Artificial intelligence in freight is simultaneously overhyped and quietly delivering - the same industry contains vendors promising autonomous everything and fleets getting real value from machine learning in their planning office. The Future of Freight plan put data and technology at the centre of the industry’s direction; the operator’s practical question is narrower: which AI applications pay their way today, in a real fleet, with the data a real fleet actually has.
The honest current list is shorter than the conference version. Forecasting: machine learning predicting volumes, arrival times and delays from historical patterns - genuinely useful where the data is clean, and the ETA prediction case is the most mature: customers served with accurate windows beat customers promised optimistic ones. Route and load optimisation: algorithms sequencing drops and matching loads to vehicles at a scale beyond manual planning - the optimisation that was operations research for decades, now sharper with learned travel times. Document processing: reading PODs, invoices and consignment notes - extracting the data without the typing, where the error rate is low enough that exception-handling replaces data entry. Predictive maintenance: learning the failure signatures in telematics and inspection data - flagging the vehicle drifting toward failure before the roadside does. And pricing and tender analysis: modelling the cost of prospective work against the network’s real shape.
The common thread is unglamorous: these are prediction and pattern tools applied to data the fleet already generates - not robots, not autonomy, just better arithmetic on the operation’s own history.
The warning list saves more money than the opportunity list makes. Data-hungry systems sold to fleets whose data is a mess: AI amplifies the input quality - a forecasting tool fed garbage produces confident garbage, and the unglamorous data-cleaning project is the prerequisite nobody sells. Autonomy timelines: driverless trucks will matter eventually, but planning a fleet around a date is not strategy - the regulatory and operational path is long, and the driver shortage is now. Black-box decisions in regulated areas: anything touching hours, safety or compliance needs explainable logic, because the operator answers to examiners who do not accept ‘the model said so’. And the integration mirage: the AI tool whose value depends on feeds from three systems you have not connected yet is a project wearing a product’s clothes.
The working approach for a real fleet: start where the data is good - telematics-fed problems like maintenance prediction and fuel analysis beat document AI fed on chaotic scans; buy the application, not the platform promise - a tool that solves one named problem with measurable output; run it alongside the current process until it proves itself - the parallel run is the only honest evaluation; and keep the human in the decision loop - AI that recommends and humans who decide is the configuration that survives contact with reality.
Useful AI in freight today is prediction and pattern-matching on the fleet’s own data: ETAs, optimisation, document extraction, predictive maintenance. Scepticism is the profitable posture toward autonomy timelines, black-box compliance, and systems that need data you do not have. Start where your data is clean, measure one problem at a time, keep humans deciding.
For most fleets the AI question is not build or buy but which vendor’s embedded feature to switch on: the forecasting inside the TMS, the optimisation inside the routing tool, the document reading inside the accounts package. That embedded route is usually right - the fleet buys the application maintained by someone whose business is maintaining it. The exception worth knowing: the fleet with genuinely unusual data and one sharp internal analyst can beat the market with a small, specific model - the bespoke ETA for one difficult contract, the maintenance predictor tuned to one unusual duty cycle. The discipline is scope: one model, one problem, one measurable output - the moment the internal project grows a roadmap, it has become a software company, and the fleet already has a business.
AI projects die unmeasured: six months in, nobody can say whether the forecasting tool is beating the planner’s spreadsheet, so renewal is an act of faith. The discipline that keeps the portfolio honest is the pre-agreed measure: before go-live, write down the current baseline - forecast accuracy by week, planner hours per schedule, document processing cost per hundred, breakdown frequency - and the number the tool must move to keep its licence. Then review on that cadence with the tool switched off in the comparison where possible: the parallel run is the only clean A/B test a fleet gets. The tools that survive this measurement earn expansion; the ones that fail it get cancelled without nostalgia. The industry graveyard is full of AI subscriptions renewed because nobody wrote down what success was, and the fleets that measure find a surprising fraction of the magic was the attention, not the algorithm.
The limiting factor in most fleets is not the tools but the people reading them: somebody must own the data, question the outputs and translate the model’s confidence into operational decisions. That role rarely needs a data scientist - it needs the fleet’s most numerate operator with the time and licence to challenge the machine. The fleets that invest in that person find every subsequent tool works better, because the scarce resource was never the algorithm; it was always the human who knew when not to believe it.