How an AI from Vienna became a global top 5 product

“Deepassist” is the name of an artificial intelligence that answers both written and telephone customer enquiries

Being among the top five in natural language technologies worldwide is a great success for Deepassist. Founded by Roland Fleischhacker, it is a major player when it comes to AI in text and speech recognition.
The AI is pre-trained and immediately usable for every channel a customer may have. A real-world example shows how Deepassist radically accelerates customer service: The City of Vienna’s “Stadt Wien - Wiener Wohnen” is one of the largest property management companies in Europe, where 550,000 customers are serviced around the clock, dealing with about 1,000 different issues. Deepassist supports the agents in real time – which greatly reduces the handling time and significantly increases the resolution rate after the first call or email. In addition, training new employees now only takes a few days instead of weeks
CEO Roland Fleischhacker explains what is behind this groundbreaking technology and where it is already successfully in use.

What does Deepsearch do in conjunction with Deepassist?
Roland Fleischhacker: We have a product called Deepassist, and the name says it all. The term ‘deep’ has stood for artificial intelligence applications for some time. When we invented the name Deepsearch in 2010, this metaphor did not yet exist. Our objective is for a machine to be able to interpret texts on a human level. Language in the broadest sense, whether written or spoken, is an omnipresent thing in both private and business environments. If language can be interpreted in an automated way, a lot of applications arise. Documents can be better classified and retrieved. This applies to email, but also to telephone conversations. It supports communication and, if necessary, also automates it. Here, ‘deep’ stands for AI and ‘assist’ for the fact that we want to help companies support processes with it.

How can you help companies?
Meat chopper: In many different ways. For example, with emails that are read and understood by the AI and forwarded to the right person for processing. That is our simplest use case.

How does deepassist detect this?
Meat chopper: It reads the text and transforms the most important elements – what is it about and the most important information – into a defined format. This makes it readable and understandable for other systems, which then process it. For example, if someone wants to cancel their mobile phone contract, there is a customer number, a contract number and the date of the desired contract cancellation. Then the information is transferred to a mail or ticketing system. This system takes this information in a structured form and starts a certain workflow. A further step is if the AI has already identified exactly what is involved and all the required information is already available. This information is then not forwarded to a person but processed automatically.

So incoming texts are answered automatically?
Meat chopper: Yes. We often even have to include a time delay, because the customer would not understand why their request is dealt with and answered immediately, within a few seconds

Does that work with all languages?
Meat chopper: In theory, yes. We have developed our own technology that works completely differently from 99 percent of other systems that are based on neural networks. Our AI is semantic and pragmatic, and does not react to static patterns but actually understands what it is about. The system can also read between the lines. We are currently patenting this technology for the American market.

Where can this type of AI be used?
Meat chopper: Mainly in customer service, but also in employee service. Especially in customer service, customers are often not able or willing enough to communicate exactly what they want. They often can’t put their request into words, but only explain symptoms or beat around the bush. Here, our system hypothesises what the customer actually wants to communicate. This works not only with emails, but also with calls in real time.

What an empathically talented person can normally do, the machine can also do?
Meat chopper: Exactly. It prepares the information, and an employee can then proactively approach the customer. This significantly improves the customer experience if the system has already told you what the customer wants without him saying it.

Can Deepassist also have conversations, or does it only provide answers?
Meat chopper: We don’t do conversations, but we can connect our system to a chatbot that is capable of having conversations

How high is the error rate with Deepassist?
Meat chopper: There is a kind of glass ceiling in recognition, which is about 95 to 98 percent.

Are there sometimes wrong answers?
Meat chopper: No, it’s more likely that the answer is ambiguous. If a telephone call is not clear, it often requires a query. But we have included many expressions from the Viennese dialect in the system, so that we also understand dialect-specific words. We also have to recognise when someone means something different from what they say. For example, many people still talk about a standing order when making payments, although they mean a SEPA mandate.

Can Deepassist be the solution to the shortage of employees at call centres?
Meat chopper: Yes, because at the moment you can hardly get staff. The customers’ expectations are growing, and the tasks are becoming more and more complex. Our system brings great advantages, especially for high-frequency and highly repetitive enquiries, particularly via email. These are, for example, complaints, changes of account numbers and the like. Manual processing of such emails takes between seven and nine minutes, in our case two CPU seconds.

How long does Deepsearch need to train a Deepassist application to make it usable for a customer? How long does the programming process take?
Meat chopper: Customers don’t want long, expensive or risky products. That’s why one of the basic premises in the development of Deepassist from the very beginning was that it should be very simple, very quick to implement and transparent. The customer service of a bank is completely different from that of an investment bank, an energy supplier or an online shop. We call this domain language. So, I have to recognise what the customer is saying, i.e. understand the domain language, and I have to know the solution processes that trigger certain requirements. We deliver this in the form of industry solutions.
industry solutions. We have pre-programmed systems that not only understand what the customer is saying, but also a solution catalogue of process templates that varies from customer to customer. At Stadtwerke Hamm, a customer in Germany, the implementation of our industry solution for energy suppliers took only five days. Other providers can only do that in months. So, there are ready-to-use industry solutions, such as for facility services of large property management companies with 15,000 or more residential units and public transport, such as railways or airlines. We are currently working working on industry solutions for insurance companies and retail banking.

Can individual industry solutions be transferred from one country to another?
Meat chopper: Yes. There are certain topics that are cross-sectoral, such as address changes or changes in bank accounts. As far as languages are concerned, it’s more a matter of culture. Italians speak much more when calling a call centre, make more small talk than people from other countries. This makes it more difficult because the longer a conversation goes on and it’s not about the actual subject of the call, the more hypotheses the system makes about what it might actually be about. Then the system needs more context to be able to identify what it is really about. We have developed a system based on semantic building blocks – similar to Lego. So, there is a box full of building blocks and we deliver a model. Individual parts of it can be rebuilt individually for the customer with the same building blocks.

Which markets is Deepsearch focusing on?
Meat chopper: Today it is the German-speaking world. 2023 is the preparatory year for us to go beyond this region.