The challenge of scaling AI at NatWest bank

'AI is a potential solution to a problem. Is there an easier way to do it? If so that's the right way' says head of AI Tom Castle

AI is already making its presence felt in the financial sector. Fraud detection systems have been around for some time for example, but there are many other ways in which automated predictive technologies could be put to good use too. To this end, NatWest bank put together a team two years ago with the goal of understanding the possibilities of automation in the context of the business. As ever the challenge is moving beyond proofs of concept to something that's genuinely useful to a highly regulated, customer-facing business.

"Customers don't come to NatWest because we have AI," said Tom Castle, head of artificial intelligence practice and development at NatWest.

"AI is a potential solution to a problem. Is there an easier way to do it? If so that's the right way given the current maturity of the technology."

Castle's team works closely with the business owners of the bank's financial products in order to align their efforts. In a keynote delivered at the Computing IT Leaders Summit last week, he went over some of the methodologies pursued and lessons learned from looking at how to scale up AI at the bank.

The company uses an agile methodology developed by UK Government Digital Services (GDS), which focuses on four key areas: strategy and vision; ways of working; education and awareness; and attracting and retaining talent.

In most respects developing machine learning algorithms and applications follows the standard lifecycle of PoC-alpha-beta-release, so it's not reinventing the wheel, but there are some important differences, Castle explained. The first is that models must be constantly monitored, improved and trained. Another difference is that governance has to be built into the development process at a much earlier stage.

"You still need to do classic software production lifecycle, but some things need to come in earlier. We've had to look at governance to ensure we have right controls in place soon enough in the process," he said.

A third difference is that machine learning algorithms need to be constantly iterated and retrained. "Once it's live, you have to keep monitoring it and improving it, so that's quite different from the standard software delivery model."

Frontline staff think 'it's going to take my job away', and for the execs it's a silver bullet - Tom Castle

Most of the challenges to scaling AI are non-technical. Under the ‘ways of working' header, there are a number of ethical considerations to be made around bias, and there's a huge role for education in getting people to understand what AI is - and what it is not.

"Frontline staff think 'it's going to take my job away', and for the execs it's this silver bullet that's been sold by whatever startup or consultancy they've been talking to," Castle said.

"Our job is to demystify and dehype it and explain what it can do. it can extract text from messages, it can analyse sentiment, make a prediction on loan defaults - it's about explaining it in a language they understand."

Where AI sits in the decision-making process

Castle's team uses a model set out in the book Prediction Machines, in which a task is typically split into five stages: input; prediction; judgement; action; and outcome. Currently, machine learning is only really beneficial at the ‘prediction' stage (although RPA is increasingly being used to take action on those predictions).

What's important, though, is to be able to gauge the accuracy of the algorithmic predictions. For example, a text-analysis algorithm might judge that a customer email is about stopping a cheque with 95 per cent confidence - but as a business is this degree of certainty sufficient to act upon that conclusion? What would be the consequences if a customer asks to stop a cheque and the system misses it? Would handling such requests be more efficiently carried out in the traditional way with a human assistant? And what if there's a false positive, with a message being improperly categorised as a request to stop a cheque when in fact it's about something else?

Castle's team uses a risk matrix to assign probabilities to each scenario. If the AI is mature enough to scale, the positive consequences should outweigh the risk of the algorithm miscategorising the input. If the probability of a false outcome is too high and the consequences of failure too great, then it's back to the drawing board.

Other barriers to scaling AI

Technical and business risk are not the only barriers to scaling AI. Even banks, which are able to pay more than most to attract talent, find it hard to compete with the likes of Google and Facebook in a city like London.

Then there's access to data. Machine learning algorithms are data-hungry beasts, but in a highly regulated industry much of a bank's data may be out of bounds or at least require some heavy-duty obfuscation and anonymisation before it can be used. Another issue arises when different departments are using different clouds, which without good governance can result in core data that is incomplete or inconsistent.

"We've seen a cloud adoption approach with different teams using different clouds, so there's stuff all over the place and getting it all together in one place is difficult, particularly if you've got your data stored in one cloud and processing in the other," Castle said.

Ultimately, AI is about taking the grind out of repetitive tasks, aiding human decision making, creating efficiencies, and reducing the cost of error.

"There are a lot of places where the business does not have an appetite for risk, so we're thinking 'how can we use predictions to help people make better, faster, more consistent decisions," said Castle, giving the example of a call centre where voice biometrics are used to authenticate the caller before an operative handles the call.

"We use the best of machine learning to identify the customer - or fraudster - but keep the people there for the more value-added part of the conversation."