Writing for people who run vehicles for a living

A missed service interval looks cheap on the day it is missed. Nothing breaks. The vehicle runs its route, the driver reports nothing unusual, and the maintenance line in the monthly report stays flat. The cost arrives later, and it arrives compounded.

Take a 200-vehicle haulage fleet in Lagos. If four vehicles a month slip past their interval unnoticed — a rate most operators would consider unremarkable — the downstream pattern is consistent: accelerated component wear, a higher rate of roadside failure, and an average of nine additional days off the road per affected vehicle once the fault finally presents.

Nine days is the number worth sitting with. At a conservative daily contribution of ₦180,000 per truck, four vehicles a month puts roughly ₦78m a year of capacity into the gap between what the schedule said and what anyone verified.

The failure is rarely negligence. It is arithmetic. A fleet supervisor covering 200 vehicles across three depots cannot physically confirm the condition of each one, so the schedule becomes a claim rather than a record. Averages sit on top of that claim and make the fleet look healthy.

What changes the outcome is not more discipline; it is a shorter distance between observation and escalation. Field officers already see these vehicles. When their photographs are read and compared against the vehicle's own history, the fourth truck stops being invisible — and nine days becomes a scheduled afternoon in the bay.

Almost every fleet in this market starts on a spreadsheet, and for a long time it is the right tool. It is fast, everyone can read it, and it costs nothing. The question is not whether spreadsheets are bad. It is knowing the point at which the sheet starts costing more than it saves.

Four signals show up reliably. First, reconciliation time: when someone spends more than a few hours a week making the sheet agree with reality, the sheet has stopped being a record. Second, version drift — depot copies, WhatsApp screenshots, a version on someone's laptop that only they update.

Third, evidence gaps. When a client or insurer disputes vehicle condition and your answer depends on a photo somebody may still have on their phone, you do not have a record; you have a rumour with a timestamp.

Fourth, and most telling: when the answer to 'is this vehicle deployable right now?' requires a phone call. That is the moment the sheet has become an index of where information might be, rather than the information itself.

What to move first: condition and compliance status per vehicle. Not routing, not fuel, not driver scoring — those are usually fine in a sheet longer than people expect. Move the thing that has evidence attached to it and expiry dates that bite, and let the rest migrate when it is ready.

There is a version of this conversation where AI 'looks at a car and knows what is wrong with it'. That version is not useful to anyone running a fleet, so here is the accurate one.

A vision model works on patterns of pixels. Trained on enough labelled examples, it learns that a particular arrangement of shadow and highlight along a door panel means a crease; that flattened tread blocks with reduced groove depth mean wear approaching threshold; that a dark irregular patch beneath an engine bay with a specific sheen is likely a fluid trail rather than a puddle of rainwater.

It is very good at three things: consistency, coverage and comparison. It applies the same criteria to the eightieth vehicle of the day as the first, it examines every image rather than a sample, and it compares today's frame against the same vehicle's frame from last month — which is where most of the value actually sits. A dent already recorded is noise; a dent that appeared since Tuesday is a question.

It is weaker at novelty and at context. An unusual aftermarket fitting, an obscured plate, a badly lit night capture — these produce low-confidence outputs, and a system that hides that fact is doing you harm. Ours attaches a confidence level to every finding and routes anything below your threshold to a human before it becomes a work order.

The practical implication for operators: do not buy detection accuracy claims in the abstract. Ask what the model does when it is unsure, who reviews that queue, and how the reviewer's decision feeds back. That workflow, not the raw model, is what determines whether a mechanic ever receives a useful job.