The pathfinder plan to get rail’s AI past the pilot and into everyday practice

“You prove it works, then it’s put out to tender and sold back to the lowest bidder. So you’re losing time, you’re losing money, you’re losing continuity, you’re losing knowledge.”

That is how Doctor Sarah Schlobohm describes the way proven ideas can be lost once they go out to tender. Britain’s railway has no shortage of good ideas about artificial intelligence (AI). What it has lacked is a dependable way to turn them into everyday practice.

The Artificial Intelligence in Rail Action Plan, published in April 2026 by GBRX, the strategic innovation and technology body for Great British Railways, sets out to change that, and its central idea is the pathfinder: a way of driving a proven idea all the way through the governance and data barriers that usually stop it, so that everyone who follows can take the same route. Global Railway Review spoke to Schlobohm, who leads the plan’s delivery as Director of the Artificial Intelligence Incubator Accelerator (AIIA), about how pathfinders work and where AI is already earning its keep.

You prove it works, then it’s put out to tender and sold back to the lowest bidder. So you’re losing time, you’re losing money, you’re losing continuity, you’re losing knowledge.” 

Doctor Sarah Schlobohm, Director, AI Incubator Accelerator, GBRX

Schlobohm came to rail from regulated finance, by way of particle physics and a chartered accountancy qualification, and one of her first tasks was to separate what is true of any AI adoption from what is specific to the railway. Data sits in both camps. It is the perennial problem, garbage in and garbage out, but in rail it is unusually fragmented, a legacy of a privatised industry whose parts were never set up to share information easily.

What a pathfinder changes

A pathfinder, Schlobohm is careful to say, is not a pilot with a new name. Rail has never struggled to produce pilots; it has struggled to carry them past the proof-of-concept stage. The point of a pathfinder is not to scale a single project but to scale the route, doing the hard work through governance and data barriers once, so the next team knows how to navigate the same territory. “We can sort this requirement once and use it in every project,” she adds.

Part of the old problem was incentives, she observes: there has been money for hackathons, pilots and proofs of concept, but rarely the funding to carry a project through to adoption. Measure success by the number of pilots produced, and the sector can look busy while nothing actually changes. “We need to focus on the outcome. Is this making the railway run better, rather than just numbers delivered?”

Fixing that means addressing the commercial break as much as the technical one. GBRX is building a new commercial route, its proposed Utilities Dynamic Market (UDM), compliant with the Procurement Act 2023, so a supplier who succeeds has a way to keep going. GBRX published a preliminary market engagement notice for the UDM on 7 July 2026, which describes a market it intends to establish rather than one already operating. Under the Act, a dynamic market is an always-open list of suppliers that have met membership conditions and can be considered for future procurements. Pair that with a clear governance pathway, she argues, and good work finally has somewhere to go, rather than being run quietly in parallel and never formally adopted.

Delivery follows a federated model, with operators keeping ownership while GBRX coordinates. Schlobohm is well placed to do it, speaking on any given day to Network Rail, DfT Operator, the Rail Delivery Group (RDG) and the Rail Safety and Standards Board (RSSB). A separate innovation function that spans the industry, she explains, can spot where a requirement solved once could serve every project, something a single operator, measured against its own objectives, is rarely positioned to do. The plan is deliberate about plugging into existing governance rather than adding new layers.

Where AI earns its keep

Asked where AI will pay off first, Schlobohm points to the pockets of excellence already delivering on individual routes and regions, and to the priority pathfinders drawn up with the sector’s own experts. The most advanced is in fares, ticketing and retail, where a data-driven approach is being applied to fares reform. The industry has inherited more than 400 ticket types, which multiply out to some 55 million fare combinations, far too many for any person to hold in their head, so GBRX is using a classical machine learning technique to sort them into a handful of industry-defined buckets. The technique is chosen because it works, not because it is the newest, and that is the point.

“It’s not the shiniest, fastest technique, but it’s a proven method that works really well. I don’t think spanners have changed much in the last couple of hundred years, but we keep using them because they work,” she explains.

One pathfinder, turning free text into structured operational data, appears twice in the plan precisely because it feeds so many other parts of the system. The bigger prize, Schlobohm argues, comes when that information flows across the whole railway rather than stopping at one team, reaching the passenger, the crew, the depot and the station alike.

Data, value and the confidence to deploy

Getting there depends on data, and Schlobohm is clear-eyed about it. Some sits locked inside rolling stock company (ROSCO) and original equipment manufacturer (OEM) contracts written before cross-boundary reuse was a consideration, but the bigger early opportunity, she concedes, is the data it already holds and cannot yet fully use for want of shared definitions. GBRX wants to offer common contract language so that, in future, the industry is not paying to buy back its own data. “It’s better for everyone if the trains run better.”

As a chartered accountant, she takes a broad view of value. Strict return on investment for AI is unreliable, and much of the benefit is qualitative, or simply the price of staying in the game. The gains come from going deep enough in one area that the data compounds, rather than scattering use cases picked by a finger in the air. Set against the familiar warning that most AI projects fail, she is blunt about the upside. “The ones that do succeed, succeed way better than you expect them to.”

For a safety-critical railway, trust matters most. The tension between deterministic and probabilistic systems is not new to AI, Schlobohm points out, and rail is far from the only safety-critical sector to have worked through it, alongside manufacturing, aviation and financial services. The fair test, she suggests, is not perfection but whether AI matches the error rate of the people doing the work today. Assurance is where the Office of Rail and Road (ORR) and the RSSB come in. “The ORR has to mark our homework,” she says. Schlobohm believes that having convinced financial regulators, she can convince the ORR.

A year from now, the surest sign of progress will be cultural: teams pairing data with instinct rather than trusting instinct alone, and a growing willingness to solve a problem once, for the whole industry. The strongest signal so far, Schlobohm reflects, is that simply naming the challenge has struck a chord. “Even that tells me the plan is working.”