Applied AI
What AI Actually Looked Like in a Fleet Office
The demos are frictionless and the pilots are triumphant. The reality of AI inside an institution is quieter, messier, and far more instructive — which is exactly why it is worth writing down.
I did not come to artificial intelligence as an engineer. I came to it as an operator with a backlog: reports that ate a day a week, policy questions buried in PDFs, vendor contracts nobody had time to read closely, and a standing suspicion that some of this work did not need a human at all. That is a very different starting point from the one most AI writing assumes, and it produces very different lessons.
Where it earned its place
The wins were unglamorous and real. Turning a month of raw fleet and financial data into a first-draft narrative for leadership — something that used to be a slow, manual assembly — became a matter of minutes to draft and then the usual care to verify. Asking plain-language questions of a stack of policy and contract documents, instead of hunting through them, quietly gave back hours. Summarizing, reformatting, sanity-checking a calculation, drafting the tedious first version of almost anything: this is where the technology paid rent.
Notice what those have in common. They are all tasks where a knowledgeable person still reads the output and owns the result. The AI compressed the distance to a first draft. It did not get the final say. In an operation where a wrong number reaches a vice president, that distinction is not a nicety — it is the whole design.
Where it did not belong
It was just as important to learn where not to point it. Anything that required accountability it could not carry, anything that touched sensitive institutional data without a clear boundary, anything where a confident-sounding wrong answer would be worse than a slow right one — those stayed human. The failure mode of these systems is not that they break loudly. It is that they are fluent and wrong, which is the most expensive kind of wrong in an institution that runs on trust.
The first principle is that you must not fool yourself — and a system built to sound convincing is very good at helping you.
The skeptic's way in
The cynic in me is the reason the AI was useful rather than dangerous. I treated every output as a claim to be checked, not an answer to be trusted. I asked what it would cost the organization if this were wrong, and I let that number decide how much verification the task deserved. I kept the human in the loop precisely where the human mattered. None of that is anti-technology. It is the same discipline I brought to a telematics feed or a maintenance invoice, applied to a new instrument.
That, in the end, is what applied AI actually looked like in a fleet office: not a transformation, but a capable, untrustworthy, genuinely useful new tool that rewarded the people willing to ask it better questions and check its work — and quietly punished the ones who took it at its word.