A railway wagon can remain idle because a piece of information, a document or a decision is missing. Skilled people work around that wagon, often spending their time searching for data, comparing versions and chasing responses. Artificial intelligence offers the opportunity to redesign this work, entrusting entire sequences of execution to the machine and giving people back time to decide and create.
When we think of AI, we almost always imagine a chatbot: we write down a wish and wait for it to take shape. We ask it for a summary, a translation or a presentation. This is a useful way to access the technology, but it leaves the organisation of work almost untouched. People continue to break down the problem, give every instruction, transfer information from one application to another and piece the result back together.
In this sense, our approach remains rooted in the twentieth century: we adopt a new technology while retaining the previous division of tasks. We treat AI like a secretary or a concierge, when we could start thinking of it as an operational colleague to whom we assign a mission, with clear objectives and limits. Responsibility for that mission remains with people.
The car changed the possibilities for travel; the plough transformed agricultural work. Innovations of this magnitude also change what a society can imagine doing. Their social costs must be addressed alongside their opportunities. I have already discussed these themes in my book on automation. Today, I would like to take the argument further, towards the space we can give back to human thought.
There are data that help us gauge the scale of the change. An experiment involving 758 Boston Consulting Group consultants, described in a 2023 working paper by Dell’Acqua and colleagues, found that, on tasks within GPT-4’s capabilities, AI users completed 12.2% more tasks and worked 25.1% faster. On a problem selected to fall beyond those capabilities, however, the probability of a correct solution decreased by 19 percentage points. These results relate to specific consulting tasks: the benefit must be verified in each operational process. mitsloan.mit.edu
The organisational lesson is clear: we need people who can recognise where to entrust execution to the machine and where to exercise their own judgement. Even complex activities can be accelerated when data, tools and controls are properly prepared. The apparent complexity of a task, on its own, tells us little about how readily it can be automated.
In logistics, a well-designed system can connect orders, asset availability, transport documents and operational information. A disrupted shipment may require contracts, routes and terminal time windows to be compared, alongside recalculating costs and updating customers. An integrated system can coordinate this sequence and update business applications within the authorisations it has received. To do so, generative AI must work alongside optimisation algorithms and technical checks: interpreting a request, calculating a plan and authorising a movement are different functions.
Let us therefore imagine equipping every manager with an AI environment built around their expertise. A fleet manager could be assigned the mission of increasing wagon availability while respecting maintenance constraints, costs and contractual commitments. The system would investigate the causes of downtime, prepare alternatives and follow through on permitted activities. The manager would intervene on exceptions, priorities and decisions that require accountability.
This would change the relationship with the central business management system. Applications designed by people who know the work could be built around it, capable of coordinating an entire process. Every action should be traceable and every result verifiable. The company’s DNA would begin to change when people’s operational knowledge became part of how its systems function.
For the railway sector, the urgency is tangible. According to Eurostat, in 2025 the main railway undertakings in the European Union recorded 368.2 billion tonne-kilometres of freight transport, down 1.8% on 2024: the fourth consecutive year of decline. The figure does not attribute the decline to a single cause. It does, however, highlight the need to make the service more competitive, including through the quality of its processes. Eurostat. The transition has already begun. In its 2026 interim report, DB Cargo reports five use cases implemented on an agentic AI platform, two already in production. It also documents the development of a system that recognises wagons’ loading status through images and AI, and preparatory work for applications to checks and invoicing associated with wagon regulations. These are concrete advances, at different stages of maturity. Deutsche Bahn Interim Report 2026
A political question remains, alongside the industrial one. A tool accessible through a chat can appear democratic while the ability to build applications on top of it remains concentrated in a few organisations. The distribution of power also depends on access to data, skills and infrastructure, and on the ability to modify the tools we work with. Pioneers can use this advantage to concentrate control further, or to distribute the capacity to design solutions. In a railway business, this means also involving those who work in yards, workshops and terminals. It means retaining access to one’s own data and building applications that are portable across models and providers, reducing dependence on the design choices of a single platform.
The energy cost must also be part of this vision. The International Energy Agency estimates that all the world’s data centres consumed 485 TWh of electricity in 2025, an increase of 17% in one year. Its 2026 central projection reaches approximately 950 TWh in 2030. These figures cover the entire data centre sector, of which AI is one component, and the future figure is a forecast. IEA
Greater capacity for execution can fuel overproduction, additional consumption and activities with no useful purpose. Putting people back at the centre means giving them the tools to choose the purposes to which that capacity should be directed. In logistics, the environmental benefit will need to be measured by considering the digital cost as well as reductions in empty journeys, waiting and waste. The time recovered should have an explicit purpose: designing new services, developing skills and understanding customers better. If every hour freed up is immediately filled with more administration, we lose a decisive part of the opportunity. Productivity may be one result of the transformation; its more ambitious value is to make frontier thinking possible.
From this perspective, rail logistics opens up a wealth of possibilities for integration between people and machines.
Give every shipment an operational mission. An agent could follow a transport operation from booking to delivery, recognising in advance when a delay threatens the unloading window. With partners’ systems connected, it would prepare alternatives and coordinate authorised activities. The person would choose which trade-offs to accept between cost and service. The customer would receive an updated commitment based on options that are actually available.
Turn railway workers’ experience into a shared asset. Including for the safety of the system. A site supervisor or a technician could describe a recurring problem to AI and help build a solution. The system would translate that experience into a proposed procedure or an application, to be tested against historical data and validated in the field. Those who know the work would take part in designing the tool, with explicit recognition of their contribution.
Connect capacity that is currently separate. Railway undertakings, wagon keepers and terminals could share selected availability through common interfaces while retaining control over their information. Agents would look for combinations to build return loads and use free operational time windows. People would agree on terms and commitments. Even small operators could gain access to a capacity for coordination that is difficult to sustain individually today.
Create a permanent laboratory for new services. Managers, technicians and commercial teams could use AI to explore flows currently underserved by rail, comparing frequencies, terminals, wagon types and loading solutions. A digital twin, meaning a model of the service fed by operational data, would help test scenarios before investing. Engineering checks would remain the responsibility of specialists. The time freed from execution would thus become time devoted to creating a market.
Move beyond the logic of limited responsibility towards responsibility for the system as a whole. The tendency I have described several times to find someone responsible for every event that occurs in a cooperative setting is a sickness in this sector that comes at a heavy cost. AI must take us where we have not reached so far: to conceive of technology as a tool for shared growth along the transport chain, ensuring that every technical detail is consistent so that transport is safe and efficient.
The market will recognise the value of those who can translate this integration into reliable transport and possibilities they were previously unable to offer. Progress will be visible when a manager can devote their day to building a new logistics solution while the machine works to turn that solution into a concrete, efficient and safe service.