Air France call-centre’s move to natural language speech recognition

The idea of improving customer service with increased automation might seem counter-intuitive. Indeed, the existence of websites such as, which outlines how to skip automated call handling systems and go straight through to an operator, might suggest that some people will never be happy interacting with an artificial agent.

However, a multi-year call-centre optimisation project by international airline Air France KLM, which incorporated natural language speech recognition, stands as one example of how a more sophisticated generation of automated systems can not only cut business expenditure – but improve customer experience too.

Before this undertaking, Air France’s call centres were, by its own admission, disorganised. It had 11 different customer service lines; and its six call centres in France were not connected, with the result that one could be overstaffed while others were being overwhelmed with calls. The upshot was long customer waiting times and misrouted calls.

Two technological improvements relieved this predicament – the first deceptively simple. Instead of allocating calls to discrete queues based on intelligent routing, the new system put them all in a single routing queue. This might seem less sophisticated, but in fact it improved waiting times.

Previously, a call might be added to a queue that has only one call in it, but that call might last 20 minutes and the new caller would effectively be stuck behind it. A queue system that distributes calls across all possible resources reduced the average waiting time by 40 seconds.

The introduction of natural language speech recognition, which permits callers to use whole sentences rather than simple, single word utterances, was more complex, and the organisation was sceptical at first. “The benefits of the fluidity of language that this technology allows were not obvious for us,” recalls head of remote sales, Florence Désert.

But after trials, those benefits soon became clear. The new speech solution allowed callers to utter actual sentences, which in turn allowed the system to divine their intention far more accurately. In time, this led to a 40% reduction in call misrouting.

It also allowed greater automation of information services and customer service functions such as flight reservations, so call centre agents were freed up to concentrate on cross-selling supplementary products. “We’ve seen a large rise in conversion rates of service calls into sales,” says Désert.

But the real test of the new system came on 12 April 2007, when French air traffic controllers went on strike, causing massive delays to the country’s flight schedules. This precipitated a flood of flight information calls, with the percentage of calls from passengers seeking information on their flights growing from 10% to 25%.

Because these calls could be automated, and because the queuing system had been optimised, the extra calls were relatively comfortably accommodated. “Callers still had good accessibility,” says Désert. “If it had happened a year before, it would have been a complete mess.”

Pete Swabey

Pete Swabey

Pete was Editor of Information Age and head of technology research for Vitesse Media plc from 2005 to 2013, before moving on to be Senior Editor and then Editorial Director at The Economist Intelligence...