Export your fuel-card transactions and your vehicle location history for the same month, upload both to ChatGPT or Claude, and ask it to flag fills where the vehicle was not near the station, the litres exceed the tank, or the fuel type is wrong. Then check every flag with the driver and paperwork before acting.
The match is the whole idea. A card statement tells you who bought what and where; a tracker tells you where the van was. When the two disagree, something is worth asking about. What the assistant cannot tell you is why. A van fuelled by a colleague while the owner's own was in the workshop looks the same in the data as a card lent to a friend, and only a conversation separates them.
The two exports and what each must contain
You need a fuel-card export and a location export, and ideally a short vehicle sheet. Ask your fuel-card provider and your telematics or tracker supplier for CSV downloads for the calendar month, not PDF statements.
| File | Columns worth having | Why |
|---|---|---|
| Fuel-card transactions | card_id, vehicle_reg, date_time, merchant_name, merchant_lat, merchant_lon (or address), fuel_type, litres, amount, odometer if entered | The purchase to be tested |
| Vehicle locations | vehicle_reg, date_time, lat, lon, ignition_on, odometer | Where the vehicle was at that moment |
| Vehicle sheet | vehicle_reg, fuel_type, tank_litres, usual_driver, home_depot | Rules need tank size and fuel type |
Two practical problems appear at once. Time zones and clocks: card times are often in the provider's system time and tracker times in UTC, so a fill appears an hour out. And merchant locations: many card exports give an address or a station name, not coordinates. If you only have addresses, ask the assistant to work at town or area level and say so, or add coordinates for your ten regular stations by hand. Anything that depends on exact distance between a station and the vehicle needs coordinates.
Before uploading, remove driver names and keep registrations and card IDs, which you can map back on your own computer. Check your plan's data controls: business plans from OpenAI and Anthropic do not train on your content by default, and on ChatGPT's individual plans you switch off "Improve the model for everyone" under Data controls. Our telematics audit tutorial uses the same preparation for yard time and tolls.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA worked example: a six-van courier firm in one month
Everything below is illustrative. A courier firm runs six vans, each with a fuel card. March has 214 fuel transactions and the tracker records a position every minute while the ignition is on. The owner uploads both files and the vehicle sheet. The first message sets rules and asks for a reading test before any analysis:
I have three files: fuel_march.csv (214 rows), locations_march.csv
(position every minute while the ignition is on) and
vehicles.csv (six vans, fuel type, tank size).
First, tell me: rows read in each file, the date range of each,
the count of fuel transactions per vehicle, and whether
timestamps in the two files look to be in the same time zone
(compare a transaction to the nearest ignition-on record).
Do not flag anything yet. List any problems with the data.The reply shows 214 transactions, six vehicles and a likely one-hour offset between the files, because every fuel time sits an hour later than the nearest stop in the tracker. The owner confirms with the card provider that fuel times are in local time and tracker times are in UTC, and asks the assistant to correct the tracker times by the offset. Without this step, the next stage would have flagged half the transactions as "vehicle not at station".
Now the flag rules. State them explicitly rather than asking for "anything suspicious":
Apply these five rules and list each flagged transaction with
the rule number, vehicle, date and time, litres, amount and
the evidence:
1. Location: the vehicle's nearest tracker position within 10
minutes of the transaction is more than 1 km from the
merchant.
2. Capacity: litres exceed the vehicle's tank size, or exceed 95%
of it twice in one day.
3. Fuel type: transaction fuel type differs from the vehicle sheet.
4. Timing: ignition was off for the whole 60 minutes either side
of the transaction and the vehicle was at its home depot.
5. Frequency: more than two transactions on one card within 24
hours.
Then give a summary count by rule and vehicle. Show the
distance and the tank figure used for each flag.A sample reply, with invented numbers, shows 17 flags: nine under rule 1, three under rule 2, one under rule 3, two under rule 4 and two under rule 5, with one transaction triggering two rules. The summary says van 4 has seven of the 17. That is the first thing worth noting, and also the first thing to resist concluding from. Seven flags on one van is a pattern to look into; it does not say who did what.
Reading the flags: what each rule usually means
Most flags have ordinary explanations, and a short key makes the follow-up fair and quick. For each rule, the likely innocent causes and the check:
| Rule | Innocent causes to rule out first | Check with |
|---|---|---|
| 1. Vehicle far from the station | Tracker dropout or missing positions; clock offset; a van swapped for the day; a colleague filling a different van or a spare can; an address-only station geocoded wrongly | Gaps in the tracker file; the day's job sheet; the driver |
| 2. Litres above tank size | Filling a jerry can or a second vehicle; a refrigeration unit or equipment tank on the same card; a wrong tank size on the sheet | Vehicle spec; receipt; whether the van carries a fuel-powered unit |
| 3. Wrong fuel type | A mis-keyed fuel code at the pump; a van changed but the sheet was not updated | Receipt; vehicle sheet |
| 4. Ignition off at depot | A fill for a van stored at the depot, with the ignition off between moves; staff filling for the next shift | Depot rota; who held the card |
| 5. Many fills in a day | Genuinely long routes; a shared card; a failed first transaction retried | Route sheet; card statement for declined items |
Take the nine rule 1 flags in the example. Four sit within ten minutes of a missing stretch in the tracker data, so the assistant could not place the vehicle at all; those go to "data gap, not misuse" once the supplier confirms the tracker was offline. Two are from the same day when van 4 was covered by the spare van; the job sheet confirms it. That leaves three. One of them is a fill 6.2 km from the nearest tracker position at a station the van had passed 20 minutes earlier, which fits a driver who went back for cheaper fuel and then rejoined the route. Only two flags remain unexplained, both on van 4, and those are worth a calm conversation.
Checking the assistant's work before you speak to anyone
A wrong flag creates a bad conversation, so validate the output first. Pick three flags and three non-flags and check each against the raw files. A reliable routine:
- Find the transaction in the raw card file. Confirm litres, time and merchant match what the assistant reported.
- Open the tracker rows around that time. Is the position gap real, or did the assistant measure to a position 40 minutes away?
- Recompute the distance for one flag. A map or a spreadsheet distance formula between the coordinates will do. If you get a different figure, find out why before trusting the list.
- Check a non-flag. Choose a normal fill and confirm the vehicle was at the station. If it was not, the rules are too lax or the join is wrong.
- Count rows. Transactions in the flagged list plus transactions not flagged must equal 214.
The most likely technical failure is a wrong join. The card file keys on card_id and the tracker on registration; if the card-to-vehicle mapping in the sheet is wrong or out of date, every flag for that van is nonsense. Ask the assistant to list the card-to-vehicle pairs it used and check them against your own records. If a driver has moved between vans mid-month, the mapping changes during the month, and the safe approach is to match by date range.
Talking to the driver: flag, then verify
The order matters. Do not start with an accusation, and do not send the spreadsheet. Start with a neutral, specific question about one transaction. Here is a filled-in message you can adapt (the details are illustrative):
Hi [first name], I'm checking last month's fuel records and
want to make sure the data is right. On 12 March at 14:20,
card 4 shows a fill of 62 litres at the station on the
ring road, and the tracker shows the van elsewhere at that
time. Could you tell me what you remember about that day?
It may be a tracker or card-time issue. Thanks, [your name]Ask the assistant to draft this message, but give it only the single transaction and tell it to be neutral. Do not paste the whole flag list into a draft that goes to a driver. The outcomes should be recorded: explained and closed; data error corrected at source; needs follow-up; or escalated under your disciplinary process. Escalation is not an AI decision, and it should only follow a process you have written down and applied consistently. If you are unsure of the rules where you operate on handling location data and employee discipline, get advice from a qualified employment adviser first.
Setting thresholds that fit your fleet
The five rules above are a starting set. Calibrate them against a month you already know. Run the rules on a month in which you are confident there was no misuse, or on one vehicle whose driver you trust, and see how many flags appear. If a clean month throws up 15 flags, the rules are too sensitive for your data quality, and you will waste time. Adjust: widen the distance tolerance to 2 km, extend the time window to 15 minutes, or exclude vehicles with known tracker gaps. If a month with a known problem throws up no flags, the rules are too lax.
Two refinements are worth adding after the first month. First, a fuel-economy check: litres per 100 km by vehicle by week, computed from odometer or tracker distance. A van whose economy suddenly worsens by 20% is a lead on fuel use or on a mechanical fault, and it does not depend on any merchant location. Second, a round-sum check: a run of identical round amounts such as exactly 50.00 may mean a card used for non-fuel items at a station shop. Neither is proof of anything, and both have innocent causes. A sudden economy drop can mean a dragging brake, a roof box or tyre pressure, which is why the fleet admin tutorial pairs fuel with service and vehicle checks.
A second example: fuel economy as a quieter check
Location matching catches fills in the wrong place. It misses the case where the fill is in the right place but the fuel does not go into the van, for example a driver filling a second vehicle at the same station. Fuel economy catches that, because the litres bought no longer square with the distance driven. Using the same six-van month, ask:
For each van, add up litres bought and kilometres driven
(odometer at last tracker point minus first). Show litres per
100 km for each of the four weeks and for the month, and the
fleet average. List any van-week more than 20% above that
van's own month average. Show the figures you used.Illustrative output: vans 1, 2, 3, 5 and 6 sit between 11.2 and 12.6 litres per 100 km in every week. Van 4 shows 11.8, 12.1, 16.9 and 17.4. Two weeks at around 17 are a 40% jump. Now there are two independent signals on one van: the earlier location flags and a fuel economy step. Still no proof, but a conversation with more to go on. The explanations to rule out first are mechanical (a fault, a tyre pressure problem, a roof rack), operational (a change in route, more stop-start city work, heavier loads) and data (an odometer that stopped updating, so the distance is understated). Ask the workshop for the service record before you talk to anyone, and check that the odometer series in the tracker is continuous.
A quick sum shows why this matters. If van 4 drives 1,800 km in a month and should use 12 litres per 100 km, that is 216 litres. At 17 litres per 100 km it is 306 litres, a 90-litre gap. At a pump price of $1.60 a litre the gap is $144 a month, or about $1,700 a year for one van. Numbers like that are why a check is worth setting up, and also why a fair process matters: a $144 difference could be a brake dragging, and fixing it benefits everybody.
Prompts that go wrong and how they show up
Three failures to recognise, each with a fix.
The vague ask. "Find any suspicious fuel transactions" returns a list that mixes sensible and silly items: a large fill the day before a long trip, a Sunday fill, a transaction at a motorway station. The assistant has no rule, so it uses a general sense of unusual, and you cannot reproduce it next month. The fix is the numbered rule list shown earlier, kept in a Project so it is identical each month.
The time-zone slip. A reply reports that 80% of transactions occurred when the vehicle was elsewhere. This looks like a catastrophe until you notice that the median gap is exactly 60 minutes. The fix is the reading test: ask the assistant to compare timestamps first and tell you if there is a consistent offset, then correct and rerun.
The confident cause. The reply ends with "this suggests driver 4 is selling fuel". The data does not say that. Add a line to your rules: "describe what the data shows; do not speculate about motive or name a person". Then ask for the explanations that would also fit the pattern, so you see the innocent ones as often as the guilty ones.
A realistic before and after on the same flag: the first run says "van 2 fuelled at a station 38 km from its recorded position; possible misuse". After the reading test and the join check, the rerun says "no tracker positions for van 2 between 13:05 and 15:40; cannot assess". The second answer is less dramatic and correct, and it points to a call to the tracker supplier instead of the driver.
Using the result to improve controls, not only catch people
A fuel audit that only hunts for misuse misses the larger gains. Once the check is routine, the same data tells you which stations your drivers use and what they cost, so you can steer them to a cheaper supplier on their usual routes. It shows which vans are fuel-hungry and due a service, and which cards are unused. It also shows what controls to add: a PIN per driver, a rule that the odometer must be entered at the pump, card limits by vehicle, and a monthly report to the driver of their own fills. Controls that make misuse hard and honest mistakes obvious reduce flags more than any investigation.
One more small example. After three months, the log shows that 11 of 38 flags came from one station that the tracker consistently places 1.4 km away, because the merchant coordinates in the card file point at the head office, not the forecourt. Correcting that one coordinate removes a third of the false alarms at a stroke. This is the kind of improvement that only appears once you keep the log, and it is why the monthly record matters more than any one run.
What to keep, and what to leave out of the prompt
Keep a monthly log: the date of the run, the files used, the rules, the number of flags, and the outcome of each. After three months, ask the assistant to summarise the log: which rules produce the most false alarms and whether the same vehicle keeps appearing. Delete the uploaded files from the assistant once the run is complete, according to the settings of your plan, and store the exports in your own system. Do not put full names, home addresses or private journeys into the prompt. If a driver is permitted to use a vehicle privately, exclude those hours from the location rules, otherwise you will flag a fill on a Sunday at the driver's own station.
Plan for the day the flag is right. You need a written process before it happens: who talks to the driver, what evidence is gathered, who decides and what the consequences are. An assistant can draft the template for that process, but a person who understands your employment obligations should approve it. The same applies to any insurer or finance provider that needs fuel records for a claim.
If you want to see what else a small fleet can automate once the fuel check is running, AI route planning for small fleets addresses the other big fuel lever, and the audit checklist shows how to keep compliance records tidy. If you would like help building the exports, rules and follow-up steps around your own cards and trackers, that is a natural fit for an AI implementation consultation.
Questions fleet owners ask about fuel-card audits with AI
Can AI prove a driver misused a fuel card?
No. It can flag transactions that do not fit the vehicle's location, tank size or schedule. A flag is a reason to ask a question, not evidence. Many flags have innocent causes: a swapped vehicle, a clock mismatch, a shared card or a location gap. Confirm with the driver and the paperwork before anything else.
What data do I need to start?
Two exports. The fuel-card transactions: date, time, card or vehicle, litres, amount, fuel type and merchant location. And a vehicle location history from telematics or a tracker: vehicle, timestamp and position, ideally also odometer. Tank capacity per vehicle helps, and a driver-to-vehicle roster helps more.
Do I have to tell drivers I am checking fuel records?
Tell them. Most workplaces should have a written fuel and tracking policy that drivers have seen, and data-protection rules may limit how location data is used. Take advice from an employment adviser for your situation. A policy known in advance also reduces disputes when a flag turns out to be an error.
How often should I run the check?
Monthly is enough for most small fleets, with a quick weekly look at the top five flags if you have many vehicles. Run it on the same day each month so the figures are comparable, and keep each month's flagged list and outcome in a log.
Further reads
- AI Dash Cams for Fleets: Cost, Privacy and Driver Consent — Dash cam costs and driver consent if you add video.
- Route Planner for Delivery Drivers: Four Apps Compared on Cost and AI — Four route planners compared on cost and AI.
- AI Software for Trucking Companies: What to Check for a Small Fleet — What to check in AI software for a small trucking fleet.
- Rental Car Companies Using AI: Seven Uses a Small Fleet Can Copy — Seven AI uses a small rental fleet can copy.
- Trucking Price Quote Template: Build One and Check Rates With AI — A filled-in trucking quote template with a cost-per-mile build-up, deadhead and fuel lines, one worked lane and an AI rate check with sample output.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: general fleet fuel-control practice; OpenAI's data controls help page and business data privacy page; Anthropic's file upload and model training help pages, checked 5 November 2026. All vehicles, drivers, figures and transactions are invented for illustration.