What It Takes to Let an AI Agent Issue Refunds: Trainline and Anthropic on the Fundamentals of Running Agents That Touch Transactions
Trainline already runs a customer-facing AI agent that processes rail refunds. We unpack the governance, token budgets, and evaluation practices its CTO and Anthropic laid out at TravelTech Show 2026 as preconditions for trusting agents with transactions.
Key Takeaways
- At TravelTech Show 2026 in London, Mike Hyde, CTO of Trainline--one of Europe's largest rail and coach booking platforms--and Anthropic enterprise GTM lead Harry Herbert discussed the fundamentals businesses need as they put agentic AI into serious operation
- Trainline already runs a customer-facing AI agent that autonomously processes refunds on eligible tickets, and the machinery behind it--controls designed around "3,000 untrustworthy engineers," token budgets, and evaluation processes--spells out the preconditions for letting AI touch transactions where money moves
- For retail and e-commerce operators weighing agents for order changes, refunds, and customer service, the session is a concrete reference for an implementation order that puts permissions, audit, and cost control ahead of feature selection
An Agent-Operations Playbook From a Transaction Platform
Trainline has partnered with Anthropic to implement AI internally and in customer-facing applications. At the recent TravelTech Show, Trainline CTO Mike Hyde shared his experience so far.
www.phocuswire.comAt a TravelTech Show 2026 session in London, Mike Hyde, CTO of Trainline--one of Europe's largest rail and coach booking platforms--and Harry Herbert, enterprise go-to-market lead at Anthropic, discussed what businesses need as they scale agentic AI. According to PhocusWire, Trainline has partnered with Anthropic and is implementing AI both in internal operations and in customer-facing applications.
The conversation deserves a wider audience than travel tech. Trainline is a transaction platform: it sells tickets, takes payment, and issues refunds. And as we will see, it already entrusts an AI agent with refunds--an operation where money actually moves. Discussions of agentic commerce, where AI agents handle the journey from discovery through purchase and post-purchase steps, tend to fixate on the shopper's agent. But the seller-side move to delegate transaction processing to agents is quietly reaching production. This session is worth reading as a first-hand account of what that operation looks like from the inside.
Hyde joined Trainline in 2021 after leading data organizations at Meta, Skype, and Microsoft. In April 2025 he was promoted from Chief Data Officer to CTO, with a mandate to embed AI and machine learning into the core of the platform. His remarks carry weight because they rest on 18 months of live operation, not plans to get started.
The Refund-Issuing Agent Is Already Live at the Customer Touchpoint
The premise to hold onto is Trainline Assistant, the company's customer-facing agent. Opened to all UK iOS app users in June 2025, the assistant answers natural-language questions about journeys and ticket conditions--and, on request, autonomously processes refunds on eligible tickets. At launch it was reported to resolve 88% of queries to completion with an average response time of 12 seconds. "The assistant is learning all the time and we're continuing to scale up the agentic AI system which supports it," Hyde said at the time.
On December 19, 2025, Trainline followed with a major app update: personalized disruption notifications, a Delay Repay alert that automatically tells passengers when they qualify for compensation and how much, and Train Swap, which moves travelers onto another train in two taps. The compensation alert enabled the processing of around 1 million pounds in claims during its summer beta alone, and by that point fewer than 10% of assistant queries were being passed to customer service representatives. "We're moving beyond selling rail tickets to supporting customers throughout their whole journey end-to-end," said Chief Product Officer Nina de Souza.
Translate this into commerce terms and it is the automation of post-purchase operations: cancellations, refunds, and compensation. In physical e-commerce, the equivalents are order changes, returns processing, and making customers whole after delivery failures. While AI-powered discovery and recommendations draw the headlines, what makes Trainline's case stand out is that it has built its agent operation up from deterministic actions where a mistake costs real money. The model architecture behind the assistant has not been disclosed; the official page says only that it is AI-powered.
"3,000 Engineers You Cannot Trust": Governance as the Starting Point
So what does a business that lets agents move money stand on? Hyde offered a thought experiment. Trainline has roughly 600 engineers. Now imagine the workforce effectively grows to 3,000 through agentic AI, with most of the additions being AI agents.
The difference is we suddenly have 3,000 engineers, and all of our systems are designed to support 600 engineers, and the 3,000 engineers are less trustworthy than the 600. You can't assume they're going to do the right thing all day, every day. Then, what kind of a security system would you need to run a 3,000+ company where you don't trust all the people?
Source: Mike Hyde (Trainline CTO)
Hyde called this the "mental model" Trainline is trying to figure out and suggested it is a good starting point for many businesses: "What would you do if you were 5x as big as you are, with a whole series of very low-trust employees?"
Conversations about agent adoption tend to drift toward which model is smartest or which tool is most convenient. Hyde reframes the problem as one of organizational security design. For human employees, companies have spent decades layering controls: hiring screens, access management, review, audit. When thousands of agent "new hires" arrive at once, the same level of control has to be rebuilt in a different form. Handing an agent the ability to touch inventory, prices, orders, and refunds is handing the till keys to a workforce you cannot yet fully trust. The implication is that designs which do not assume correct behavior--something close to zero trust--become the standard.
Five Fundamentals and Token Budgets as Financial Discipline
Hyde's fundamentals are five: process, design, governance, security, and financial control. These things are not seen as sexy or headline-grabbing, he noted, but "they're the foundation that lets you adopt whatever the latest model is. The fundamentals don't change as fast as the models do." Models turn over in cycles of months, and investments in optimizing for one depreciate with the next generation. Investments in control design carry over to whichever model comes next. It is the principle of investing in the slow-changing layer rather than the fast-changing one.
The most concrete of the five was financial control--managing AI cost and tokens. Trainline is starting to use a system of individual token budgets for engineers, plus special budgets for projects known to be token-expensive. "But it might be worth it because we're building something new," as Hyde put it.
For businesses that handle transactions, this discipline is not optional. Put agents on customer service and transaction processing, and AI workload and token consumption grow in proportion to transaction volume. Just as cloud cost management matured into FinOps, agent operations need budgetary rigor of their own. Trainline's two-tier setup--personal budgets plus exceptional project budgets--is one of the earliest working examples of managing the unit economics of running agents.
From Tool Rollouts to Job-Level Automation: The 18-Month Inflection
Trainline has been implementing internal AI capabilities for about 18 months, and Hyde says it hit an inflection point in the past three to four months. The initial approach--roll out tools and see what happens--delivered productivity gains of 10% to 20%. The company has since concluded that "it's not enough to just roll AI out; you get to a point where you step back and realize we need to rethink things more fundamentally, like how will we work in the future." It is now building automations that do not merely save someone an hour or two but take on an entire task or job through a series of agentic steps.
| Phase 1 (until 3-4 months ago) | Phase 2 (after the inflection point) | |
|---|---|---|
| Approach | Bottom-up: roll out tools and see what happens | Top-down automation driven by leadership |
| Unit of automation | Saving individuals an hour or two | Replacing entire tasks and jobs with chains of agentic steps |
| Outcome | 10-20% productivity gains | Rethinking how the company will work (in progress) |
| Guiding question | 'Which tools should we use?' | 'How will we work in the future?' |
Hyde named three areas where AI's impact is showing up: building more intelligent customer products on the company's data, understanding and adapting to changes in how people search for and discover travel, and transforming internal operations. The second deserves attention. The migration of discovery into AI chat and AI search hits rail tickets and physical goods alike. Framing AI's impact across three layers--discovery, transactions, and operations--travels directly into retail and e-commerce management.
Herbert added the vendor's view of how customers have changed. Eight months ago, companies were experimenting with scattered tools while employees used personal ChatGPT licenses at work. Organizations now ask how AI should be built into workflows and which tools are actually suited to delivering ROI. What successful companies share: "top-down vision and mandate, and then from the bottom-up is where you find the use cases." Herbert also stressed building processes and evaluation practices (evals) to judge whether a new feature is actually relevant to the use cases and customers being served, observing that "the 'Eureka' moment for a lot of clients is when they see AI working with their own data and their own systems."
The Line Where "Almost Right" Fails
The same TravelTech Show hosted a panel on the industry's readiness for agentic commerce. Travelport CTO Andrew Jordan argued that as travel distribution shifts from deterministic to probabilistic, a firm line separates the realm of discovery, itinerary building, and inspiration from the realm where money is handed over--and on the far side of that line, "you can't get it almost right." On the same panel, Herbert said safety and security are critical in the booking journey and stressed building trust through guardrails.
That line maps onto Trainline's implementation. What its assistant processes autonomously is refunds on eligible tickets--an operation whose qualifying conditions are explicitly defined as rules. Distinguish the discovery layer, where probabilistic suggestions create value, from the transaction layer, which demands deterministic precision, and delegate the latter only under controls and constraints. At a stage where consumer trust in delegating purchases to agents is still forming, that ordering of design decisions is itself a source of trust.
The panel also looked ahead to consumers bringing their own AI agents to book travel, and to the plumbing of connectivity standards such as MCP, ACP, and UCP that would support it (Kismet CEO Jason Cincotta). There is, however, no announcement that Trainline has adopted any of these standards. Its agent today is an implementation contained within its own app, and external agent connectivity should be kept separate from the confirmed facts and treated as the industry's next question.
What Retail and E-Commerce Operators Should Take Away
First, design how far agents may reach into transactions before deciding what they should do. Permissions, audit, and limits need rebuilding on the assumption of a large, low-trust new workforce. Post-purchase operations such as refunds and order changes are a realistic first placement for agents in money-moving territory, because eligibility is easy to encode as rules and easy to audit. That Trainline built its production experience on refund processing rather than flashy discovery features speaks for this ordering.
Second, adopt the discipline of managing AI cost as budget early. Token consumption grows with transaction volume, and left unmanaged it obscures whether the agent operation pays for itself. The two-tier structure of individual budgets plus special project budgets can be copied tomorrow.
Third, put evaluation processes in place before widening what agents are trusted with. Without a systematic way to evaluate outputs, there is no yardstick for judging whether a given operation can be delegated. Whether a company moves past the 10-20% productivity gains of tool rollouts to job-level automation depends on that evaluation foundation, combined with top-down intent and bottom-up discovery of use cases.
Conclusion
Agentic commerce is maturing not from flashy agents that buy anything on a shopper's behalf, but from unglamorous, dependable transaction operations like refund processing--and from the control design behind them. What Trainline's 18 months show is that the precondition for putting AI into systems where money moves is not model intelligence but the fundamentals: process, design, governance, security, and financial control. Models will keep changing; the fundamentals change far more slowly. Whether the party you entrust with a transaction is a person or an agent, trust is produced by the design of controls.
