When I started the Information Strategy course, I had BIG reservations about AI. However, our AI Strategy Lab project, also linked to reservations, changed how I feel about AI in the end.
Planning and booking a trip can sometimes be annoying. What was meant to be an enjoyable process ends up becoming a frustrating one. Because you are busy jumping between websites, comparing flights and hotels, checking dates and cancellation policies and repeatedly entering the same information. I may have had reservations about AI, but the booking/reservation pains are enough to get me to be open about what AI can do.
Our travel idea was immediately interesting to me. However, I learned that relating to a problem personally does not automatically make it a good business opportunity. We still had to prove that the problem mattered beyond our own experiences and in which aspects AI added value.
We settled on Layla.ai as the business to focus on. I had never heard of Layla before and, honestly, wished I had. However, while we were looking into Layla, we found many flaws/processes that could be improved. Our initial idea was broad, including swipe-based discovery, personalised itineraries, and social travel. I liked the creativity of our initial idea, but it quickly became clear that we were trying to solve too many problems at once. Through iteration, we narrowed the idea to the gap between an approved itinerary and a completed booking. Rather than turning Layla into a different product, we proposed extending its existing planning capability so controlled Agentic AI could execute an already-approved itinerary.
A significant part of my contribution involved refining this problem and challenging whether we actually had evidence for it. At one stage, we considered statistics suggesting that around 82% of travel bookings were abandoned. We also found one customer review suggesting that a Layla booking had taken them a couple of weeks to complete. This review solidified why we proposed our solution, but it was just 1 customer review. Initially, both facts seemed useful because they supported the story we wanted to tell. However, one customer review could not represent Layla’s average booking experience, and the broader abondonment statistic did not prove that Layla itself had a conversion problem.
This made me realise how easy it is to start with a conclusion and then search for numbers that make it sound credible. We therefore looked for stronger evidence around customer effort, one-stop booking and digital friction. We stopped claiming that Layla had a proven conversion problem and instead argued that its planning-to-booking hand-off creates a plausible source of customer effort and potential transaction loss. One of my biggest lessons during this process was that a strategy can become stronger when you are willing to make a smaller, more defensible claim. Evidence should be able to challenge the idea, not simply support it.
We identified Layla as an Orchestrator because it coordinates different travel providers around what the traveller has approved rather than owning the underlying services. Layla was recently acquired by Expedia group so that gave it a more competitive stance as well against other incumbents like Booking.com. Applying concepts such as multihoming, price sensitivity, subsidy sides, and many sides also changed how I thought about pricing. Charging travellers an additional AI free when they already received free options from Layla itself could create friction exactly where we were trying to remove it, which led us towards a commission-funded model. At the same time, I learned that scalability and profitability are not the same. AI may reduce the cost of execution, but that does not mean every automated transaction created value.
The project also changed how I think about competitive advantage. Earlier, I saw Agentic AI itself as the differentiator. The future-market exercise challenged that assumption. If other OTAs and general AI assistants can eventually deploy capable agents, simply having an agent will not be much of a moat. For Layla, a more defensible position would need to come from traveller trust, supplier integration, reliable execution, and proprietary feedback loop learning. This taught me that technology can reshape competition without necessarily being the competitive advantage itself.
The prototype challenged another assumption that greater AI autonomy is automatically better. We developed a risk-weighted system in which human judgement increases with the consequences of the transaction. If a booking remains within agreed boundaries, the agent can do more. For example, if the price or itinerary changes materially, the traveller must approve again or the system should not only know how to act, but also when not to act.
When it comes to the general group assignment process, I found that collaboration was another important part of my learning. Towards the end of the assignment, I only realised then that sitting together and working on the project was much faster than everyone independently completing sections. We could challenge each other, identify inconsistencies, and make decisions immediately. Having worked professionally for more than seven years, I have high expectations around communication, presence, and shared responsibility. However, this also made me reflect on my own approach. I cannot assume that everyone shares my definition of good collaboration. In future projects, I would communicate expectations around availability and ways of working much earlier. Similar to how we discovered it, we cannot assume anything about AI/technology without figuring out its limitations first.
Looking back, my biggest takeaway is therefore not that AI is something to fear, nor that it should be used simply because it is available. AI strategy should begin with the problem rather than the technology. I started the project asking, “What can Agentic AI do?” and finished by asking much harder questions: Does this solve an important problem? Will customers delegate the task? Can the business capture enough value? What happens when competitors have the same capability? What could go wrong, and when should we stop rather than scale?
This is why our final recommendation was a controlled pilot rather than simply launching Agentic AI, just because it was technically possible. I still like the idea of making travel easier. Like being able to say, you know what I interact with on socials and what my Pinterest board looks like, plan my dream trip based on XXX budget and handle the annoying parts with the booking. What we presented is a starting point, to really amazing potential.
Turning frustration into a viable AI strategy requires more than a clever agent. It requires evidence, economics, trust, governance and a willingness to discover that your original idea might be wrong. The project ultimately moved me from fearing what AI might do, to thinking much more carefully about what we should allow it to do, and under what conditions.
