How AI is revolutionising customer service in the travel industry

The travel industry keeps growing. Global travel and tourism revenue is estimated at £802.82bn this year and projected to reach £1,043.12bn by 2030. Behind those figures sits a customer service reality that grows with them: more bookings to fulfil, each bringing requests, modifications, unforeseen events and operational issues, all while ensuring that every guest has the best possible experience on a trip that matters to them. 

A single trip can also involve several actors — from the travel agency and accommodation provider to transfer companies and activity operators — meaning that one issue often requires coordination across multiple parties. 

AI has turned out to be a powerful ally in keeping pace. However, despite customer service being one of the most popular use cases for AI agents, there remains a wide chasm between a demo or pilot that works and the reliable operational system at scale needed to transform how customers are served. 

The real opportunity is not simply to replace individual tasks with automation. It is to combine autonomous agents that can resolve straightforward cases end to end with service professionals who are supported by AI when dealing with more complex or higher-stakes interactions. 

From what I have seen, there is no single right approach. Where resolution is straightforward, companies can prioritise rapid, accurate automated resolution, increasingly through autonomous agents that take ownership of a case from beginning to end. When something has gone wrong, what matters most is having experienced service professionals, aided by AI, to navigate the interaction. 

These are two different implementation angles, but both are essential. The lessons behind them are equally important: it all starts with classification, language no longer needs to be a barrier, the real test is production, and people remain the differentiator. 

It all starts with classification 

It sounds unglamorous, but automated classification is the foundation of everything else. 

Many customer service processes exist the way they do for one simple reason: a request is only understood when a person reads the email or picks up the phone. Classification and resolution have traditionally happened together, in the same step and often by the same person — unless companies were willing to subject customers to endless IVR menus or unreliable self-classification forms. 

Automate the first step and the two activities can be decoupled. That opens up real process-reengineering opportunities: routing according to complexity, creating specialist teams for particular case types, and introducing automation wherever a process is sufficiently mature. 

In a large travel operation, this may involve classifying demand across dozens of intent categories. The objective is not only to identify what the customer is saying, but to understand why they are contacting the company and what operational process is required to resolve the request. 

This was not practical until relatively recently. Earlier attempts at intent detection using machine-learning-based natural language processing required endless training phrases. Intents frequently collided with one another, and the systems were difficult to build and even harder to maintain. 

Large language models have changed the game on both fronts. They have accelerated implementation because they do not require the same volume of extensive training, and they have improved accuracy because they are more sophisticated not only in generating language, but also in understanding it. 

Early implementations often depended on heavyweight models. Today, lightweight models can increasingly perform many of these classification tasks effectively, reducing latency and the cost of running them at scale. 

There are broadly two ways to proceed: wide, by classifying all demand so that it can be routed more effectively; or deep, by fully automating a limited number of high-volume case types. 

Both approaches have value. Going wide creates order and visibility across the service operation. Going deep can generate earlier operational impact by automatically resolving a proportion of repetitive cases. In practice, the strongest transformation strategies are likely to combine the two. 

Breaking the language barrier 

Language matters particularly in conversational channels. 

An email can wait a few minutes while a translation tool is used. A live chat cannot. In the traditional service model, a real-time conversation depends on having the right multilingual specialist on shift at the right hour. Outside that window, the customer either waits or is passed to someone who may not be a native speaker. 

Real-time, two-way translation can remove that constraint. Lightweight language models can enable a service professional to write in one language while the customer reads and responds in another, allowing the conversation to continue naturally and without interruption. 

But using a generic translation tool is not always enough. A term that means one thing in everyday language can mean something entirely different in travel. The context provided to the model therefore matters enormously. Industry terminology, booking language and company-specific vocabulary all need to be understood correctly if translation is to be operationally reliable. 

When implemented well, this approach can improve both resource allocation and the customer experience. Service professionals are no longer restricted to handling contacts in a narrow group of languages, while customers can receive real-time assistance in their preferred language. 

The technology does not eliminate the need for language skills or cultural awareness. It does, however, allow organisations to deploy expertise more flexibly and make support available to more customers at more times of day. 

Distrust the demo 

Anyone can put together an impressive AI demo in a week, and companies are frequently approached by vendors showing apparently seamless examples. 

The distance between that demo and a production system is where most initiatives stall. 

Production means real data, real volume, integration with legacy back-office systems and telephony platforms, and dealing with latencies that a controlled demo rarely reveals. It also means meeting the operational efficacy and predictability that customer service requires: the system cannot fluctuate unpredictably in its behaviour from one interaction to the next. 

So how should companies cross that distance? Gradually. A sensible starting point is often deterministic, with AI doing little more than classifying cases while the resolution logic remains explicitly defined. This can take an organisation a long way while keeping operational risk tightly controlled. 

However, every new use case and every corner case then requires another rule to be defined manually. Diminishing returns arrive quickly. Covering more ground requires not only embracing generative AI further, but also giving the system greater agency. It must be able to interpret a problem, decide what information or action is needed and use the tools available to reach a resolution. 

That makes the system more flexible, but flexibility also introduces risk. Guardrails therefore need the same level of investment as the use cases themselves. That means human oversight, strong monitoring, traceability and auditing, checks that keep every resolution within defined boundaries, and reliability that can be measured and demonstrated. 

This is why customer service is now moving towards agentic AI. The customer explains the problem once, and the system either resolves it by reasoning and using the tools at its disposal or passes it to a specialist with the full context attached. 

Owning a case end to end can go surprisingly far. In more advanced implementations, an agent can consult a booking in real time and, when the answer is not already available, contact the relevant supply partner, interpret the response, extract the necessary data, update the booking and respond to the customer. 

The significance is not simply that individual steps are automated. It is that the system can close the operational loop without requiring a person to move manually between disconnected systems and communications. 

The work only people can do 

In customer service, the moment of truth is often when something goes wrong in the middle of a trip. 

That is when experienced professionals make the greatest difference: untangling bookings involving several parties, finding answers where policy alone does not provide one, and listening to a customer who needs to feel heard. 

The role of AI should be to give those professionals every possible tool to deliver excellent service. Cases can arrive classified, triaged and enriched with context. AI can absorb administrative work, prepare information and identify sentiment so that urgent or sensitive requests are routed appropriately. 

There is sometimes an assumption that average productivity will fall as automation absorbs routine cases and leaves service teams dealing only with the more difficult ones. That does not necessarily follow. 

Partially resolved cases can move faster because part of the work has already been completed. Instant responses can also eliminate duplicated contacts: when the answer arrives within seconds, the customer has no reason to call or send another message to follow up. 

This is why measuring automation only through the number of cases fully resolved by AI can be misleading. Its wider contribution includes reducing handling time, avoiding repeat contacts, improving routing and giving specialists better information before they engage with the customer. 

Understand every request from the first second. Speak every customer’s language. Judge AI in production, not in the demo. And place people where their judgement, expertise and empathy make the greatest difference. 

That, in one sentence, is how AI is revolutionising customer service in travel: autonomous agents owning what can be automated, and empowered experts delivering the service only people can.