How Natwest built an AI team and gets projects into production

How a major bank builds AI capabilities and gets algorithms into production

Share

High street bank Natwest has been steadily building out its artificial intelligence team over the past two years, starting with very specific, individual projects and now growing into a formal AI unit.

Speaking during Finovate Europe last month, Tom Castle, head of AI Practice and development at Natwest, said that this team delivered roughly a dozen proof of concepts in 2018, but in 2019 started to think about how to scale that capability.

iStock
iStock

The result was a systematic approach to driving AI adoption within the bank, where Castle and his team developed "a framework to scale our AI ambition" based on work done by Prediction Machines.

This starts with internal scouting for opportunities across the bank, where Castle and his team sit alongside front line workers to talk about "things they recognise" - that could mean taking an invoice and extracting information from it, or sitting sitting with all agents to determine the next best actions. "Building prototypes and solutions they understand is a better way to talk about AI than broader ideas," he advised.

The second, 'beta' stage involves model development in collaboration with governance and compliance. "That process is based around traditional software development lifecycles," Castle said, with data governance processes built in early in the process.

Lastly, there is going live.

"The roles and responsibilities of having a machine learning model live is different to traditional software, so at the moment it really is more about keeping the lights on, that really starts to change with machine learning algorithms that need monitoring, retraining," he said. "Our organisation doesn't have those roles right now."

An example of this process in action was an email categorisation algorithm Castle and his team developed for Natwest customer service staff.

"One of our big problems in the commercial bank is the vast majority of customer interactions are via phone or email, that takes time," he explained. "So can we use AI to help process those quickly and effectively?"

Castle and his team subsequently developed topic and category models to scan all incoming emails to pick out known topics and automatically process those emails more effectively with smarter routing.

Challenges

Castle also talked about the first and last mile challenges around AI development within a major bank.

"The first mile is to get access to lots of good quality data," he said. "The challenge is that lots of banks have historically been treated data as a liability... At the other end of the spectrum is the last mile, which is the ability once you have built something to get that through regulatory and compliance and into production."

Another challenge was, of course, hiring.

"That talent question is definitely a challenge. A lot of data science doesn't need a PhD, and in some situations actually we have found it easier to retrain business people to understand these technologies than it is to train a data scientists to understand the business," Castle said. The Natwest Data Academy, which was set up this year to train 1,000 staff new data skills, is an example of this in practice.

The resulting team is roughly "a third business, a third data science and a third core technologists," he added.

Earlier in the day, Shezad Khan, global head of advanced analytics at HSBC spoke about similar challenges.

"It has been a nightmare for the past eight months," he said. "I have been trying to recruit for 10 data scientists over the past eight months. HSBC has a list of suppliers we only work with and if you want to apply for a job at HSBC you fill out a form, take a situational judgement test, then a series of interviews. So for data science, where there is high demand for people like that I have been losing out on a lot of candidates because of that process."

Attitudes to AI

Castle also talked about having three broad groups of people within the bank and how their attitudes to AI differ.

First he echoed a common complaint in enterprise AI circles.

"There was an expectation of silver bullets at the executive level," he said, driven by startups and consultancies over-promising in the market. "What we are trying to do is start with a real business problem and then find the right solution or partnerships to solve that."

On the other end of the spectrum, Castle spoke about the 'killer robot syndrome', where "the longer you talk about AI the more likely you will end up talking about killer robots".

He said this thinking is most prevalent amongst junior staff members at the bank who are most worried about their jobs, which stems from a "lack of understanding" about the technology.

Lastly, there is what he calls 'the stuck middle'.

"I think we have been quite lucky that there is a fair bit of understanding at the executive level," but that "there are people that have been there a while who have a vested interest in maintaining the status quo, and they tend to control the budgets, so educating them is a challenge," he said.

"Recommended For You"

Data and AI trends 2018 UiPath launches three immersion labs to show off its AI capabilities