Seven ways to bring AI into a fleet. Six give you another tool. One gives you the work done.
Telematics with AI, ChatGPT with a spreadsheet, old fleet software with an 'ask AI' button, hyperscaler platforms, internal builds, shadow AI - and a seventh way: you buy the finished work. How AI really enters fleets today, and which of these routes it took into yours.
82% of fleet operators report higher operating costs over the last year (Geotab, press release of 29 October 2026). Almost all of them already run some telematics, some fleet software and have someone who "tried ChatGPT on the lessor's spreadsheet". The dashboards knew about the rising costs for a long time. It did not change the fact that service invoices are still re-typed by hand.
When I talk to people who run fleets, AI is almost always already in the company. The question is not "whether", but "which way it came in" and whether anyone is in control of it. I counted: there are seven routes. Six of them leave you with another tool to operate. One leaves you with the work done.
1. Telematics platforms with AI
Geotab, Samsara, Webfleet and dozens of smaller vendors. The data from the vehicle is there: location, mileage, driving style, fault codes, increasingly battery health. The models are there: driver scoring, failure prediction, event detection. The problem is what stands at the end of that chain: a screen.
The screen shows that vehicle 47 has 40% more mileage than the contract assumed. It will not call the lessor, recalculate the buy-out or ask the driver whether the route has changed. A person does that, if they have time to look at the screen. This year's SambaSafety survey (740 fleets, insurers and brokers) says it plainly: 53% of fleets struggle to interpret or act on their telematics data, even though almost half of them have been collecting it for at least six years. The data was never the problem. What happens to it afterwards is.
2. General models with a fleet layer
ChatGPT, Claude, Gemini, Copilot. You paste the lessor's statement, ask which cars are the most expensive to run and get a sensible answer in thirty seconds. Someone in your company already does this, even if you do not know it.
It is a great tool for the first analysis. By the twentieth identical analysis, every month, for four hundred vehicles, with files in four formats from three lessors, it stops working. The model does not remember that last month you agreed the tyre invoice goes to the branch cost centre, not head office. Every conversation starts from zero, and the responsibility for the result stays with the person who pasted the file.
3. Old fleet software with an "ask AI" button
Vehicle register, fuel cards, claims log - the same programme the company has had for eight years, now with a new button in the corner of the screen. The button answers questions about data the system already had.
A fair question to ask the vendor: is AI the core of this product, or a feature meant to defend the licence renewal? The answer shows in one thing: can the system do something it could not do before - for example split a consolidated invoice per vehicle on its own and flag the lines that need checking - or does it just answer faster about what is already in it?
4. Hyperscaler platforms
Copilot Studio, Gemini Enterprise, Bedrock AgentCore. IT was told to "do something with AI" and built an agent on fuel-card data. It works, it reports, it has an owner in IT.
If that is the case in your company, your fleet already has an AI programme. The person who manages that fleet may not know. This is not a complaint about IT. It is a picture of a tool built far away from the problem it was meant to solve, where nobody asked the fleet manager what really eats their Thursday afternoon.
5. Internal builds
The company has a capable developer or an external software house that built a fleet agent in three months. Building is easy today - easier than ever. Keeping it running is not.
Who fixes the agent when the person who wrote it leaves? Who reacts when the lessor changes the settlement file format and the insurer changes the claims report layout? Who is accountable when the agent assigns a fine to the wrong driver and the letter reaches the wrong manager? An internal build turns an operational problem into a software-maintenance problem, and that is rarely the core competence of a company that runs a fleet rather than builds software.
6. Shadow AI
The cheapest and most common route. A driver list pasted into a public chatbot because "it's faster". Together with plate numbers, home addresses and policy numbers. A scan of the registration document uploaded from a phone to "pull out the VIN".
Nobody approved it, nobody knows where the data went, and officially the company "does not use AI". This is not a technology problem. It is a signal that people have work to do and reach for the first tool that takes it off their desk - even at a risk they cannot see.
Most fleets use four or five of these routes at once
This is not a choice of "one out of seven". A typical mid-sized company fleet has telematics from one vendor, fleet software from another, someone with ChatGPT in a browser tab, a pilot project in IT and a few people quietly using AI on their own phones. Each route has a licence, an owner and a learning curve. None of them takes the work off the desk.
The work always looks the same: a consolidated service invoice with 60 lines to split per vehicle and cost centre; a fleet policy where you must check that every replaced car is covered; a fuel-card statement where you must find the fill-ups that do not match the mileage; a fine where you must name the driver within seven days. None of the six tools above does this end to end. Each of them helps the person who does.
7. You buy the finished work
The seventh route looks different. You send the service invoice, the policy, the fuel-card statement. You get the document back checked and split per vehicle: lines assigned to vehicles and cost centres, deviations flagged, the obvious things done. You get a question only where a person has to decide - for example when the invoice concerns a car that, according to the register, was returned to the lessor two weeks ago.
The models, the tools and the risk of maintaining them sit with the vendor. You pay for an account, not for an implementation project, and you do not need to know which model runs underneath or who fixes it when the lessor changes the file format.
That is how mAIster, the AI agent in VivaDrive, works. It is not a button in an old system or a dashboard with one more chart. It is an agent that receives a document and returns a result, and every step it took stays in the log - who allowed it, what it did, what came out of it. Because "human in the loop" is not enough if nobody knows where in the loop that human stands.
Which of these routes did AI take into your fleet?
Count honestly. If the answer is more than two, you are probably paying for several tools while the work is still done by the same person as three years ago. The seventh route can be tested on your own data: vivadrive.ai/pl/rejestracja.
Sources: Geotab, press release of 29 October 2026 (survey of fleet operators: 82% report higher operating costs, 31% name fuel as the main cause); SambaSafety, fourth annual telematics report (740 fleets, insurers and brokers; 53% of fleets struggle to interpret or act on telematics data; 46% have used telematics for at least six years) - via Insurance Thought Leadership, 13 October 2026.