Banks Struggle to Manage Technical Debt When Dealing with AI, Data Science

Many technologists tend to think that technical debt accrues when a shortcut was taken and a cheaper but inferior platform was chosen or built to solve a problem. It’s where robotic process automation (RPA) has received a bum rap, as bots are created to tackle repetitive tasks, but there’s limited thought given to redundancies and to the coding. As a result, a constant string of fixes and recoding is necessary, and after a while, the cost of upkeep is higher than the cost of a better, longer-term solution.

But technical debt can also be a necessary by-product of innovation, as sometimes things need to get broken before the final fix comes to the forefront. This is where machine learning development has become challenging for the TD-minded technologist.

David Hardoon, senior advisor for data and artificial intelligence at UnionBank of the Philippines, says that from an IT perspective, technical debt in the world of machine learning and data science has been referred to in terms of code refactoring, but the industry is only now coming to realize that there are no guarantees when exploring the world of AI-driven models.

“At the end of the day, the nature of machine learning and data science-based solutions is that you have no guaranteed certainty with the outcome,” Hardoon said, while speaking on a panel at the WatersTechnology Innovation Exchange.

When implementing an extract, transform, and load (ETL) workflow, it could be simple for the IT team to get up to 99.98% accuracy, but expecting that same accuracy over and over is “impossible” as machine-learning models continually learn.

This, according to Hardoon, is a new dimension of technical debt—not everything invested will immediately translate to value, or in a way that IT teams are accustomed to. One of the main challenges is getting the organization to focus on an objective, and the operationalization of data science.

“Machine learning, data science, AI is this weird beast, and I used this analogy the other day: I call it the jam between the two slices of bread that makes the sandwich delicious,” he said. Essentially, the way Hardoon sees it, when you combine these disciplines, they’re not just about technology, and they’re not just about the business either—they’re interwoven and “it is something in between.”

So the key then becomes understanding the technologies that are available and understanding how they will sit in the organization and what strategic areas they will address.

“I’ve seen amazing solutions that are really, truly technological marvels using the most advanced technological data science, machine learning, [and] deep learning techniques, but when push came to shove, it completely missed the point,” he said. 

The Data Science Equation

Technical debt from a data scientist perspective is more like a knowledge gap, according to Peili Chang, chief data officer at AXA Philippines. It can become challenging to manage technical debt if the talent pool is shallow. 

“Unfortunately, because this market is booming so fast, there is clearly a lack of talent in the whole industry. So, you may even have trouble hiring the right data scientists,” he said. “Maybe the manager, or the hiring person, may not be able to assess properly the person who is being hired. So hence, I end up with people that know a little bit of data science,” Chang said.

Then this person jumps into doing data science, using whatever is available and may even get by with developing data science models well. But still, the knowledge gap for how those models are being applied to the business isn’t there. 

“I would say, as well, the open-source programs and modules that they offer to us—we have a lack of understanding [of] how it works exactly. That for me is also one of the troubles in terms of technical [debt], but more to the data science itself,” he said.

This could be where the role of the data translator fits in. Chang said without a data translator, it will be harder to avoid failures. This person would have sufficient knowledge from the business and bridge the gap between IT and the data science team. 

“If you don’t have a data translator, how do you avoid those failures? Well, that’s quite complicated because what I see, personally, is we have a lot of trials, a lot of MVPs [minimum viable products], or proof-of-concepts being delivered here and there, [but they] never went live because simply we could not make it happen. So I think it’s really important to have it,” he said. 

Philip Yu, an honorary professor at the computer science department at The University of Hong Kong, agreed, and added that there needs to be more collaboration between teams to make sure that a project is staying on track to address the required business need. You need to include “people who are the users, who are going to use your AI model so that at the end of the day, we can identify the problems and formulate the questions to address,” he said. “I think it’ll be best if we have all these people, including data scientists, end-users, and also the data providers [in one group] so that they can work together and identify what they need and what they want to have.”