FLEETDRIVE
Telematics is only part of the picture
The telematics providers deserve credit. The leaders in the space now say plainly that fragmented data and incomplete asset records are what stop AI from delivering. They are right.
They are also only one source among many. Berg Insight puts telematics penetration of business fleet vehicles in Australia and New Zealand at 27 % in 2023, rising to 40 % by 2028. On a fitted vehicle, the device records location, odometer, driving behaviour and fault codes. It does not see the lessor invoice, the warranty claim, the tyre bill, the infringement notice or the cost centre. Explaining why one depot costs 30 % more per vehicle than another requires all of those sources, joined on the same vehicle.
One vehicle. Five different records
This year we worked through the data of a multinational fleet of about 10,000 vehicles across four countries: five sources, five file formats, three currencies, three languages. Damage appeared under eleven different labels. Only two of the five sources carried a VIN. The finance file matched all invoices. Reconciling invoices is not the same as being able to attribute and compare costs by vehicle.
The work that followed was not AI. Every source was landed in one governed store in the customer’ s own cloud, which the data industry calls a lakehouse. Identity was resolved first, so that a lessor’ s asset number, a telematics ID and the registration on a spreadsheet pointed to one vehicle. Meaning came next: 45 raw charge labels were mapped to eleven concepts a fleet manager would recognise. That map, the ontology, gives the model consistent definitions to work from, so an answer can be traced back to the records that produced it. Within a week, the fleet company could see which fuel and energy spend could be tied to a vehicle, 43 % in one market, which could not, and which lessor files were missing the make, model and fuel type an emissions report depends on.
When data. world tested GPT-4 on business questions over a company database, it answered correctly 16.7 % of the time. Given a knowledge graph of the same data, the rate tripled.
Less reconciliation. Better decisions
Alphabet’ s 2025 survey of 740 European fleet managers found only 27 % could accurately quantify fleet CO2 and 26 % tracked it in Excel; with fuel, lessor and vehicle data joined, the figure comes from the record.
Some fleet management organisations here have begun offering clients exactly this: a question in plain English, a ranked answer with the figures behind it. The people who built those services will tell you the chat window took weeks and the data behind it much longer. That work is starting to be offered as a service: send in the exports, get back one record per vehicle and a list of what is missing. Require the system to flag missing data, show its sources and withhold answers it cannot support. That gap list tells the fleet what to connect next.
Start with the questions
Before buying the AI, write down three questions you cannot answer today and list every system the answer would need. You do not need to connect everything before you start. You need trustworthy data for the question you are asking. Ask your lessor and telematics provider for a complete, exportable record of your own vehicles, regardless of who hosts it, and know where it goes once it leaves. Require every figure to carry its source. Then put the model on top.
It is the less visible work that makes AI useful and its answers easier to check.
“ The vehicles are connected and the intelligence is not. Join the records first. Until then the AI is answering from whichever file it happened to be given, and it will not tell you which one.”
Barry Pryce is CTO and Co-Founder of FleetGuru. ai.
Trusted by leading fleet companies and 12,000 + service providers across Australia and New Zealand, FleetGuru simplifies bookings, approvals, payments, and your fleet’ s data into one intelligent platform. Smarter fleets, everywhere.
ISSUE 61 OCTOBER 2026 / WWW. AFMA. ORG. AU 23