I would however, argue that the picture changes drastically if we focus on solving a single, isolated problem, e.g. only booking business trips, no pizzas. To begin with, understanding requests from a single domain is an easier problem to solve. Although very impressive language models like Open AI’s GPT-2 have recently been released, enormous domain-specific amounts of data are required to train these models. Data for the travel domain – unlike unspecific text – is not easily available. Limiting yourself to a narrow domain means these demands can be reduced, given the fact that typical business travel requests are fairly repetitive. Another important aspect is the fact that users are not very tolerant of errors produced by machines. Human communication uses very subtle details. In order to have a good user experience, it is important to get these details right. While it may be acceptable for a machine translation to be slightly off (human intelligence is able to fill in the gaps or interpolate the errors), there is no tolerance toward these details in an automatically generated trip booking. If it does not match what has been asked for, the experience is painful or useless. Finally, it also makes a difference whether you put AI directly into the face of the customer or use AI to supercharge humans in performing their tasks.

Based on our experience at Comtravo, AI is an extremely valuable tool to improve the first 80% of our tasks, which is then combined with humans that can take care of the remaining 20%. While this is still far from being trivial, limiting the application of AI tech to a narrow domain increases the chance of coming up with a solution that actually works.

Over the next few years, I expect to see many new impulses in the travel industry that will leverage data, use analytics and machine learning and that will ultimately make booking a business trip as simple and convenient as ordering a pizza.