A senior engineer from Booking.com stood on stage at QCon London and did something rare: talked about AI integration without the hype. No significant breakthroughs. No paradigm shifts. Just the messy, practical reality of weaving AI into a 20-year-old system that handles millions of bookings every day.
The talk mattered because it answered the question most businesses are actually asking: how do you add AI to infrastructure that already works?
The Challenge of Legacy Systems
Booking.com is not a startup. It is a two-decade-old platform built on layers of decisions made when AI meant basic recommendation algorithms and machine learning was a niche academic field. The architecture was designed for reliability and scale, not for plugging in transformer models and expecting everything to work.
The engineer's insight was refreshingly honest: integrating AI into established systems is not about replacing what exists. It is about finding the places where AI genuinely improves outcomes without breaking what already works.
That sounds obvious until you consider the temptation to rebuild everything around new capabilities. Booking.com resisted that urge. Instead, they identified specific friction points - search relevance, pricing optimisation, customer service routing - and applied AI narrowly, measurably, and iteratively.
What Actually Worked
The talk outlined several practical lessons that apply beyond Booking.com's specific architecture. First: start with data infrastructure. AI models are only as good as the data pipelines feeding them. Booking.com spent significant time cleaning, standardising, and validating data before training models. Unglamorous work, but essential.
Second: measure everything. They did not deploy AI features based on promising demos. They A/B tested relentlessly, compared performance against existing systems, and killed features that did not meaningfully improve metrics. This is harder than it sounds - it requires cultural buy-in to abandon clever solutions that do not deliver results.
Third: avoid black boxes in critical paths. Some parts of Booking.com's system need to be explainable - pricing, fraud detection, compliance-related decisions. They used AI in those areas cautiously, maintaining fallback logic and human oversight where stakes were high.
The Human Element
One of the more interesting points was about team structure. Booking.com did not create a separate AI division that operated in isolation. They embedded machine learning engineers directly into product teams. That meant AI capabilities evolved alongside product needs rather than being handed down as finished tools.
This organisational choice shaped how AI got integrated. Engineers building booking flows understood the edge cases, the regional differences, the compliance requirements. When they added AI features, those constraints were already accounted for. The alternative - an AI team building models in isolation and handing them off - creates friction and misalignment.
What This Means for Other Businesses
Booking.com's experience offers a useful counterpoint to the "rip it all out and start fresh" narrative. Most businesses operate on legacy systems. Rebuilding from scratch is expensive, risky, and often unnecessary. The more practical path is selective integration - finding the specific problems where AI delivers clear value and applying it there first.
That requires discipline. It means saying no to impressive technology that does not solve a real problem. It means investing in unglamorous infrastructure work before deploying flashy models. And it means accepting that AI integration is a years-long process, not a single project.
The talk also highlighted something often missing from AI discussions: the importance of institutional knowledge. Booking.com's engineers understood their system deeply - where it was flexible, where it was brittle, where changes would cascade unexpectedly. That knowledge guided their AI integration strategy more than any external framework could.
For businesses looking at AI and wondering how to start, Booking.com's approach offers a roadmap: fix your data, embed AI expertise in product teams, measure ruthlessly, and integrate incrementally. Not significant advice, but in a field drowning in hype, practical guidance is worth more than bold claims.