Could AI be responsible for the debanking rise?

Recently, a client of mine had a payment to a legitimate, FCA-registered crypto platform blocked. It happened again shortly after and their account was then put under review.  

When they queried this, the bank wasn’t able to tell them why. Was that decision made by a person? Or did an algorithm flag a pattern that nobody ever stopped to question? Honestly, I don’t know and neither does my client. Right now, there’s no actual way for either of us to find out. 

That, to me, is the bit that should worry all of us. 

At the end of September of this year, futurist Brett King told The National that “99% of banking can be automated.” He even went further, suggesting that in an agentic bank, humans just slow things down. 

Now, I don’t think he’s entirely wrong. There are plenty of examples I can point to in banking where repetitive tasks have slowed things down to the point where, speeding them up with machines could help make the entire process more efficient.  

However, and I cannot stress this enough, when it comes to deciding whether or not to close a bank account, slowing things down and doing the due diligence is the entire point.  

And that, it seems, is exactly the kind of job banks are now lining up to automate. 

Earlier this year, in May, Standard Chartered said that it plans to cut more than 15% of roles  by 2030. To put that into perspective, that’s around 7800 jobs gone. Those functions, interestingly enough, include risk management and regulatory compliance. Similarly, HSBC, according to Bloomberg, is weighing cuts that could affect around 10% of its workforce. 

Now, at the same time we’re seeing a rise in debanking. In fact, a freedom of Information request found that UK banks shut down more than 400,000 accounts last year with banks pointing to financial crime as the reason for the majority of those cases.  

When you put these two trends together and the question pretty much asks itself.  What happens when the humans who check who has a right to a bank account are gone? 

The machines are already doing the flagging. The latest Bank of England and FCA survey found 75% of UK financial firms are using AI, with AML and fraud among the areas where they see the biggest benefit. That, on its own, isn’t a problem as, it’s fair to say that monitoring millions of transactions a day simply can’t be done by hand. 

The problem is how often those systems are making mistakes.  

Industry research suggests that around 9 out of 10 AML alerts turn out to be false alarms. In other words, most of what gets flagged just happened to just look a bit odd even though it was perfectly fine.  

As it currently stands, a lot of those alerts are still landing on an analyst’s desk, which is great because someone still has to look at the customer’s history and decide that, actually, this is fine. That person is the only thing standing between a false positive and a closed account. 

Take them away and a 90-95% false alarm rate could very quickly, and very easily result in many many more businesses being debanked.  

I’m already seeing an early version of this play out in fintech, an industry I’ve worked in for over 10 years now. The UK Cryptoasset Business Council found that roughly 40% of payments to crypto exchanges are now being blocked or delayed.  

Then there’s the law, which has a flaw. Since April, banks have had to give customers at least 90 days’ notice and a specific reason before they close an account. I’ve written about this in a piece for American Banker before where I outlined why I think that’s a good step, and I still do today.  

There is a caveat, however, that I’d be remiss not to call out. If a bank has reasonable grounds to suspect the account is being used in connection with serious crime, it can skip the notice and the explanation altogether. 

What is an AML algorithm designed to do, if not raise suspicion? 

So let’s say a system flags an account, marks it as high-risk and starts an exit. The 90 days, the explanation, the time to go and find another bank, all of that could be void. 

‘Reasonable grounds’ was written with a person in mind. Someone who’s looked at the facts and formed a view. I’m really not sure a risk score can do that on its own. 

I want to be clear here. I’m not saying banks should stop using AI in compliance. They shouldn’t, and realistically, they won’t. A good model will get through a ton of transactions far quicker than any team of analysts ever could and we all love to cut costs. 

My issue is with who gets the final say on closing an account. I, personally, don’t think it should be AI. 

What I’d like, instead, is to see is a human-in-the-loop approach whereby AI does the sorting, putting the cases that actually look concerning at the top of the pile. Then, a real person, someone human, accountable for the decision, looks at it and decides whether that account closes. They need the authority and final say to say no. 

That person should also be the one who writes the customer’s explanation, in plain English. What was flagged, and why it mattered. 

The good news is we’re not at full automation yet. The same Bank of England and FCA survey found only 2% of AI use cases in UK financial services run fully autonomously.  

What’s less reassuring is that only about a third of firms said they completely understood the AI they’re using.  

If banks are determined to shrink their compliance teams, this is the one role I’d protect first. The solution here would be for regulators to make it official. If a bank wants to rely on the serious crime exemption to close an account, a named person should legally be required to have reviewed the zed file and signed it off. 

That really isn’t a big ask, in my opinion. Without it, I truly don’t believe that reasonable grounds means very much at all. 

Brett King may well be right in his belief that humans slow banking down. For a lot of what banks do, that’s probably a cost worth cutting. But when the decision is whether a legitimate business gets to keep its bank account, a bit of slowing down and taking a holistic approach is exactly what we should want.