"I’m Improving the Digital Assistant"
An interview with Emmanuel Drosos
Sharing is caring. That’s also true at Deutsche Telekom. Why should every department do its own thing instead of sharing know-how and expertise to save resources? That’s exactly what makes Emmanuel Drosos tick. A computer scientist by training, his job is to help AI learn and get better at its various tasks so customers can benefit from the enormous synthetic brain power.
Tell us what your team is working on.
We run through various use cases that revolve around the question how different products can tap into our AI capabilities. Those capabilities can come from different vendors and can change quickly. What vendor A does well today might be provided in a better way next year by vendor B. That means we have to be able to switch gears fast, create new bundles of offerings and services for our entire product portfolio, particularly improving digital assistants.
So your team is a distribution hub for AI?
We aim to be the go-to place within Deutsche Telekom where people come to ask for AI: easy to use, adapted to and pre-configured for your needs, optimized and trained. Those basic topics have to be tackled in a smart way, and we have to be clear how people can access all AI capabilities in one spot. Are there functions I need time and again? Take anonymization of data before I hand it over to a vendor for processing. Learning is another topic that touches all AI skills. They need initial training and after that an ongoing learning process. That’s especially true for the things Deutsche Telekom wants to get out of AI.
That sounds pretty theoretical. Can you give us a real-world example?
In Austria, we have a big use case called “Voice of the Customer.” It’s about integrating voice data, for instance from Interactive Voice Response (IVR) dialogs when a customer talks to a digital assistant. We use two AI skills to anonymize and transcribe that data, and another one to infer the intent of the caller. All that information is fed into a big data system where it can be mined to boost the digital assistant’s smarts and usefulness.
How do you collaborate with the eLIZA Project?
We’ve been a part of this project from the start. All activities and customer interactions are gathered in a big data project in real time and are available for use cases. We combine that data to improve our digital assistant. The assistant itself is modular, consisting of a growing set of different AI skills. You can think of them as lego blocks. The way we have them play together, configure and train them -- that’s the know-how we’re building piece by piece through this project. It’s our job to make sure the pieces fit together and meet to a certain standard so you can address and combine them. For starters, we’re doing this for T-Mobile Austria, but everything we do is purpose-built so it can be rolled out to other national subsidiaries of Telekom.
Should we picture this kind of work like an interface or API?
Everything we build will be available as a Rest API, which is today’s state of the art API. Our architecture complies with our architecture chapter and is based on modern principles such as microservices and Docker. We’re quite far along with our proof of concept for T-Mobile Austria, with several parts already in production. Our approach also provides a template or vision for the enterprise as a whole.
Why do you work with microservices?
We like to leverage current and proven technologies. We have the luxury of not being too beholden to legacy systems, unless we want to integrate existing systems into some use cases like in Austria. There, we’ve been able to use data from a host of eCare channels in real time since last year. But if we’re not tied to a pre-existing set-up we try new things, including the technologies we use.
Tell us more about the process to develop an AI skill.
That’s a wide field where various squads work on their own respective topics. One, for instance, is investigating and evaluating emotion analysis and developing it into a proof of concept. In the end, we identify one or two vendors who are the best fit for our needs. We integrate those into my team for another proof of concept. The vendor takes care of the technical integration, making sure all the parts work together, and we provide the overarching vision.
Does each country have its own use case and its own AI capabilities?
If you look at eCare, we currently don’t have the exact same use cases for Austria and Germany. But that’s only the launch phase. Long-term, I expect that we need similar use cases for eCare. That doesn’t mean you have the same vendor for each capability. The differences you see are also due to local requirements or preferences.
So it’s not about everybody doing their own thing and in the process re-inventing the wheel...
We need AI for various applications, not just a monolithic digital assistant but for many other products. It’s not necessary that everyone be intimately familiar with these complex topics requiring tons of expert know-how. You don’t have to deal with different vendors to negotiate a deal and go through the same steps that somebody else has already taken with an adjacent product.
Why should somebody join Deutsche Telekom if he or she is interested in AI?
We’re a big company with fascinating topics, and the combination of running a business, state of the art technology and innovative ideas is pretty unique. It’s an amazing task to implement ideas in this domain, there honestly aren’t many other places where you can do that. It’s one thing to try an AI algorithm in some university lab, but it’s something else to do it in a big enterprise environment. That’s why we love working here.