordinates. A chat interface asks her to compose a query, wait, and read a paragraph. That is a step back in.

What would help is the product noticing first

The useful thing is surfacing what she would have found an hour later anyway. The vehicle idling somewhere vehicles do not idle. The route deviation that is going to turn into a late delivery. The alarm that has fired eleven times this week and never meant anything.

That work is ranking. Deciding what belongs at the top of a list currently sorted by time.

Which is also where the risk sits, because ranking means something goes below the fold. A product that quietly buries the one alert that mattered has done more harm than the version that never ranked anything.

You need the states before you need the model

None of it works until a vehicle is in exactly one of a known set of states. Moving, stopped, idle, delayed, offline. Five words, with the lines between them drawn once and in one place.

Every team I have watched try something clever on top of location data runs into this early. The data says where a vehicle is. The product needs to say what it is doing. Until those five words mean one thing each across the map, the alerts, the trip views and the compliance reports, a model sitting on top has nothing stable to be right about.

The prerequisite for AI in an operational product is a data model that somebody argued about for two weeks.

Where it already pays, without the label

Three places, none of them exciting. Cutting alert volume so the remaining alerts get read. Turning free text in tickets into the structured fields nobody fills in by hand. And extraction, which is what I worked on before I got into fleet software, pulling numbers off documents that arrive as photographs of documents.

That last one has been working quietly for years and has never needed a conversation interface.

The scale question

At two hundred thousand vehicles, anything you run per vehicle you are running two hundred thousand times. Cost and latency stop being infrastructure concerns and start deciding what the feature can be, the same way rendering cost did. Clustering and lazy rendering were not visual preferences on those maps. Without them nothing was usable at all.

Any AI feature in this category has that waiting for it. Better to find it in the requirements than in the first invoice.

What I would measure

How long it takes an operator to find the one vehicle that needs her, before and after. What share of the things you surface end in an action instead of a dismissal.

And the uncomfortable one: how often something that mattered got ranked below something that did not. If you are not measuring that, you have not really shipped a ranking feature.