← Multimodal Sources 🕐 6 min read
Multimodal Sources

How Booking.com Built AI that Converts Millions

> We stumbled into agentic architectures before agentic architectures were a thing because the first use case we started with was conversational recommendations.

Show: Beyond the Pilot · Publisher: VentureBeat · Host: VentureBeat editorial

Episode URL: https://www.youtube.com/watch?v=h_tcfmz_tlE

Publish date: 2026-04-13
Duration: 2748.0s
Default source credibility: HIGH — Named F500 practitioners on-record with production metrics. VentureBeat editorial vetting. Treat vendor-sponsored segments as MEDIUM.

  • Booking.com uses an agentic architecture to handle billions of daily predictions, evolving from simple ML to complex, multi-tool systems.
  • Natural language search revealed unmet customer needs, like a “hot tub” filter, providing direct signals for product development.
  • Using LLMs for customer service topic detection doubled accuracy, freeing human agents for high-value interactions and boosting retention.
  • Their tech stack includes an orchestrator, moderation layers, and specialized agents, balancing large batch models with smaller, low-latency models for real-time tasks.
  • The build vs. buy strategy focuses on owning core travel expertise (fine-tuning, RAG) while avoiding irreversible, “one-way door” decisions on foundational tech.

Extracted quotes

# Credibility Speaker Org Timestamp Topic Quote
1 HIGH Pranav Pathak (Director of Product Machine Learning) Booking.com 09:13 12-agent-workers We stumbled into agentic architectures before agentic architectures were a thing because the first use case we started with was conversational recommendations. One of our core philosophies was we don’t want to use a generic recommendation system. Whatever we recommend as Booking should be bookable on Booking.com, the price should be right, and the recommendations should be right within the context of the customer. So we started with an architecture of you have to call a tool if this is the intent you detect.
2 HIGH Pranav Pathak (Director of Product Machine Learning) Booking.com 03:17 02-corporate-tools When we did the free text filters thing, the most popular filter that our customers used was hot tubs. We didn’t actually have hot tubs as a filter on the platform. So this was our cue around, this is where demand is, this is what we want to build, and now the hot tub filter is live.
3 HIGH Pranav Pathak (Director of Product Machine Learning) Booking.com 04:52 12-agent-workers We’re seeing after LLMs, a 2x increase in topic detection there. And if we can get the topic right, we can get the tooling right. We can increase the number of self-service that happens on the platform, which means we can actually make more of our CS agents available to have actual conversations with customers.
4 HIGH Pranav Pathak (Director of Product Machine Learning) Booking.com 15:37 07-adoption-challenges At this point with everything that’s happening with AI, we are a little bit averse to walking through one-way doors. We want for as many of our decisions to be reversible as possible. We don’t want to get locked into a decision that we cannot reverse two years from now. So anything that forces us through a one-way door is a lot more scrutiny than anything that’s a reversible decision.
5 HIGH Pranav Pathak (Director of Product Machine Learning) Booking.com 16:30 02-corporate-tools We would rather use a very large model that then is much slower, but we’ll do that as a batch job because we can cache all of that data. But then for things like search or for recommendations, we want to be fast. So there we’ll use much smaller models, but then much more task-focused. So a model for intent extraction, a model for search query understanding.
6 HIGH Pranav Pathak (Director of Product Machine Learning) Booking.com 14:06 07-adoption-challenges Is this something that is the core of our business that we are specialists in, that we know how to do better than others? If yes, then we’re going to do that. Versus is this something that every business needs to some extent and someone else is going to have the specialization for doing that better than us? Then we don’t want to venture into that territory.

Per-quote detail

1. Pranav Pathak — Booking.com (09:13)

We stumbled into agentic architectures before agentic architectures were a thing because the first use case we started with was conversational recommendations. One of our core philosophies was we don’t want to use a generic recommendation system. Whatever we recommend as Booking should be bookable on Booking.com, the price should be right, and the recommendations should be right within the context of the customer. So we started with an architecture of you have to call a tool if this is the intent you detect.

  • Credibility: HIGH — Practitioner describes the specific business constraint (bookable inventory) that led to their early adoption of an agentic, tool-using architecture.
  • Topic tag: 12-agent-workers

2. Pranav Pathak — Booking.com (03:17)

When we did the free text filters thing, the most popular filter that our customers used was hot tubs. We didn’t actually have hot tubs as a filter on the platform. So this was our cue around, this is where demand is, this is what we want to build, and now the hot tub filter is live.

  • Credibility: HIGH — Practitioner provides a specific, tangible example of how natural language input revealed a product gap that was subsequently filled.
  • Topic tag: 02-corporate-tools

3. Pranav Pathak — Booking.com (04:52)

We’re seeing after LLMs, a 2x increase in topic detection there. And if we can get the topic right, we can get the tooling right. We can increase the number of self-service that happens on the platform, which means we can actually make more of our CS agents available to have actual conversations with customers.

  • Stat: 2x increase in topic detection accuracy after implementing LLMs
  • Credibility: HIGH — Practitioner provides a specific metric for a production system (customer service topic detection) and explains the business impact.
  • Topic tag: 12-agent-workers

4. Pranav Pathak — Booking.com (15:37)

At this point with everything that’s happening with AI, we are a little bit averse to walking through one-way doors. We want for as many of our decisions to be reversible as possible. We don’t want to get locked into a decision that we cannot reverse two years from now. So anything that forces us through a one-way door is a lot more scrutiny than anything that’s a reversible decision.

  • Credibility: HIGH — Practitioner articulates a clear strategic principle for technology adoption in a rapidly changing field, prioritizing flexibility and avoiding vendor lock-in.
  • Topic tag: 07-adoption-challenges

5. Pranav Pathak — Booking.com (16:30)

We would rather use a very large model that then is much slower, but we’ll do that as a batch job because we can cache all of that data. But then for things like search or for recommendations, we want to be fast. So there we’ll use much smaller models, but then much more task-focused. So a model for intent extraction, a model for search query understanding.

  • Credibility: HIGH — Practitioner describes a specific architectural choice, trading off model size and latency based on whether the use case is real-time or batch.
  • Topic tag: 02-corporate-tools

6. Pranav Pathak — Booking.com (14:06)

Is this something that is the core of our business that we are specialists in, that we know how to do better than others? If yes, then we’re going to do that. Versus is this something that every business needs to some extent and someone else is going to have the specialization for doing that better than us? Then we don’t want to venture into that territory.

  • Credibility: HIGH — Practitioner provides a clear, actionable framework for making build-vs-buy decisions for AI capabilities.
  • Topic tag: 07-adoption-challenges

Extracted 2026-04-13T23:36:23 via scripts/podcast_mine.py (Gemini gemini-2.5-pro).