Banking has been lying to itself for 25 years

Twenty-five of those years spent in financial services. Long enough to know where the bodies are buried.
I say that as someone who spent the first 15 of those years helping construct the lie.

an article by Bryan Carroll, digital bank builder, top 30 Global Fintech Leader

Bank of Ireland, National Bank of Abu Dhabi, Rabobank, Ulsterbank, in truth all great organisations. I helped drive the build of the data warehouses. I signed off on the segmentation models. I sat in the meetings where we told ourselves that if we just collected enough, connected enough, analysed enough — we would finally know our customers.

We didn’t. We just got more confident about being wrong.

In 2019 I moved to Vietnam with a literally a suitcase and an idea. I co-founded TNEX – Vietnam’s first digital-only bank – serving the unbanked and underbanked consumers and micro-SMEs that every model I’d ever built had consistently underestimated, or never seen at all.

I made the move because I thought the problem was the customers. That serving people with thin credit files or irregular income was the hard part.

It wasn’t. The hard part was exactly the same in Hanoi as it had been in Dublin and Abu Dhabi.

Banks don’t lack data. We lack understanding.

And there’s a specific failure mode that comes from that gap – one we almost never talk about. Not defaults. Not fraud.

Mistiming.

A mortgage offered six months too early. A credit limit increased at exactly the wrong moment. A loan approved for someone who is eligible on paper but quietly, invisibly fragile in reality.

The model said yes. The moment was wrong.

The customer’s life got harder. The bank never knew why.

I want to tell you about a woman I met in Hanoi.

She ran a small food stall. Cash business. Thin margins. Every day dependent on the day before going right.

By every measure we had, she qualified for credit. Income history, repayment behaviour, the works. The model said yes.

But I sat with her for an hour and what she described wasn’t a growth opportunity. It was survival under pressure. School fees coming. Input prices up. She didn’t want to expand. She was managing uncertainty with everything she had.

A loan at that moment wouldn’t have been banking. It would have been weight added to someone already at their limit.

Two months later – same woman, same credit file – her cash flow had stabilised. The buffers were back. The same loan, at that moment, would have genuinely helped her.

Nothing about her identity changed between those two visits.

Her position did.

I’ve thought about that woman probably more than any model I’ve ever built. Because she’s not unusual. Every bank I’ve ever worked in has thousands of her – in Dublin, in Abu Dhabi, Botswana, Russia, in Hanoi, Saudi Arabia, now in Manila. The question is whether the system can see what I saw when I sat across from her.

Can it tell the difference between eligible and ready?

For most of my career, the answer was no. Not because we didn’t care. Because the technology wasn’t there. Data was thin. Signals were slow. Systems were rigid. So we made do with proxies. Static scores. Eligibility criteria that captured who someone was on paper but nothing about where they were in their lives right now.

Here’s what’s changed.

We can now read behaviour in motion. Transaction sequences, spending patterns, cash buffer movements, income timing – fed into a model that reads them the way you’d read a sentence, understanding not just the words but the direction things are moving.

The output isn’t a score. It’s a position.

Four dimensions: Stability, Confidence, Readiness, Stress. Not labels. Not categories. Coordinates that shift as behaviour shifts.

We’ve seen position deteriorate up to 30 days before delinquency appears. The score looks fine. The PD hasn’t moved. But the behaviour is already telling a different story — buffers thinning, variability increasing, small liquidity events starting to stack.

I have shown this to credit officers who’ve been doing this for thirty years. The good ones go quiet for a second. Then they say – I have always known this was happening. I just couldn’t see it.

That’s the point. The signal was always there. We just didn’t have the tools to read it.

Thirty days is enough time to help someone. It’s enough time to restructure, to reach out, to adjust. Thirty days before the missed payment, before the relationship breaks, before the harm is done.

We don’t wait for the score to move anymore. We watch the trajectory.

And on the other side — when readiness builds, even in a customer with a thin or patchy history, we can see it forming. The offer aligned to that moment lands differently. The repayment quality is better. The relationship goes somewhere.

We don’t approve people. We approve positions in time.

I have been in enough boardrooms — in Europe, the Middle East, Southeast Asia — watching AI get presented to leadership to know what happens.

The questions are always the same. Accuracy – Speed – Cost – Compliance.

Nobody ever asks: are we solving the right problem?

Because the answer to that question is uncomfortable. If your underlying model treats customers as static risk containers, AI doesn’t fix that. It scales it. Faster underwriting of the same wrong signals. More confident decisions built on the same incomplete picture.

AI amplifies whatever assumptions you start with. The banks that understand that — really understand it, not just as a line in a strategy deck — will build something genuinely different. The ones that don’t will spend a fortune making their existing mistakes more efficiently.

The technology isn’t the hard part. The honest conversation about what you’ve been missing is.

After 25 years I’ve stopped thinking the gap between what banks know and what they understand is a technical problem.

It’s a question problem.

We have been asking: who is this customer?

We should have been asking: where are they right now?

The first question gives you a profile. The second gives you a person.

The woman in Hanoi had a perfect profile. What she needed was a bank that could see her position.