AI for trucking companies: what works, what doesn't, and how to start

Not self-driving trucks and not a chatbot bolted onto dispatch. The useful version of AI for a carrier is quieter: it reads the records you already have, joins them, and tells you where the money went. Here is what that takes.

By Djuro Kolundzija, founder of Nerivo. Started his own trucking company in 2012 and grew it past 100 trucks.

What AI is good for in a trucking company

A fleet of 20 trucks produces more records than anyone reads, and a fleet of 500 produces twenty-five times as many: every stop, every gallon, every toll, every invoice, every shop ticket. The losses are in there, and they are not hidden on purpose. They are hidden because finding them means putting two records from two different systems side by side, for every load, every week. Nobody has that job.

AI is good at exactly that job, if it can reach the data. In practice it does three things well:

• Answers questions from your own records. “Which customer pays slowest?” “What did unit 214 cost last month, and why?” Asked in plain words, answered in dollars, with the rows behind the figure.
• Runs the same checks every night. Unbilled detention, short pays, repeat repairs, fuel that does not add up, cost-per-mile outliers, idling and speeding. The checks do not get tired or busy.
• Drafts the follow-up. The detention line for the invoice, the dispute email with the rate con terms quoted, the note to the driver with the numbers in it.

What it is not good for, yet: anything that needs judgement about a customer, a driver or a lane. It finds and counts. People decide.

Why “just use a chat assistant” falls short

Most owners who have tried AI started the obvious way: a general-purpose assistant, a subscription, and an attempt to point it at the TMS. It is worth doing once, because you learn the five walls fast.

1. Every platform connects differently, and most barely connect. ELDs have APIs. A TMS sometimes has an MCP connector an assistant can use directly, and even then a slim one. Maintenance software has an API at best. Fuel cards usually have neither: you download a statement and import it. Each source is its own bridge, and somebody has to build and keep it working.
2. Where a connector exists, it comes with limits. An MCP or API call returns a page at a time, fifty or a hundred rows, with rate limits on top. A month of fuel transactions for 40 trucks is two thousand rows; a quarter of loads is more. The assistant never sees the whole picture, so it answers from a sample and does not say so.
3. Nothing joins the data. The ELD knows the truck sat at the dock. The TMS knows the load and the customer. The invoice knows whether detention was billed. Three systems, three IDs for the same load, and the join is the entire finding. A chat window does not do joins.
4. It only works while someone is typing. There is no overnight run, no Monday morning list, no alert when a short pay lands. The person who asks the questions becomes the system.
5. It is priced per seat. The owner, the dispatcher, the safety manager and the bookkeeper each need their own subscription, and each of them pastes the same exports in separately. Four people, four copies of the data, four answers that do not quite match.

None of this is a reason to avoid AI. It is a reason to put a data layer underneath it: one place where every system is connected once, every record is matched to the right truck, driver and load, and every person in the office asks questions of the same numbers.

How to get started: the week-one plan

• Day 1 — connect what can be connected. The ELD or telematics has an API: a read-only key, ten minutes, and you have location, engine hours, idle and speed for every truck. If the TMS offers an MCP connector, add it; if the shop software has an API, add that.
• Day 1 — import the rest. Fuel card statements almost never connect; download the transaction file and import it. Same for loads, invoices and payments where the TMS has no connector, repair orders, and the pay register. A CSV per system is enough to start; a built integration can follow.
• Day 2 — have the agent match the records. Every fuel line to a truck and a load, every stop to a load, every repair to a truck. This is the step that general assistants skip, and it is where most of the value is. Expect a few hundred lines that match nothing; those are findings too (a card used on a truck that was not there).
• Day 3 — write the instructions down. Pick three questions, not thirty: the three that cost the most in your fleet, usually unbilled detention, short pays and cost per mile by truck. Then write each one up as a standing instruction for the agent (what to check, which records to join, what counts as a flag, who gets it) so it is saved as a task, not retyped every week.
• Day 4 — put it on a schedule. Whatever you use has to be able to run those tasks on its own: nightly or weekly, with the result sent to the inbox. If the tool can only answer while someone is typing, it is not finished.
• Day 5 — first report. The three answers, with the records, to the people who can act on them. Then the same report every week without anyone asking for it.
• Week 2 onward — widen. Add the checks you did not start with. Let the dispatcher and the safety manager ask their own questions of the same data. Measure what was recovered.

Six questions to ask on day one

Each of these needs two systems joined. That is why nobody was answering them before, and why they are the first things to ask.

What to look for in any AI tool for your fleet

• It shows the records. Every figure opens to the ELD times, the invoice, the fuel line that produced it. A number you cannot check is a number the customer will not accept either.
• It reads, it does not write. Read-only access to your systems. Nothing an AI does should be able to change a load, a rate or a pay line.
• It runs on its own. Nightly, with a report in the inbox, not a chat window waiting for a prompt.
• It is priced for the fleet, not the headcount. Everyone in the office on the same data for one fee.
• It is honest about missing data. A month with no maintenance records on file should say “none on file”, not “$0”.

Where Nerivo fits

Nerivo IQ is that data layer with nightly checks, driver scorecards, custom reports and the assistant built on top. It connects your ELD by API, takes fuel card, toll, TMS and shop data by import or connector, whichever each one allows (read-only; what we do not integrate with yet, we build within the first ten days), matches every line to the right truck, driver and load, runs the checks every night, and answers questions from the whole fleet’s data, not a page of it. One fee for the fleet, every manager on the same numbers, and every figure opens to its records.

Questions people ask

See also: Detention pay: rates, free time, and the loads nobody billed · Fleet cost per mile calculator