Sanchit Juneja has spent nearly two decades building data and AI products in Israel, Singapore, the Netherlands and the U.S., with stops at National Instruments, AppsFlyer, Gojek and Booking.com. At Booking.com, where he was director of product for data and AI, he spent five years on the company’s data and tech modernization program. His team built the horizontal platforms for data, machine learning and AI from scratch, and customer-facing groups like flights and hotels relied on that foundation. The work included Booking.com’s LLM ecosystem and an early AI trip planner. When adoption of the trip planner fell short, his team saw that the chat interface users interacted with was holding it back as much as anything in the models underneath. Juneja now leads AI-driven transformation at Tripadvisor.
He explains to us the context layer in plain terms, argues that agent quality will depend on how well a company has modernized its data, and offers advice for a product leader handed an AI mandate on what to ignore during their first six months.
You’ve built data and AI products in Israel, Singapore, the Netherlands, and the U.S. What did each market teach you about shipping AI that the others didn’t?
When you’re young and ambitious, you have the option to follow wherever opportunity takes you, and I was lucky enough to get a chance to work across these countries and learn from the different cultures. Each country offered something new, be it The US’ ruthless meritocracy, Singapore’s seamless efficiency, Netherlands’ business pragmatism, or Israel’s infectious energy.
For example, working at National Instruments inculcated in me my abiding love for deep tech which has carried on in my various roles throughout my career; working with Appsflyer taught me to scale a SaaS unicorn from the ground up; whilst working at gojek taught me how emerging market startups scale
What was uniquely common across all these markets was their embrace of tech-driven value–all of these countries pride themselves on supporting tech startups and have created marquee companies, so it was great to have lived and worked across, and to imbibe those learnings in my personal ethos.
At Booking.com you led one of the largest LLM ecosystems in travel. What did that look like organizationally, and what was the first product decision that turned out to matter most?
A lot of e-commerce orgs startup as a hotchpotch where just getting things done is the priority and no focus in put on the right ways of doing things (rightly so) because capturing the market and user’s share-of-mind is so important. This is where B.com was in the 2000s
By the time the 2010s were at an end, the meteoric double-digit growths had crawled to single digits, and a lot of young hungry startups had cropped up, hungry to dethrone the king (which was us). It is at this time, B.com realised that the next set of growth would only happen if dedicated investments are done in Data and Tech Modernization; this was the 5-year plan I was part of in my journey at the organization.
As is the case in most of these multi-year, highly ambitious programs, the first choice is the simplest yet the most pivotal. How do we untangle the “data swamp” and create seamless data products that would set up the org for success? This is where a vertical-horizontal view of the org was agreed upon, where verticals are groups which the customer directly interacts with (flights, hotels, etc.) and these product groups ruthlessly focused upon domain-specific problems, and horizontals were what my team and I focused on setting up from scratch–the platforms, which create horizontal strategic capabilities on Data,ML,AI to hyper boost the growth of the org.
You talk about the context layer as the piece most teams are missing. Can you define it in plain terms and explain what breaks when it isn’t there?
Context is a much-abused term in the industry right now. Every org is “pivoting” to become a context org, as it is the in thing right now, but, let’s focus in the basics, which will help us define what context layer is in the simplest terms.
The data standardization worldview for big data existed and was widely accepted as the evolution of Big Data in the tech ecosystem before the advent of AI. In this world, Data is defined and maintained by an org in a Canonical Layer so that data can be treated with the same sanctity we treat products (and thus get the same repeatability of experience). In the data consumption part of this world, a corollary of ML and Analytical standards came up to afford repeatability of use. The catalog layer on top of these 2 consumption paradigms with these Products defined and cataloged is what is called as the Semantic Layer.
The Context Layer builds on top of this. Semantic Layer is only aware of Product schemas and what they do to afford easy discovery and reuse, but what if we could add operational awareness on top of the semantic layer, give depth of usage, frequency, unique user insights tying up the metric layer to other user signals (typically highlighted as a data structure in a graph DB) so that a powerful querying engine on top of this data ecosystem–an LLM–can create unique user products using this capability–this is the Context Layer.
What would happen if this layer isn’t there? Say you’re building an AI trip planner aiming to personalise a users’ journey end to end, without the context layer in mind you would lose sight of the user, start sending him spam on stale data, and lose sight of the further bookings a user will do down-the-line.
A lot of companies got an LLM pilot working in 2024 and 2025 and then stalled. From what you saw at Booking.com, where does the pilot-to-production gap usually open up?
Keeping true to the first principles is the secret sauce. `The biggest areas of stall occur when something is being built not to solve a user pain-point for other reasons (say tech fascination, blind hypothesis, etc.). That is why high discipline in what problems you aim to solve, and treating AI as a way of solving user problem, not as a magical panacea, is what keeps the org grounded
Let me cite a real-world example–AI trip planning end-to-end has been the most sought after use-case for a travel app. Being able to get a virtual concierge or travel agent who knows you and can suggest you things of delight, is the dream. When we started out, we captured this dream under a complex layer of chained ML models, with an LLM sitting on top–being totally myopic to the User Experience at the top, which was the most limiting factor of them all–a simple chat interface. Needless to say, the adoption of the product wasn’t as streamlined as expected, but that made us realise the key gap in what we need to achieve–thus a renewed focus on dynamic frontends was realised.
How does technical product management change when the thing you’re managing is probabilistic? What do PMs coming from classic software get wrong at first?
Every investment super cycle has had a lot of waste, but it has always changed user behaviour. The AI super cycle is no different.
Crux of product management is to be the user’s voice and grounding the big org machine towards user benefit. What is changing now is that the user might not be the one interacting with our product, it might be an agent, or an LLM directly or a new form factor itself. How do we evolve our product sense then? If an agent is buying from an agent, do we need to invest in the optimized frontends we have perfected over multiple iterations? Those are some of the key unsolved problems a lot of VC money is flowing to.
I don’t have a perfect answer to this, no one does. What has always helped in previous super cycle is to keep the user at the center of all our decisions. This should be what we do in this cycle as well.
Travel was one of the earliest consumer categories to put LLMs in front of users. What has changed in how people search for and plan trips, and what hasn’t moved as much as the industry expected?
This area is highly dynamic. The first iteration of our products aimed to marry our well-hones traditional ML expertise with the user-intent capturing capabilities of the LLM–this is what AI trip planners of the world came to. But, as we discussed above, this planner still relies on the human to do the rigour of the search, what is changing now is that a human would have multitude of Agents in his service who will do this job for him/her. This changing paradigm is what we are aiming to solve for in the next iteration of our products. Look at how Airbnb is looking to launch on AI lab on evolving user experience (dynamic frontend), and how the founders of Kayak are floating a new AI experience startup, and you will find out where the wind is blowing.
There’s a growing view that agent quality in 2027 will come down to data infrastructure. Do you agree, and what should a company fix first?
This has always been the case; your data consumption quality depends on your data. This was true for Traditional ML models and is true for SOTA ML models as well. Agents are abstractions built on top of these SOTA models, so it is compounding truer for these as well.
There is no other way around it but to focus on a thorough Data Modernization of your org, if you have not done so. If you’re just sitting on an old archaic system, and not harnessing the power of your data ecosystem, prepare to be disrupted.
For a product leader who has just been handed an AI transformation mandate, what would you tell them to ignore for the first six months?
Ignore the hype is always good advice. Getting swept up in the glamour of cool things and the ever-present information bombardment can evoke a sense of FOMO. This is what we are aiming to avoid–be disciplined, have a plan, execute that plan, listen to user feedback, adjust according to feedback, rinse and repeat. This is what we plan to do!
