2026-11-10 · ← News
Singapore Airlines doubled chatbot satisfaction by building the service around the model
Singapore Airlines says moving its Kris chatbot to OpenAI models doubled CSAT. The result rests on airline data, validation and human escalation around the API, not on the model alone.
The image could not be loaded.
Singapore Airlines says customer satisfaction with its Kris chatbot doubled after it replaced hand-mapped dialog flows with OpenAI models. The case shows that customer support gains come from the service built around a model.
Kris left scripted dialog trees behind and doubled its CSAT
The old Kris required teams to map each topic and conversation path in advance. The new version uses the OpenAI API, understands natural-language requests and retrieves information from Singapore Airlines systems. The airline says the change doubled the chatbot's CSAT.
The same technology summarizes calls for service representatives. The case study says this cut average handling time by about 10%. Flight Recommender, which turns a natural-language request into travel options, nearly tripled its monthly users over the past year.
Enterprise value comes from connecting data, rules and human review
The airline also runs OpenAI models inside Jarvis, its governed internal AI platform, which it says is used monthly by more than 60% of employees. It reports productivity gains of up to 20% in selected software-development activities. These figures come from Singapore Airlines and OpenAI, not an independent audit.
For support teams, the service architecture matters more than the model name. Orchestration, restricted data access, answer validation and escalation to a person turn a general chatbot into an operational tool. Those layers also need testing whenever fares, baggage rules or schedules change.
The case study reports outcomes but leaves the full bill off the page
A twofold CSAT gain is notable, but the published material does not provide the baseline, sample size or the cost of inference and human review. Nearly tripling Flight Recommender users does not reveal how many recommendations became bookings.
OpenAI is both supplier and publisher of the case study. The numbers indicate direction rather than a transferable benchmark for other airlines. Handling personal and travel data also deserves scrutiny because a wrong answer here carries more weight than a poor email draft.
Complaints, escalations and completed bookings will determine the value
The next useful evidence must come from operations: resolution without an agent, incorrect-answer rates, human escalation and conversion from recommendation to booking. CSAT is helpful, but it captures only part of the service.
If Singapore Airlines sustains quality through disruptions and policy changes, it will offer a credible pattern for other regulated services. If it cannot, a smoother conversation will merely move a harder complaint to the airport desk.
Lilith's verdict
Passengers do not grade the model. They remember whether the chatbot explained a delay correctly and called in a human before the gate closed. Singapore Airlines wins only when that scene still works on the day the schedule falls apart.