Copilot is not an AI strategy: how infrastructure prevents standstill with AI

Artificial intelligence (AI) only delivers value to the infrastructure sector when companies maintain control over their own data, security and governance. More pilots are not the answer, but better decisions in core processes, according to PwC experts Mathieu Rosier and Ragnar van der Valk.

Ask an organisation how things are going with AI and you usually hear two figures: how many employees are working with Copilot and how many pilots are running. Both are rising. Yet remarkably little is changing in the work that really matters: at companies in the infrastructure sector, for example, around maintaining, renewing, tendering and prioritising assets.

In the infrastructure sector, we see an important cause for this: most leaders measure AI progress by adoption and experiments, whilst the value lies in core processes. You only get there with your own foundation of data, security and governance. It is precisely this foundation that many organisations have partly relinquished over recent decades.

Meanwhile, time is running out. Europe is at the beginning of a gigantic cycle of renewals. According to the PwC Global Infrastructure Outlook, annual investments in European infrastructure will rise by 42 per cent by 2050, and transport (bridges, tunnels and rail) accounts for around thirty per cent of this. At the same time, the people who know how those assets were built and why certain choices were made are retiring. It is precisely they who can assess whether an algorithm draws the right conclusion. That window is short, and every year that AI remains in the pilot phase, it gets smaller.

Adoption measures activity, not impact

Copilot and similar tools make office work faster. They help with taking minutes, creating summaries and writing emails. That is useful and it helps employees become familiar with AI. As a starting point, there is nothing wrong with that. As a measure of value creation, however, there is, because it does not address the challenge of renewals. A faster summary of a meeting does not bring a bridge into maintenance any sooner.

That it can be done differently is demonstrated by ProRail and Schiphol. Schiphol uses AI to predict not only how many travellers will arrive, but especially when. With these insights, travellers are distributed across available security capacity via 15-minute time slots. Together with NS, ProRail is equipping regular passenger trains with sensors that recognise deviations in the rails and link them to a location. This shifts maintenance from fixed schedules to the actual condition of the track. Neither is an office application. Both change how you utilise and maintain the infrastructure itself.

These opportunities lie everywhere in core processes. Some examples:

• Prioritising maintenance and renewal based on the consequences of failure, not just the technical condition. A rail connection or energy supply that fails brings down more than itself.
• Contract management: unlocking agreements from extensive contracts and calculating what a change or incident means.
• Tenders: gathering and structuring the large amount of information about assets, processes and governance more quickly.
• Risks in major projects: identify what went wrong in similar projects early on, so you and your contractor can build agreements on solid foundations.

A good AI portfolio makes that distinction explicit. The question is not how many initiatives are running. The question is which of them change how you make decisions about your assets and what that delivers in relation to the costs.

Pilots do not fail on the model, but on the foundation

There are enough proofs of concept in the sector. An application that works reliably in practice and meets the quality requirements of critical infrastructure is something entirely different. Pilots rarely fail on technology. They fail on data that is incomplete or dispersed, on security that only looks in at the end, or on governance.

AI is not a black box that solves problems automatically. Anyone who wants to scale must redesign processes, give employees new skills and get data in order. A foundation of data, security and governance is therefore not a precondition that you arrange later. It is the only route from isolated pilots to broad application in the organisation.

Invest in what does not become outdated

Leaders in the infrastructure sector are faced with a real dilemma. A railway line or terminal lasts for decades, whilst an AI application is sometimes outdated within months. If you step in too early, you pay for technology that delivers too little. If you step in too late, you fall behind. That dilemma applies to models and tools, but not to the foundation. The model you choose today, you will replace in a few years. The data about your assets, the agreements about who manages them, the security architecture and the governance of decisions last just as long as the infrastructure itself. Anyone who is afraid of betting on the wrong AI horse should invest precisely there. A good foundation makes you independent of any model in five years' time.

This requires distinction in your portfolio. What already works, you apply now. What is still in development, you explore in a focused way. What only delivers value in five to ten years, you monitor. The foundation beneath it, you build in all cases, because you use it for every time horizon.

You cannot build AI on knowledge you do not possess

Companies in infrastructure outsource a great deal: design, construction, maintenance and inspections. That is often efficient, but it has a side effect that is only now becoming fully visible. Data about assets, maintenance and contracts are partly held by suppliers, in their systems and their formats. And the technical knowledge to assess that data is disappearing with the experienced employees who are retiring.

AI can help build the foundation. With a so-called 'company memory', you can unlock implicit knowledge and informal processes into formal knowledge. AI agents help with capturing and structuring knowledge, and subsequently with making it part of more formal processes, procedures and decision criteria. A digital twin, a virtual copy of physical reality, combines the location of cables and pipes with sensor data about temperature, load or traffic intensity.

A simple test to make visible where you stand

There is also a simple test to make visible where you stand. Can you request the inspection and maintenance data of your three most important assets tomorrow, in a usable format, without calling your contractor? If not, then your first AI task does not lie with a model, but with a contract.

This will become more important in the coming years. The market is moving towards performance contracts, billed on reliability, availability and resilience, and towards platforms on which clients, contractors and technology suppliers share data, risks and value. Then data ownership is no longer a technical detail, but a negotiating position. Anyone who does not arrange it now will negotiate from a position of dependency later.

But that only works if it is established in advance that the data and knowledge belong to the organisation itself and do not end up with a supplier again. That begins with three contract questions:

• who owns the data that a contractor collects,
• in what format is it delivered, and
• do you still have the expertise yourself to assess whether it is correct?

AI can make you more dependent instead of more resilient

For critical infrastructure, there is another layer. A large part of the technology comes from abroad, as do the AI models and cloud platforms on which new applications run. Recent geopolitical events show how quickly energy flows, shipping routes and critical supplies can falter. Resilience, redundancy and security therefore weigh just as heavily as efficiency. Anyone who deploys AI in core processes creates new dependencies at precisely the places where failure hits hardest.

That is not a reason to avoid core processes. It is a reason to weigh resilience from the beginning. Thinking from a minimum viable company (MVC) helps with this. Which processes must continue to work during an incident or crisis? On which systems and suppliers are they dependent? And what happens if an AI application fails? Anyone who does not know this cannot make a responsible choice about where AI belongs and where it does not.

First value or first foundation? Both simultaneously

There are two legitimate objections. The first: anyone who wants to build a complete data foundation first will spend years in a programme without visible results. The second: anyone who puts AI in core processes too quickly in critical infrastructure takes irresponsible risks.

Both points are valid. That is why we choose neither extreme. Choose two or three applications in core processes that really matter, such as maintenance prioritisation or contract management, and build the foundation precisely for that. Then every investment in data, security and governance delivers immediate value, and the foundation grows step by step with each subsequent application.

Core process here does not mean that AI decides independently about a bridge or a power grid. It means that AI supports better decisions, with a human who assesses, adds context and remains responsible.

Two questions for the next board meeting

The infrastructure sector does not have the luxury of continuing to treat AI as an experiment. For leaders, it therefore begins with two questions:

• Which of our AI initiatives actually change core processes and which only make office work faster?
• Where is our foundation of data, security and governance, and do the data about our assets actually belong to us?

If you mainly mention Copilot in answer to the first question and have no clear answer to the second, then you know where the real task lies. It does not lie in more pilots, but in regaining control over your own assets. AI does not make that control superfluous. AI makes it indispensable.