Information strategy – AI flight rebooking agent team 6

No ratings yet. The project and my role

After a long period of bouncing around and debating many ideas our team decided to develop an AI powered reboking agent to handle flight disruptions for KLM. There could be many reasons stated for why we have gone with this project but I firmly believe that the main reason was personal. All of us have experienced flight delays or cancelations and it is not a pleasant experince to have. We wanted to build something that directly addresses that stress. The reference point that acted as our main rationale for starting this project was the recent mass flight disruption in Shipol. In January, the airport has experienced a major snow crisis. This posed a logistical nightmare for the flight crews, staff that were forced to handle 3,200 flight cancellations and most importantly the 300,000 passengers that were left stranded. Standard rule-based rebooking systems already find another flight for 90% of the passengers but only taking into account the booking information. If the passenger writes that his/her father has surgery the next morning, or her child has another booking, then the rebooking system ignores the messages and finds the next available seat. Our solution was to allow a language model to scan all messages sent by passengers, but the same fixed rules would allocate the seats. Furthermore sensitive cases like minors, medical, or visa concerns would remain in the hands of KLM staff members. We have tested a rebooking agent that was supposed to run a simulation based on 550 synthetic passenger datasets booked on three different cancelled flights to New York, Nairobi, and London.

At the start, it wasn’t very clear what each of us would do since we had not known each other yet. Once the idea took shape, I took responsibility for the financial and economic feasibility given my academic background. In practice, this meant working out what the agent would cost and save compared with KLM’s current rule engine. Once one of our teammates ran the simulation, I had to interpret the numbers and create key takeaways from them to our main presentation as well as the pitch. This included the cost per passenger, a sensitivity analysis on how many cases go to staff, an estimate of what the agent would mean at the scale of the January crisis, and how a vendor would price such a tool.


Important decisions and iterations

The most important decision was how much freedom to give the AI. We chose to let the model only read messages, not choose seats. This made the agent easier to test and cheaper to build, because we didnt need to spend a lot of money on AI tokens.

Our first version failed our own decision rule and got a “Stop” verdict. The rationale behind it was that the rule engine scored around 89% and intially we wanted to have a 10% higher score, which was logically not attainable. We changed that target once and documented it. The second version rebooked 91–95% of passengers correctly, compared with 88–92% for the rule engine, and honoured far more of the needs that only appeared in messages. However, only the New York flight met both our 92% target for correct rebookings and our 80% target for honored needs. Nairobi and London fell just short. Mainly, there were too few alternative seats, especially to Nairobi. Since the agent was better than the rule engine everywhere but not good enough by our own standard, we decided to improve the agent and test it again rather than launch the original version. The third verion would reduce the effect of priotizing certain passengers, send impossible deadlines to staff earlier, and be tested on passengers that the agent has never seen. This is why we recommended a three month long shadow pilot, in which the agent would be used as tool for the staffs decision making process.


What suprised me me most during the work on the project

One major discovery was that flight disruption management is already heavily automated through systems provided by Amadeus. This required us to identify clear functional gaps in existing software, leading us to focus on reading unstructured passenger context as our primary value proposition.

The second major surprise concerned financial impact. Contrary to our initial assumption that the agent would reduce costs by mitigating EU261 compensation payouts, the simulation showed identical compensation totals 240,000 eur for both systems because EU261 liabilities depend on total delay rather than preference satisfaction. In fact, accommodating additional passenger preferences added compute expenses of up to €9 per passenger (120 euro per flight). I also had to correct my own assumption that about 10% of cases would go to staff. In reality the simulation measured almost 30%, mostly because of mandatory human cases and their travel companions.

Finally, because KLM is an airline rather than a software company, we neede to establish a defensible competitive moat. This was reflected in our slides that highlighted KLM’s proprietary internal assets, such as: customer data, fleet infrastructure, and Star Alliance partnerships for real-time seat availability.


How feedback and evidence changed my work

Feedback from our professor led us to drop the efforts in cost cutting through the EU261 reimbursment regulation. This has significantly changed the direction of our work from priotising financial preformance to focusing on improving the service itself.

Throughout the course lectures many of the sessions have impacted our work and prompted us to rediscuss our approach. In session 1 through through Furr & Shipilov (2019), “Digital Doesn’t Have to Be Disruptive,” we decided to build a targeted module on top of existing software infrastructure rather than replacing the entire legacy systems. This has significantly reduced capital expenditure and operational risk while keeping financial predictability. Additionally, applying framework concepts from Weill and Woerner reinforced our recommendation to preserve an omnichannel delivery model for customer support.


What I have learned from collaboration

Being a part of the Dutch education system for 3.5 years, team collaboration is nothing new to me. However, this project was unique because team members came from different institutions and academic backgrounds. Combining engineering expertise with economic and financial analysis significantly improved the depth and quality of our final outcome. This has proved to me that collaboration in teams with diverse, specialized knowgledge backgrounds results in a better quality of work.


How did the project change my understanding of AI business strategy?

Before starting this course, as well as at the beginning of the lectures, my approach to AI was fairly negative. I was strongly concerned that artificial intelligence would take a lot of jobs, especially in the white collar sector. Now I see that throughout our project, we have not only not eliminated people working, but only improved their quality of work and the service itself. The research and results from this project made me hopeful about the future and ultimately how AI could potentially contribute to society as a whole if utilised correctly.