Orchestrated by the local serial entrepreneur and IT expert Davor Runje and the entrepreneur and marketing expert Hajdi Ćenan, after vHeart startup that profiled itself on the London insurtech and healthtech scene – they have come up with another. This time they’re focused on the financial industry. Their airtAI solution is meant for small and mid-sized banks that are interested in applying new technology, such as artificial intelligence (AI), in order to get better understanding of their clients, and thereby be better in preparing and communicating their products and services to them.
airtAI will enable such insight to banks through data processing. Their solution processes transactions of the existing bank’s clients and, based on that, predicts how those clients will manage their money in the future. For example, what is the likelihood of them topping up their car on a specific gas station, or doing some weekend shopping or going to nature. And then, based on this ground model, models that solve specific business problems are built, such as, for example, what is the likelihood of a client needing a certain loan or investment opportunity.
How does airt recognize users needs?
Techniques of deep learning are basis of all AI successes we’ve witnessed over the past few years, explains Hajdi who took over the CEO role in airt. Those techniques require vast amounts of data and a lot of processing power to crunch that data – however, for many real-world problems, not having enough data is still the biggest problem. As she points out, it is not difficult to find millions of photos of cats and dogs or translations of sentences from English to German, but it is a problem to find good datasets of rare objects or languages.
In order to overcome that problem, the modern AI systems process huge amounts of data that is available but is not directly related to the problem at hand. For example, if you want to build a system that translates from Swahili to Croatian, first you will train the system by using many available untranslated texts in Swahili and Croatian, and only then will you train the translation on pairs of sentences in Swahili and Croatian.
That same methodology we use in banking: we first use huge amounts of data to better understand individual clients and their habits, and only then solve particular business problems, from offering a specific loan to credit card fraud. For example, if your spending in a drugstore and a baby center suddenly goes up, and your utility bills are still relatively low, that is a pretty good signal that you need a mortgage to buy a bigger place due to a new baby in the family.
There are many such examples, adds Hajdi, and the main comparative advantage of their system is that it reveals those patterns and regularities without having to specify them. In practice, that means that the system built for one market can quickly and easily be adapted to any other market, no matter how different they might be.
Processing one’s banking data
On a technical side, Hajdi explains that the system currently supports scaling on clusters up to ca. 1000 CPU cores and 100 GPUs, in other words, they have hardware power of a data center with several dozens of physical processors. Scaling on a cluster has significantly sped up the process of data preparation for training, from the initial 100 hours to just one hour for a billion transactions by using 128 cores.
In that context, I was curious how long on average would take to process data from a single user that has a bank account for 20 years.
An average client has around 200 transactions per annum or 1000 in the past five years. If you have a bank with around 1 million clients, that is circa 1 billion transactions per year. By using a cluster with 128 CPU cores, data preparation takes about an hour, and training and hypervalidation of the model on 8 GPUs around 20 hours. Any data older than 5 years is usually of low quality, so we don’t usually even consider using it.
And in case you are wondering what about personal data protection and GDPR, banks already have approvals from their clients for processing such data; airt has very little to do with it. Even though they are enhancing the communication towards the bank’s end clients, they do not change the bank’s existing communication mode, their solution is not in direct contact with the end clients.
The future of banking is in new technologies
airt’s first model was set up for production in 2018 when they started working with one bank and one credit card processing company, Hajdi explains:
In this year and a half, we’ve continually tested and upgraded the models, so today we are ready to offer a scalable solution for several banking problems, already proven in practice. The banks’ experiences are great because they can measurably cut costs or increase revenues, but their clients’ experiences are also great since now their bank is becoming a true partner that is following their needs and offering them specific solutions when they need them the most. Possibility of personalized communication, enabled primarily by the digital channels combined with such advanced analytical systems, is the future of banking.
In anticipation of that future, this financial solution seems like a good move. We are already six months into PSD2 regulative being in effect, which has opened banking data to third parties. This means that the fight for clients and their better experience is no longer just among the banks, but there are new, more agile competitors out there, such as fintech startups and even bigtechs who are now entering banks’ protected territory, which is putting the banks into an unenviable position. If they want to keep their status and position, they need to immediately start using the data they have in a much more efficient way and start applying new technologies.
Of course, easier said than done – banks, like any other corporations, are limited by slow processes of implementing such solutions while, on the other hand, AI experts are both very scarce and expensive, says Hajdi. And that is where startups like airt or Worig come in handy, as they operate as somewhat of a bridge between the quickly changing tech world and large, slow systems, such as banks.
AI solution within weeks
airtAI has built its solution for banks on the latest deep learning techniques that were primarily built for language processing, but have been adapted to process banking transactions. Most of the solution was built by Davor Runje, airt’s CTO, himself, with the assistance of their senior data scientist Frane Kačić Peko. A particular problem to overcome was the (pre)processing of historical data – one of the major problems in implementing AI solutions in general, Hajdi points out.
airt specifies the exact type and format of data that needs to be uploaded into their solution that can be installed and deployed in the bank’s on-premise data center (no need for connectivity to an external network), in order to start model training and get the desired output. The entire process can be done in weeks, unlike months (or even years) currently taking for implementations of AI solutions.
At the moment, airtAI can do the proof-of-concept implementations in 3 months, and their goal is to automate the processes even further and cut that time down to a single week.
Preparation of human and hardware resources
As we can see, implementing AI solution takes time, but also needs certain prerequisites, most importantly hardware and human resources. Large banks can either afford to hire large numbers of AI experts and develop their own AI solutions in-house or to hire large (consultant) firms to custom-build solutions for them. However, small and mid-size banks don’t have that luxury, says Hajdi, so even though many are trying to do something internally, most are not satisfied with their solutions’ levels of quality and sophistication and, ultimately, results they get from them.
That is exactly the reason why airt is focusing on banks up to 10 million clients, as a solution like theirs would significantly facilitate the implementation of new technologies.
Considering there is a vast number of data centers and using cloud computing is rather standard already, training the models on-premise is quite an unusual approach. Nevertheless, when it comes to banks, it is a rather logical and necessary step due to security reasons. In order to train the models on-premise, banks need to have certain hardware and software configurations available but, in case they don’t already have it, Hajdi says it is not a large investment for the banks, as we’re talking about equipment that costs around 10k EUR.
As for the people in the banks necessary for overlooking and managing the solution, we’re talking about a data scientist that can export and format the data according to our specifications and then upload them to our docker image, which needs to be installed into the bank’s data center by someone from the bank’s IT department.
Raising awareness of AI to get initial opportunities
In only two years of cooperation and development, airt has come to a point where they can offer banks a scalable solution for several types of problems. On the other hand, considering that their biggest investment in solution development was – time, any further upgrade and maintenance of the solution does not require significant capital investments – meaning that they can offer the solution for a (much more) affordable price, concludes Hajdi.
As a matter of fact, for any skeptical clients, we are even ready to split the risk – they can get the solution free of charge, and we will only take a cut from increased revenues.
And, as we well know, banks are very traditional businesses, slow to change, and even slower in applying new technologies. Such a client is always a challenge, regardless of the industry, and even airt didn’t get the first chance easily. However, as Hajdi puts it, even when you do get a chance, it is all up to you and your delivery and results.
Luckily, our results have opened some new door and it is much easier nowadays than it was in 2017, when we first started pilot discussions. Apart from applications in the financial industry, in these past three years in which we’ve been working with AI, building the community and raising public awareness about it, we do see better acceptance and understanding of the need for AI. And it is getting easier to find people within companies that do the same internally but, of course, locating the right person can take time.