In this episode of The Inference, Paul Curtis, CTO and Ecommerce Director at easyJet, joins host Michelle Hawley to discuss what it actually takes to move AI from a proof of concept into production inside one of Europe's largest airlines.
Drawing on easyJet's experience running over 150 POCs, Curtis shares how the company has narrowed its focus to AI use cases that deliver measurable value, from empathy-aware customer service bots to hyper-personalized holiday content tailored down to the individual. The conversation covers how composable architecture and MCP adoption have become essential prerequisites for agentic AI, why total cost of ownership is the question most executives are still avoiding and how the build-vs-buy calculus is shifting as SaaS products lose ground to custom AI solutions.
Host
Guest
Michelle Hawley
Paul Curtis
What Stood Out From Our Chat
Table of Contents
- Focus AI Investment on Measurable Business Value
- Keep Humans in the Loop When Emotions Run High
- AI Agents Depend on Modern Enterprise Architecture
- Give Agents Autonomy Where Failure Is Manageable
- Executives Must Calculate AI’s Total Cost
Key Takeaways
- easyJet narrowed its AI efforts after running roughly 150 proofs of concept.
- Modern APIs and composable architecture are prerequisites for scalable AI agents.
- Enterprises should grant autonomy first where the consequences of failure are limited.
easyJet once had roughly 150 artificial intelligence proofs of concept running at the same time. Few produced meaningful business value.
That experience helped push the European airline toward a more disciplined AI strategy focused on practical use cases, modern architecture and clearly defined limits on autonomy, according to Paul Curtis, easyJet’s chief technology officer and ecommerce director.
Speaking with VKTR Editorial Director Michelle Hawley on The Inference, Curtis discussed how easyJet is moving AI into production, where agentic systems could reshape travel and what executives still misunderstand about AI costs.
Focus AI Investment on Measurable Business Value
Like many large enterprises, easyJet initially explored a wide range of potential AI applications. Curtis said the airline eventually realized that experimentation alone was not enough.
“At one point we had something like 150 POCs running at any one time, but so few of those then actually turned into something tangible with real business value,” Curtis said.
The company has since concentrated on areas where AI can produce measurable results, including customer service and personalization.
In customer support, newer AI systems can handle more complex, multi-intent requests than traditional chatbots. A traveler might complain about a delayed flight, an incorrect seat and a missing onboard product in the same conversation. AI can identify each issue, determine whether compensation may apply and adjust the response based on the customer’s sentiment.
Curtis said easyJet is even seeing cases where chatbots and voice-enabled AI demonstrate more empathy than their human counterparts. That capability can help the airline provide more timely service while improving customer satisfaction.
easyJet is also using AI to move beyond broad customer segments toward personalization at the individual level.
“You can almost have a personalization experience down to a unit of one person,” said Curtis.
Offers, hotel descriptions and travel recommendations can be adapted based on a customer’s past behavior and recent browsing activity. A hotel page might emphasize children’s activities for a family while highlighting restaurants and spa services for a couple.
Related Article: The Inference: Bad Bots, Big Bills and the AI Scaling Problem
Keep Humans in the Loop When Emotions Run High
Curtis said enterprises should not treat human involvement as a sign that an AI system has failed. While some requests, like adding baggage to an existing booking, can be handled autonomously, more sensitive situations still require people.
His company, said Curtis, looks for signals that an interaction should be escalated, including:
- Elevated stress or frustration in a customer’s voice
- Language indicating a serious or sensitive event
- Requests that reach the limits of the AI system’s knowledge
- Situations where a customer explicitly asks for a person
“If we determine that actually there's a high degree of stress, that we think actually this customer really needs to speak to a human, it doesn't matter how empathetic our bot is, they're going to want to speak to a human,” Curtis said, adding that even an empathetic AI response may frustrate someone who is already distressed. Effective systems need to identify escalation points early and transfer the conversation without forcing customers through repetitive automated steps.
“There is nothing worse than being highly stressed, having a very difficult experience at a time when you're meant to be enjoying yourself on holiday and then having to speak to an agent that's kind of repeating back to you parrot fashion what you've just told it, and you just need to speak to a person” Curtis said.
AI Agents Depend on Modern Enterprise Architecture
One of the largest gaps in enterprise AI is the distance between polished vendor demonstrations and the reality of production environments, Curtis said.
A voice agent may sound natural in a demo, but its value depends on what it can access behind the scenes. Enterprises need reliable APIs, accurate data and systems that can trace information back to its source.
“The big challenge, particularly for large enterprises, is that expectation gap between what the AI companies will sell you versus the reality of actually putting them into production,” Curtis said.
easyJet’s earlier investment in composable architecture made it easier to connect agents to existing services through APIs and Model Context Protocol servers. That foundation allows the airline to define what agents can access, what actions they can take and where their authority ends.
Organizations tied to monolithic platforms may have fewer options. They could be forced to rely on the AI products offered by their existing software vendors, deepening vendor lock-in and limiting their ability to experiment with other tools.
“That modernization piece is going to become more and more important as companies look to grow their adoption of AI,” Curtis said.
Give Agents Autonomy Where Failure Is Manageable
easyJet operates in a heavily regulated industry where safety shapes every technology decision.
Curtis said enterprises should begin autonomous AI deployments with use cases where the consequences of failure are limited.
“An example for us where I think we would be more than happy to trust an agent to do work on our behalf is where if they get it wrong, no one dies,” Curtis said. “And that sounds a bit dramatic, but when you run an airline, no one dying is a very key requirement for running our business.”
Managing onboard retail inventory is one example. An agent could analyze sales history, product availability and aircraft weight limits to decide how many items should be loaded onto a flight. A poor decision could reduce revenue or leave the airline with the wrong product mix, but it would not create a safety risk.
Before granting agents greater autonomy, enterprises need:
- Predictable and reliable outputs
- Clearly defined operational boundaries
- Complete records of agent actions and decisions
- Fast escalation when an agent reaches its limits
“Every action an agent takes, we need to be able to look back, see what it did, understand why it made a particular decision,” Curtis said.
This controlled approach allows organizations to increase autonomy gradually rather than turning entire workflows over to AI at once.
Related Article: The Inference: The Leadership Mindset Needed to Scale AI
Executives Must Calculate AI’s Total Cost
Curtis said one of the most important questions executives should ask is whether they understand the total cost of ownership for AI.
“I think there's a lot of hype around how AI is going to transform their businesses, how they can move much more quickly, but actually there is a cost element associated with AI as well,” Curtis said. “And I don't think we've fully quantified what that is yet.”
Model usage, token consumption, integration, oversight and infrastructure can all increase expenses. At the same time, AI may change the traditional build-versus-buy calculation.
Businesses often pay for enterprise software suites while using only a fraction of their features. AI-assisted development could make customized internal tools more economical, giving companies greater control over functionality and intellectual property.
“Certainly for us we are more and more leaning into that build versus buy conversation and how does AI really change that engagement and how we think about that total cost,” Curtis said.
While the opportunity is substantial, enterprises must evaluate AI as an operating model, architecture and cost decision — not simply another software feature.
