Interview: How AI can transform fleet operations

What are the most immediate benefits AI can offer to van and HGV fleet owners in day-to-day operations?  

There is a broad range of elements that AI can offer – for example in predictive maintenance, driver assistance through in-spec or after-market Automated Driver Assistance Systems (ADAS) in vehicle, in weather alerts, traffic alerts and rerouting and in Automatic Number Plate Recognition (ANPR) in tolling and parking applications.  

Alongside these existing applications, future applications include load and route optimisation using hauliers' own data, behaviour mapping and intervention systems to help with safety and reduce accidents, predictive maintenance using AI to interrogate data streams from the vehicle plus external inspection information.  

Other future applications could include intelligent integration between fleet vehicles and a third-party logistics company system to enable true interoperability. All of these applications are possible to develop today.  

How can AI-powered route optimisation help fleets reduce fuel costs and improve delivery times?  

By taking the data that exist in the haulier’s systems, potentially augmenting this with synthetic data from simulations and external data sources such as road information or weather, then training AI systems to optimise for certain conditions.  

For example, the algorithms may be trained to optimise for shortest delivery time or most fuel-efficient route or one that maximises the driver utilisation within regulation, or a combination of these. Importantly, the set of synthetic scenarios that can be modelled during training is expanding all the time, and can now include rare events or disruptions, allowing AI to learn to make decisions that are resilient against those. 

Overall, the system level can be tuned to get optimal sustainability, safety, and delivery performance even when things are not running smoothly.  

Driver safety is a major concern — how can AI enhance driver behaviour monitoring and accident prevention in commercial fleets?  

There are many data sources for monitoring driver behaviour – speed, braking severity, cornering, dashcam data etc plus in cab systems such as measuring driver facial expression, blink rate etc plus external sources such as weather data, traffic density, road conditions, etc.  

These are all rich sources of information to train an AI system along with outcomes – if a near miss or an accident happened what factors were involved? This data can be used to train an AI to spot the potential for an accident in the moments leading up to this.  

Additional data can also be accessed through infrastructure sensor systems that enable near-real time traffic conditions. Combine this with elements such as human behaviour modelling – we use the term Human-Machine understanding (HMU) – then you can develop interventions that can be utilised to prevent accidents happening.  

AI is a very powerful tool and HMU is the skillset to enable the AI prediction to consider how humans in the environment will behave – allowing effective intervention for prevention of a hazardous outcome.  

The AI-HMU approach is cutting edge, but it is possible to develop systems that could have significant impact in commercial fleets.  

What role does predictive maintenance play, and how does AI help fleet owners avoid unexpected breakdowns or costly repairs?  

Predictive maintenance is a likely early win for AI in fleet management. A wealth of data exists for vehicles from telematics and sensor data plus existing maintenance information (usually digitised already) and an understanding of external conditions such as road state and weather etc.  

This is a good source of data to train AI models to look for patterns that lead to failure. This can be by looking back on where reactive maintenance was needed – what led to this need? Or where planned maintenance showed need – or not – for component change or refurbishment.  

The system can be trained to look for patterns and then highlight when the early stages of these are occurring to allow preventative intervention rather than the expense and disruption of failure in the field.   

Furthermore, AI can be used to optimise the supply chain (for example by ordering parts as needed, at the least cost). Predictive maintenance is often talked about, but very rarely implemented at scale.  

The key barrier is the scope of the predictive model; it’s often set too wide, failing to cover unusual events, closed-loop interactions, or human behaviours, and therefore not giving an optimal solution. More successful systems tend to be narrow – trying to model specific types of fault or smaller decision spaces where there is known good data.  

Once the value is proven at this smaller scale, targeted datasets can be collected to widen out AI training to include more behaviour models and deploy in a wider set of scenarios.  

How can AI be used to streamline compliance with regulations like driver hours, emissions, and vehicle inspections?  

These regulatory compliance needs are not new and checklists and compliance to rules is part of the day-to-day operations of the fleet. AI can help with scheduling or tracking these, but so can many existing approaches - so AI is likely to be additive rather than transformative in this area.  One area where an AI system is likely to be additive is in spotting patterns of non-compliance, or more general anomaly detection, and suggesting intervention ahead of the fault re-occurring.  

There’s often concern about upfront investment — what kind of ROI can fleet owners realistically expect from implementing AI solutions?  

This is a challenging question – and it depends on what systems are being developed and implemented. Elements like predictive maintenance are likely to have a relatively short ROI when the total cost of unexpected failure is considered. It will come down to the use case and the economics of each case. Does the business understand the cost of these use cases?   

And we should also bear in mind that ROI is not always purely a monetary analysis, other elements such as safety, operational efficiency and carbon footprint reduction may or will form part of the equation. In the end the economics need to stack up. Businesses often have a qualitative, or ideally quantitative, understanding of where the major areas for improvement are.  

AI is a powerful tool to deploy against these in many cases. Often there are some obvious areas to examine that can yield early impact and build confidence and acceptance of this type of technology. Alongside the technology investment, it is essential to assess the challenge of gathering, cleaning and maintaining data from the requisite sources to train and operate the AI models.  

Data preparation can cost twice as much as the core AI technology, so this must be carefully evaluated before proceeding. Finally, the investment in people, process and organisation needs to be carefully considered. The best technology in the world will fail to be successful if the business processes and the users are not considered when implementing.  

How should fleet owners begin evaluating which AI tools or platforms are right for their size and operational needs?  

Our AI Implementation guide was created to equip transportation organisations with clear, actionable steps for successfully implementing AI and provides guidance on tools and platforms, but fundamentally it should start with a pressing problem.  

This issue is hurting my business, either in terms of competitiveness or inability to seize an opportunity. Identify those challenges and then look at the solution – AI is not necessarily the answer by the way! Identifying the right outcome to solve the problem at hand is. But if it is a situation where there is a rich source of data then potentially AI is a powerful tool.  

Once there is an understanding of what areas are the most tractable, then look at what products and services are available.  

Will these work out of the box? How much and what type of data is needed to train the system, is this open source? Is the data traceable and of sufficient quality at training time and inference time? Will it all come from your systems? Will the real world ‘drift’ away from the training data over time, will this be monitored where required?  

There are many providers of tools and often these will work well. In some cases, off-the-shelf packages won’t be suitable, because they can’t be easily integrated into legacy operations. Then it is a case of developing new tools, either with the right partner or potentially with internal teams, where it makes sense to build that capability.  

Can AI help with sustainability goals, such as reducing emissions or preparing for the transition to electric vehicles?  

Optimising routes and lane choices brings sustainability and fuel consumption benefits alongside bottom line efficiencies. As wider sets of “unexpected” situations are considered in plans, predictive systems and responsive strategies, AI becomes a more natural solution for this type of optimisation.  

AI might also help analyse driving behaviour (or help simulate what drivers will do) for fuel benefits. And across electrified transport there are many applications already being deployed, from learning optimal charging strategies, to planning infrastructure locations, to analysing and predicting battery health. Managing EV fleets will also present a new set of challenges - this will be more complex than managing diesel fleets.  

Fortunately, AI can help us with the optimisation problem, allowing fleet managers to balance battery range limitations, charging station availability and energy-efficient routing is essential. It is also important to account for fleet drivers who may or may not have access to overnight charging at home.  

Finally, alongside emerging ‘flexibility’ market-based decision making, and direct large scale energy control (where AI may have more regulatory issues), there is room for plenty of automated decision making in moving energy to where it is likely to be needed, whether that’s in grid scale storage, transport system batteries, or novel emerging storage technologies.  

What are the biggest mistakes or misconceptions you see when it comes to adopting AI in the fleet management space?  

AI is not just ‘plug and play’. Behind the core inference algorithm, you might experience as an end user (or buy from a supplier), at scale there are training systems, data systems, data ownership and privacy processes, and assurance processes covering safety, ethics, performance, accountability, transparency, sustainability, and regulatory aspects.  

These need to be embedded into the wider organisation rather than being part of a dedicated AI function. Keeping AI systems adaptable needs to be a conscious decision – the best AI optimisation today might be different later when commercial incentives or the wider world changes, and there needs to be a commercially viable mechanism to (have a supplier) ‘retrain’ or adjust goals – ultimately to stop a successful AI deployment becoming a barrier to future innovation. And fleet data can be sensitive, potentially including driver behaviour, location tracking, customer details. Data governance should be built in at an early stage to avoid expensive rework later in scaleup.  

Looking ahead, how do you see AI transforming the logistics and commercial transport sector over the next 5 to 10 years?  

AI will continue to advance in the vehicles themselves - driver assistance, autonomous delivery, platooning, collaborative autonomous agents, human/robot teaming and the like, but will also start to be aided by wider decision-making systems on routes, inventory, and scheduling - based on more live inputs such as weather and geopolitics, improving both resilience and efficiency.  

This applies to transport of both people and goods, taking effect within logistics depots down to package handling level, alongside fleet operations. AI might also take on some of the complexities of optimising between organisations – for instance sharing your charging infrastructure with a third-party fleet, or lending your energy storage to a party that needs it before you do.  

Looking further into the crystal ball - given the success of large language models in interfacing with humans, we can imagine a suite of chained foundation models for completing large scale goals like ‘move this cargo from A to B’ – these components might be LLMs where natural language is specifically required, but also dedicated large ‘spatial’ models that understand and can predict the physical world, augmented perception models that perceive what other actors are trying to do, reasoning and decision making models, fine-tuned robotic models that encapsulate the behaviour of particular hardware systems, and so on. 

Where these systems are making decisions autonomously without humans in the loop, or with only limited supervision, this is termed Agentic AI. Using Agentic AI can be powerful both when deployed as an individual model or more impactfully when a series of AI agents working together break down complex problems where one overarching model would be unwieldy or impractical. However, this requires greater level of assurance in design and deployment (as discussed above) given its autonomous nature.  

These might ultimately be widely applicable modules provided by a technology supply chain but chained together or co-optimised in novel ways to achieve commercial tasks. Deploying this Agentic AI approach at scale would need organisational transformation across people, processes, technologies and policies, but those that are fastest to take advantage may be best placed to succeed.  

Published On: 12/06/2025 15:00:45

Comments Section

If you are a Logistics UK member login to add comments.