--- title: "Hangzhou airport's AI hotline now takes a third of the calls, and complaints fell a fifth" date: 2026-11-03 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-app-hangzhou-airport-ai-hotline language: en --- # Hangzhou airport's AI hotline now takes a third of the calls, and complaints fell a fifth > Hangzhou Xiaoshan's 96299 line became China's first large-model airline call centre. The measured gains are real, and so are the gaps between them. A woman realises her bag is gone only after she has already left the terminal, taxi booked, train to catch. Every second of that uncertainty is logged somewhere in a call centre quality report. At Hangzhou Xiaoshan International Airport, the same story ended with a Hermès bag back in her hands 19 minutes after she first called 96299, worth about ¥115,000 (≈ US$16K / HK$125K). A cleaner found it in a restroom while a human agent on the line kept the caller calm. Nothing about that rescue was automatic. What changed underneath it was. The hotline is now run by a large model (大模型) architecture, and it quietly absorbs roughly a third of the traffic before a person answers. ## Key takeaways - Hangzhou Xiaoshan's 96299 hotline went live as an AI call centre in May 2025, the first in Chinese civil aviation to put a large model into a live phone line. - Connection rate rose from 83.82% before the upgrade to 95% after it, and in the first three days of trial operation the queue wait fell from 15 seconds to 3. - The vendor and the airport report 94.2% semantic parsing accuracy on flight-status queries and a 300% jump in response efficiency over the legacy system. - In the 2026 summer peak, the AI robot handled close to 30% of inbound calls, and complaints about the airport handled by the centre fell about 20% year on year. - What the model does not do is decide. Rule-matching, a standardised knowledge base and a human handover path still carry the hard cases. ## The hotline that asked you to press 1 The old system was not a modern call centre by any international measure. It launched around a decade ago and ran on multi-level key menus, without any support for a passenger simply speaking a question. Staff relied on their own memory and manual lookups across several separate systems, and flight, gate and baggage data lived in silos that a router could not reach. The replacement was not a quick swap. Preparatory research started in 2023, the project was formally approved in June 2024, and Hangzhou airport built it with Zhejiang Airport Digital Technology. The new hotline went live in the early hours of a May morning in 2025, with DeepSeek as the underlying model and a "perception, decision, interaction" service system on top. The architectural choice matters more than the model name. The voice agent runs a dual-model design with three fallback layers: a scripted multi-turn flow first, a curated standard-answer bank second, and the general model only as the last resort. Below that sits a three-tier aviation knowledge base, built specifically to constrain where answers can come from, which is the point where hallucination would be most expensive. ## What the measured numbers say The reported gains cluster around three figures, and they come from the operator's own data rather than an independent audit. - Connection rate: 83.82% before the upgrade, 95% afterwards. - AI automatic handling: 67.2% of interactions after go-live. - Semantic parsing accuracy on flight-status queries: 94.2%, with response efficiency reported as 300% higher than the legacy system. There is also a smaller number worth watching. In the 2026 summer travel peak, the AI robot took close to 30% of inbound calls and handled simple enquiries; human agents took more than 500 calls a day and concentrated on delays, cancellations and rebookings. Complaints about the airport handled by the centre fell by about 20% year on year, with on-site resolution running at 60% to 70%. The direction of travel is not in doubt. What is worth flagging is that these are operator-reported figures from a single site, produced by the airport and its vendor, with no outside audit attached. ## Where the machine stops Two design decisions do most of the real work, and neither is a model upgrade. First, the routing boundary was set in advance, using real historical call recordings. One rule the airport published is blunt: during a mass flight disruption, the system works from a scripted flow rather than open-ended generation, because a wrong answer about a cancelled flight turns into an angry passenger and another call. Second, the knowledge base is treated as an operational asset with scheduled maintenance, not a static file. Staff reviewed real call samples, extracted more than a hundred high-frequency enquiry scenarios, tested them one by one before launch, and set explicit rules for when a call is handed to a person. Calls are transcribed in real time and archived, so a supervisor can replay a recording by timestamp rather than by searching old logs. The bag case shows the shape of the division of labour. The model was never going to find a Hermès bag in a restroom. It was going to make sure the human who could, was reachable and had the record in front of them. According to Li An, chief scientist at BrainNet (脑机网), the test of a service hotline AI is not how fast it answers but whether the human on the other end gets a complete record in one glance. ## Honest limitations Every number above comes from Hangzhou airport and Zhejiang Airport Digital Technology, with no independent audit available, and the reporting mixes trial-run figures with full-season ones. Queue-wait numbers illustrate the problem: the airport's own May 2025 release and the civil aviation trade press both cite a drop from 15 seconds to 3 seconds for the first three days of trial operation, while a later civil aviation newspaper report puts the same metric at 35 seconds falling to 30 seconds over a longer window. Different measurement windows, same system. Treat any single queue number as a snapshot rather than a baseline. The scope is also narrow in ways that limit how far this travels. One airport, one hotline, one dominant disruption scenario. The harder customer-service problems — a passenger arguing about compensation, a medical emergency on the line, a lost-traveller case with legal ambiguity — are exactly the ones where the scripted path and the human escalation both have to work, and the public reporting does not break out how often that handover succeeds. Whether the same architecture would survive a hub airport with ten times the traffic during a system-wide grounding is untested in public. ## What readers can do now - If you call an airport hotline, ask a question in your own words rather than pressing through the menu. At least one major Chinese airport now handles natural-language queries and routes to an agent when it should, and the manual fallback is deliberately kept open. - Keep your own ticket reference and the exact wording of what the agent tells you. When the call is transcribed and archived, which the published system does, a precise record is what makes a later claim or appeal tractable. - Do not assume a voice agent can improvise during a crisis. Published airport practice treats mass disruptions as a scripted path, so confirmation of a cancellation or a gate change should come from an official channel, not from a phone call alone. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-hangzhou-airport-ai-hotline Free to quote with attribution and a link to the original.