How an airport chatbot’s migration to LLM pushed automation to 98%
98%
of inquiries handled automatically after migrating to LLM (vs 89% in the rule-based model year-on-year)
12 237
active conversations between March 9 – June 9, 2026 (up from 11,573 y/y)
~2 seconds
average response time generated by the generative model – comparable to the rule-based model
Introduction
Wrocław Airport was the first airport in Poland to deploy a passenger service chatbot – our solution has been accompanying travellers for over 7 years, answering questions about flight statuses, directions and check-in procedures 24 hours a day, 7 days a week. During that time, the chatbot – initially operating on a rule-based model – handled tens of thousands of conversations and became a permanent part of the airport’s communication with passengers.
As conversational AI technology evolved, we jointly decided with Wrocław Airport to rebuild the chatbot and migrate from the rule-based model to a modern large language model (LLM). This was not a decision made overnight – for over 2 years we operated in a hybrid model, gradually expanding the scope of responses handled by the LLM and testing the new approach with trusted passengers. Only after confirming the stability and effectiveness of the new model did we make the decision to fully transition. The project – exceptionally complex, spanning multiple technological layers and external integrations – had one goal: to respond even more effectively to passengers’ growing expectations, without losing the stability and speed that had built trust in the tool over the years.
Our 8-year partnership with KODA shows us every day how technology can have a real impact on passenger experience. Since switching to large language models, Chatlot has gained remarkable resourcefulness, handling complex queries with ease. Our virtual assistant has become one of the key tools in building a great experience for travellers.
Solution
The new model understands the meaning of passenger queries even when they are imprecisely worded – and, crucially, it can recognise and handle several intents within a single message, rather than guiding the passenger through a rigid question-and-answer path.
We integrated the solution with flight search engine APIs and the airport’s internal systems, giving the chatbot access to an up-to-date knowledge base and allowing it to direct users precisely to the right information – from flight status, through check-in queries, to the newly added lost property reporting feature.
Although responses are generated by the model in real time rather than selected from a fixed pool of scenarios, model optimisation kept the average response time at around 2 seconds – comparable to the rule-based model. The passenger gets a bot that understands more, without any loss of conversational speed.
Implementation
Analysis of the existing solution and target architecture (August 2025)
We began designing the LLM architecture in August 2025. The task involved analysing nearly 7 years’ worth of rule-based chatbot features and scenarios, and planning how to migrate them to the new model. We defined feature priorities across three phases:
Phase 1, must-haves without which we could not go live – flight status with survey, flight search, directions/parking, airport alert keyword responses;
Phase 2, additional features to follow shortly after the switch – lost property reporting, airport services and venues;
Phase 3, features added in the days and weeks that followed – travel FAQ.
The key architectural challenge was designing a classifier built on two cooperating prompts – a new approach that required precise calibration to reliably route users to the correct responses.
Production and feature adaptation (November 2025 – January 2026)
From November 2025, work began on the new chatbot structure, prompt engineering and API feature development. The most critical features – flight status, flight search, security alerts – were rebuilt and adapted to the new LLM-based model. In parallel, we worked with the airport to prepare and update the knowledge base.
External integrations (December 2025 – January 2026)
The key change in integration was the switch to direct API queries to eSky. This required moving the integration logic from an external service directly into the bot’s structure within the platform. eSky queries demanded careful data formatting and iterative calling. At the same time, we verified the flight status notification system, which needed to work correctly with the new bot.
Testing and deployment (January – February 2026)
Internal testing covered verification of all features – flight status, search, knowledge base, alerts and lost property reporting. Before go-live, we ran a client training session on the new approach to bot management and prepared a detailed transition plan.
Post-deployment development
Following migration, we continue to develop the chatbot together – it is a living tool in a genuinely dynamic industry, requiring constant monitoring to deliver the best possible results.
Results
Comparing the same three months year-on-year (March 9 – June 9) allows us to precisely assess the impact of the migration, eliminating the influence of seasonal passenger traffic. And all of this against a backdrop of growing conversation volume.
Automation jumped from 89% to 98%. In the first three months of the new model, automation rose from 89% to 98%. Given that passenger traffic continues to grow and conversations with Chatlot are increasing, this is invaluable support not only for passengers themselves, but also for the team serving them.
Faster responses, fewer messages. The generative model can identify and combine several threads within a single passenger message, meaning users reach the information they need faster and without repeating themselves. Despite an increase in the number of conversations, the total number of messages exchanged fell from 202,593 to 189,004 – confirming that Chatlot’s core task of making it easier for passengers to get answers has been successfully accomplished.
No compromise on speed. Although responses are generated in real time rather than retrieved from a fixed pool of scenarios, the average response time remains around 2 seconds – the same as in the rule-based model.
New functionality quickly adopted by passengers. Within the first 3 months of launching the lost property reporting feature, 230 people had already used it – proof that passengers are happy to embrace new features when the chatbot is a reliable, efficient channel they can trust.
As an airport, we prioritise solutions that directly translate into passenger comfort and the highest quality of communication. Chatlot, powered by KODA’s generative model, has become the natural first line of mature dialogue with travellers. A 98% automation rate while maintaining a two-second response time proves that well-designed and well-implemented technology can actually exceed the expectations set for it. And through our partnership, we are once again setting the trends in the industry together.
Future development
Migrating to the LLM model opens up new possibilities for Chatlot that would not have been achievable with the rule-based model. The chatbot already operates multilingually thanks to the LLM’s capabilities – meeting the needs of international travellers. The long-term vision is for Chatlot to continue evolving as a multifunctional airport assistant – a single point of contact for passengers at every stage of their journey.
Wrocław is a pioneer in this space, creating one of the first commercial LLM deployments in an airport chatbot in the world. This experience lays the foundation for rolling out the solution to other airports.
Working on this project means more than simply delivering technology. The airport team is deeply committed to passenger experience, and our assistant supports them brilliantly in that mission – both in everyday operations and in moments of crisis. It proved its worth in challenging situations such as the pandemic and flood alerts, instantly delivering critical information to travellers.
