AI Agents: Transforming Rail Passenger Assistance

The data exists but rail can't use It

The ORR's recent call for better data use in passenger assistance hints at a fundamental architectural flaw.

Rail operators already hold massive datasets on booked assistance, station layouts, lift availability, and staffing levels. The problem is these systems don't connect. And they certainly don't adapt in real time.

The failure point is the interoperability of the data and the barriers at handover between systems, between people, and between organisations.

When compliance becomes the problem

I saw this pattern clearly in the Accessible Travel Research Study I led for a UK train operator in 2019. One story kept appearing in the feedback.

A passenger books assistance in advance. Their train runs late. The receiving station marks them as a "no-show" because the system doesn't update when reality changes.

The record shows complete. Nobody breached procedure. Yet the passenger is stranded.

This is what happens when assistance is treated as a task to close rather than a service to sustain.

The ORR's 2024-2025 data shows 11% of passengers received none of the assistance they booked, despite a 20% increase in requests to 312,576. The system is buckling under demand it can't orchestrate.

The one-way data problem

When we built the DataWharf agentic assistant, we designed it to enable operators to track traveller progress, flag missed connections, and notify staff automatically. The logic works but unfortunately the data flows available in the rail industry currently don’t allow for this end-user feature.

Rail creates open data feeds, but they're one-way streets. Our agent can read them but can't push information back. It can't tell the system when a passenger's plans change or when assistance needs to shift to a different platform.

The ORR identified that electronic handovers reduce information loss compared to phone calls. But even electronic handovers fail when the architecture doesn't allow bidirectional flow.

Feeds about staffing, lift status, and live arrivals either aren't open or aren't structured for real-time orchestration.

What AI agents actually need

AI agents could personalise the entire passenger experience. Spoken interfaces. Natural language in any tongue. Customised help levels. You could outsource your entire journey to an agent that makes optimal decisions based on your specific needs.

But only if the infrastructure exists to support it.

We also know that LLMs and agents can hallucinate when they don’t have the data they need, offering inaccurate information unless they are explicitly told not to. We’ve curated the DataWharf agent only to offer factually accurate support, and to be honest when it doesn't know something.

That requires trust. And trust requires fail-safes.

I foresee the most in demand part of an agent’s job would be to support a customer during disruptions yet this is the one time when the rail industry turns a little chaotic and provides multiple different information sources saying multiple different things.

The result is that the agent might need to tell passengers not to travel by rail and suggest a bus instead.

Rail operators can't expect vulnerable passengers to rely on AI agents until the data connections behind the scenes are bulletproof. Someone has to monitor those feeds. Systems have to plug together seamlessly.

The fragmentation risk

We're heading toward a world where every passenger might have their own agent. Some optimised for operator interests. Others for passenger interests.

Each working with different data quality. Different prompt engineering. Different advice.

This could recreate the exact coordination failures we're trying to solve.

The only solution is a single source of truth. All agents working from the same information and available tools from rail operators. Organisations like RSSB, GBRX, and National Rail need to create standards that make these systems interoperable yes all working from the same truth.

McKinsey estimates railway companies could unlock $13 to $22 billion annually through AI adoption. But few are implementing at scale because the foundational data architecture doesn't exist yet.

Awareness is the bottleneck

The technology barrier is real. None of these bidirectional feeds exist today.

But they won't get built until people realise there's a problem to overcome. Rail operators need to understand that AI agents aren't just another customer service channel. They're a fundamental shift in how passengers interact with transportation.

When agents work properly, they don't just answer questions. They orchestrate entire journeys across multiple operators, adapting in real time to changes.

That requires rethinking data as a live, networked information layer rather than siloed databases.

The ORR is pointing at the gap between data collected and data used. What they're really highlighting is that the architecture is wrong.

We have plenty of compliance. What we need is orchestration.