AI Agents for OTAs: Build or Buy?

If you run an OTA and you are weighing whether to build your own AI travel agent or buy one, the most useful data point of 2026 is what Expedia did.

On 31 July, Expedia Group acquired Layla, a Berlin company founded in 2023, to accelerate its AI trip-planning and booking strategy. Expedia has thousands of engineers, the deepest traveller data in the industry, and more reason than anyone to own this capability outright.

But still, they bought.

That does not make building wrong. It means the build decision is harder than it looks from the outside, and the company best equipped to build chose not to start from zero.

‍

Why building a conversational agent in-house seems cheaper

Spoiler alert, it’s not. Every in-house AI project starts with a demo that seems to work perfectly fine.

Someone connects a language model to a spreadsheet. It looks impressive. It is also the easiest mile of the work, and it creates a dangerous impression: that the hard part is behind you, and now it’s a matter of time until it will be ready to go live.

Our engineering team puts it plainly: a prototype is easy. Making a conversational AI agent in travel work well, at scale, while optimising operating costs is hard. The distance between the prototype and being ready to go live is where most AI projects end up costing three times more than originally scoped or fail before helping a single traveller.

Most build business cases are written after seeing the demo. But the demo is the easy part. Everything that follows it is what the budget should actually reflect.
‍

The first version of any conversational AI agent is not perfect at all. Not because the team didn’t do good work, but because the agent needs feedback to learn and correct, and most often teams don’t plan and allocate time for this stage. Most leaders expect the implementation to be plug-and-play.

Language is complex and varies a lot based on your target audience. Travellers talk differently and will ask twisted, oddly specific and sometimes fragmented questions that you cannot anticipate. We saw this in the early days of Maya too.

‍

What actually goes wrong

The finish line moves

Normal software has a specification. You build to it, ship, and you're done.

An AI agent has no specification. It has an accuracy level, and you can’t know which one you need until real customers are talking to it. Teams set out to build a chatbot and discover, around month four, that they have started a permanent quality operation instead.

Which means the project you approved and the project you now own are different projects, and only one of them had a budget.

Someone has to keep up with evolving technology

The best model for a given cost is a moving target, and it moves in weeks. A model that was the obvious choice last month is often no longer the efficient option. Somebody has to notice that, evaluate the replacement, and migrate.

That is not a launch task. It is a standing job, and it doesn’t appear on any project plan.

No team owns the AI implementation in 6 months

This is the one that ends projects, and it is worth being specific about how.

Nothing dramatic happens. The launch goes fine, and the team that built it moves on to the next project, because that’s what teams do. But the person who was going to keep an eye on it has a quarter to close. Wrong answers start accumulating, and someone in customer service starts telling colleagues not to trust it. Six weeks later, it’s switched off, with no analysis behind it.

We spoke recently with a travel research firm that had built on a general-purpose chatbot platform before coming to us. Two things stopped them: they never reached an accuracy level they were comfortable with, and the work of building, monitoring and improving it consumed time that nobody in the company actually had.

The project didn’t fail on a technical decision. It failed because the maintenance never fit into anyone's week.

The bill grows in areas you can’t see

An agent that is not carefully built spends far more to produce the same answer than one that is.

A typical operator's availability API returns an enormous response. Pass all of it to the model to answer "is there space on the 14th" and you pay for every word, on every question. Trigger three calls where one would do and you pay three times, and your traveller waits three times as long.

Then there are the conversations that spiral, a small number of chats where the agent is doing something it was never meant to do, quietly consuming a real share of the month.

None of this shows up in a pilot, because a pilot handles ten conversations a day.

When building is the right call

We’d rather be transparent, so here is the honest version. Build if you have a dedicated team of 5 to 8 developers and quality testers 1 project managers  and a solid infrastructure that can run and maintain it permanently, not just deliver a project.

Build if your travel business model is unlike anything on the market and you don’t have direct competitors. If two of those two are not true, maybe your team should consider working with experts in AI for travel.

What does buying actually look like?

Buying does not mean licensing a chatbot and switching it on. That expectation is where most of the disappointment starts.

What you are buying is two things: software that builds and runs a conversational AI agent on your own data, and the team that operates it. Working with a specialist such as Maya moves four responsibilities off your roadmap.

The travel data model already exists. Destinations, trip types, seasons, availability, amenities, and the relationships between them are modelled before your project starts. Your team connects your data to a structure that already understands travel, instead of designing that structure first. This is the part in-house projects consistently underestimate.

Implementation happens with your team, not by your team. Your people supply data access, product knowledge, and the decisions about tone and policy. The engineering work, prompt development, feedback processing, connecting your travel tech tools, and deployment sit on the provider’s side.

‍

The feedback on answers is ours. Every conversation is monitored. Wrong answers get labelled, knowledge gaps are flagged, and the conversational AI agent gets better day by day. Your team reviews and gives feedback, usually a few hours a week in the first months, then consistently less.

Your team does not run the feedback processing but retains the learnings and instructions.

We make model decisions so you get the most efficient model at the best cost.  We choose which AI model runs each part of the conversational AI agent, and we change it when a better option appears. That decision never becomes an item on your engineering roadmap.

What stays yours. Your data, your traveller relationship, your commercial terms and your learning and instructions layer. The conversational AI agent runs in your branding, on your website, and answers only from sources you control.
‍

What are you actually trying to own?

That is the better question, and it usually dissolves the first one.

Most OTAs do not want to own a conversational AI system or any piece of software. They want to own the traveller relationship, the supply, and the margin. The agent serves those. It is not one of them.

Expedia's decision reads the same way. They didn't buy Layla because they couldn't build a conversational AI trip planner. They bought it because time and accumulated product learning were worth more than building it themselves.

Three questions before you decide

Who in your team owns the conversational AI agent in month seven? Name a person, not a team. If you can’t, you have already chosen to buy, you just haven’t written it down yet.

What accuracy would make you comfortable putting this in front of your customers, and how will you measure it? If nobody can answer, the project has no finish line and won’t get one later.

What does year one cost, including your own team's hours? Not the licence. The hours. Put them in the same table or the comparison means nothing.

Answer those three honestly, and the build-or-buy decision usually makes itself, sometimes in favour of building. Either way, you will know why, which is more than most teams have when they start.

I'm Ben (Benjamin) Manzi, co-founder and CCO of Maya. We've built over 100 conversational AI agents for travel agencies, OTAs, tour operators and experience providers since 2023.

Not sure where to start with AI? Book a consultation call with one of our AI experts. No cost, no commitments, just value for you and your team.

‍

Frequent Asked Questions
‍

1. What is the "build or buy" decision for a travel agency or OTA AI agent about?

It's the choice between developing a conversational AI agent in-house or licensing an AI agent solution from a specialist, such as Maya. Building means structuring the data model from scratch, engineering, quality monitoring, and permanently switching AI models. Buying means a vendor runs those responsibilities while the OTA keeps its data, traveller relationship, branding, and commercial terms.

2. Why do in-house conversational AI agent projects often cost more than planned?

A demo that connects a language model to a spreadsheet is the easiest tenth of the work, so teams underestimate what remains. Making a conversational AI agent work well while controlling costs at scale is hard, and the gap between demo and production is where most projects fail or end up costing three times more than originally scoped.

3. Why do in-house AI agents often get switched off within a year of launch?

Unlike normal software, an AI agent has no fixed specification. Teams learn the accuracy level they need only once real customers are talking to it, often around month four. Ownership frequently lapses after launch: the building team moves on, wrong answers accumulate unnoticed, and roughly six weeks later the agent is quietly switched off with no analysis behind the decision.

4. What hidden costs make a conversational AI agent more expensive than a pilot suggests?

Costs hide in places a small pilot never reveals: passing an operator's full, oversized availability response to the model for every question, triggering three API calls where one would do, tripling cost and traveller wait time, and a small number of conversations that spiral into unintended, resource-heavy exchanges. None of this shows up when a pilot handles only ten conversations a day.

Q6. What does buying a conversational AI agent from a specialist like Maya actually include?

Buying from a specialist such as Maya includes a travel data model already built for destinations, trip types, seasons, and availability; implementation done with your team rather than by it; a vendor-run quality operation that monitors and corrects every conversation; and vendor-managed AI model choices. In return, your team gives up direct control of the product roadmap.

Q7. When does it make sense to build a conversational AI agent instead of buying one?

Building makes sense only if you have a dedicated team of developers and quality testers who can run and maintain the agent permanently, not just deliver it as a project, or if your travel business model is genuinely unlike anything on the market with no direct competitors. If neither is true, working with a specialist in AI for travel is the more realistic path.

Q8. If I buy a conversational AI agent, do I keep control of my traveller data and relationship?

Yes. Buying a conversational AI agent from a specialist such as Maya keeps your data, traveller relationship, and commercial terms yours — the agent runs in your branding, on your website, and answers only from sources you control. What you give up is direct control of the vendor's product roadmap, not ownership of your data or customer relationship.

Q9. We tried a general-purpose chatbot platform and never got the accuracy we needed — is that a common problem?

Yes. A travel research firm that had built on a general-purpose chatbot platform ran into exactly this before switching providers: it never reached an accuracy level it was comfortable with, and the ongoing work of building, monitoring and improving the agent consumed time nobody in the company actually had. The project didn't fail on a technical decision — it failed because maintenance never fit into anyone's week.

Q10. I run a small OTA with no dedicated AI engineering team — should we build or buy a conversational AI agent?

Without developers and quality testers to maintain the agent long term, the article's guidance points toward buying rather than building. A specialist like Maya already has the travel data model built, runs the quality-monitoring operation, and manages AI model choices, leaving your team to supply data access and policy decisions instead of full-time engineering.

‍