Too Tech to Fail (Part Three)

Over the past two days, I have looked separately at two uncomfortable developments in banking. The first is that banks are moving more and more of their critical infrastructure onto a handful of global cloud platforms, while the second is that artificial intelligence is moving from helping humans make decisions towards making those decisions itself.

Individually, both transformations make enormous sense because cloud gives banks computing power, scalability, resilience and access to technology they could never economically reproduce themselves, while AI gives them intelligence, automation, personalisation and the ability to detect, decide and respond at extraordinary speed.

Put them together, however, and something much bigger happens because we create the intelligent bank, and the intelligent bank is not simply a digital bank with some artificial intelligence sprinkled on top. It is an institution in which the infrastructure is increasingly external; the intelligence is increasingly autonomous; the processes are increasingly automated; and the decisions are increasingly made at machine speed.

That creates an extraordinary opportunity to reinvent banking, but it also creates an extraordinary systemic risk because AI and cloud are not two separate technology trends anymore. They are becoming one technological architecture.

AI needs enormous computing power, enormous quantities of data and enormous connectivity, and cloud provides all three. As banks become more dependent upon AI, they therefore become more dependent upon the infrastructure underneath AI and, increasingly, that infrastructure sits inside a small number of global technology companies.

Banks are moving computing into the cloud, putting data into the cloud, running AI in the cloud, buying AI services from companies connected to those clouds and using cybersecurity, software and customer platforms that increasingly depend upon exactly the same technological ecosystem.

This is where the argument becomes much more interesting because what looks like hundreds of independent technological decisions inside hundreds of independent financial institutions increasingly converges onto the same underlying foundations.

Historically, banks diversified financial risk by holding different loans, trading different assets, operating in different countries and dealing with different customers, while regulators worried about the connections between those institutions because financial contagion allowed problems to travel from one bank to another.

The intelligent bank introduces something different because the commonality increasingly sits underneath the institutions themselves. The banks look different, but the technology underneath them increasingly looks the same.

That means a financial system containing thousands of supposedly independent institutions increasingly depends upon common cloud providers, common AI infrastructure, common software libraries, common datasets, common cybersecurity tools and increasingly common models. Diversification above the surface therefore disguises concentration underneath it, and that is the first great systemic problem created by the intelligent bank.

The second is that the infrastructure does not simply host the bank anymore because, increasingly, it thinks for the bank.

AI is moving into fraud detection, cybersecurity, credit underwriting, liquidity management, investments, trading, compliance, customer service and payments, while the next generation of AI agents will negotiate, buy, sell, borrow, invest and transfer money without waiting for a human to approve every individual decision. This creates an institution that is not merely automated but increasingly autonomous, and that distinction matters enormously because automation follows instructions whereas autonomy interprets objectives and decides how to achieve them.

Tell a traditional system to reject a transaction when a predefined rule is triggered and you know what it will do, but tell an intelligent system to minimise fraud losses, maximise returns, optimise liquidity or reduce credit risk and the system increasingly decides how to achieve that objective itself.

Now connect thousands of those systems to the same markets, the same information and increasingly the same technological infrastructure and intelligence starts creating correlation.

A geopolitical shock occurs, markets move and AI systems across thousands of financial institutions detect increased risk, causing credit models to tighten lending, trading models to reduce positions, treasury systems to increase liquidity, fraud systems to raise their thresholds and autonomous agents to move money towards safer assets.

Every machine is behaving rationally, but collectively they are creating instability.

That is the critical difference between the intelligent financial system and the one we have known before because the global financial crisis demonstrated what happens when thousands of humans and institutions believe the same thing at the same time. The intelligent financial system industrialises that behaviour and executes it at machine speed. Humans panic together while machines optimise together, and the outcome looks remarkably similar except that the machines will have acted before the humans have finished discussing what is happening.

This brings us to speed, which is perhaps the most important change of all because banking governance was built for human time.

Committees meet, executives discuss, risk officers review, regulators consult and boards approve, while decisions move through organisations according to carefully constructed authorities and controls. AI does not operate like that because it analyses in milliseconds, cloud deploys globally in minutes, algorithms trade continuously, cyberattacks propagate automatically, customers communicate instantly and autonomous agents increasingly transact continuously. Technology therefore compresses the time between event, decision and consequence, which means that the central risk is no longer simply whether the bank makes the right decision but whether humans remain capable of intervening before thousands of automated decisions create an irreversible outcome.

Imagine that a bank discovers an AI model is behaving incorrectly.

In the traditional world, management investigates the problem, convenes the appropriate people, decides what to do and intervenes. In the intelligent bank, the model has already made thousands of additional decisions while those conversations are beginning, those decisions have triggered responses from other machines and those machines have generated further responses elsewhere in the financial system. What started as a model error therefore turns into an automated chain reaction, not because the technology stopped working but because it continued working perfectly.

That introduces something I think we have discussed far too little in banking: decision resilience.

Banks have spent years building operational resilience through backup data centres, disaster recovery systems, redundant networks and business continuity plans, but intelligent banking creates another requirement because we now have to ask what happens when the computers are working perfectly and making the wrong decisions.

That is considerably harder than recovering from a failed server because the danger does not announce itself as a failure. The cloud is running, the network is available, the APIs are responding and the AI model is operating exactly as designed, but the collective outcome is destabilising the institution.

In other words, artificial intelligence introduces a new category of financial failure in which everything works but the outcome is wrong.

This is why the conversation around AI and cloud needs to move beyond digital transformation because, for the past decade, bank executives have been encouraged to move faster, migrate to the cloud, open the APIs, automate the processes, connect the data, deploy the AI and remove the friction.

All of that remains correct, but it is only half of the strategy.

The other half is understanding what happens when the connections fail, knowing who stops the automation when it behaves unexpectedly, recognising where common infrastructure creates concentration and ensuring that intelligence does not turn into correlated behaviour across the entire financial system.

I call this intelligent resilience, and intelligent resilience starts from the assumption that cloud providers will fail, AI models will fail, networks will fail, software will fail, data will become corrupted, cyber defences will be breached and perfectly rational algorithms will occasionally produce completely irrational systemic outcomes.

The objective is not to pretend that every failure is preventable because it isn't. The objective is to make sure that inevitable failures remain failures rather than becoming catastrophes.

That requires banks to understand technological concentration with the same seriousness that they understand credit concentration, common AI models with the same seriousness that they understand common market exposures and autonomous decision-making with the same seriousness that they stress-test capital and liquidity.

Most importantly, it requires banks to understand where control actually sits because an institution might own the customer relationship while renting its computing infrastructure, buying its software, consuming external data, connecting through external networks and using artificial intelligence developed by another company.

At some point the board has to ask a very uncomfortable question: how much of the bank can we outsource before we no longer control the bank?

That is not an argument for rebuilding the past by bringing everything back inside the institution.

Reconstructing giant proprietary data centres, writing every piece of software internally and developing every AI model yourself would simply recreate the expensive technological fortresses banks have spent decades trying to escape. Outsourcing makes enormous economic and technological sense, but the strategic question is no longer whether banks should outsource. The question is what they must never outsource to the point where they lose control of their own survival.

That is the dividing line emerging in intelligent banking because the losers will treat AI and cloud as another technology transformation and congratulate themselves on how many applications they have moved, how many AI use cases they have deployed, how many processes they have automated and how much cost they have removed. The winners will ask much harder questions about where they are concentrated, where they are correlated, where they are dependent, where machines are making decisions faster than humans can intervene and what happens when something everyone believes is resilient suddenly isn't.

Those questions matter because banking is moving into a fundamentally different world.

The digital bank connected people to money, while the intelligent bank connects machines to money, and once machines start making financial decisions, executing those decisions through global cloud infrastructure and interacting autonomously with other machines, the financial system starts operating at a speed and scale that its existing governance structures were never designed to manage.

And there is still one critical component missing from this picture because, so far, cloud has globalised the infrastructure and AI is globalising the intelligence, but now the money itself is becoming global, digital and programmable. Crypto, stablecoins, tokenised deposits and CBDCs therefore change the argument again because technology is no longer simply changing how banks operate. It is beginning to challenge the national borders around which banking, regulation and central banking were built.

So far, cloud has globalised the infrastructure and AI is globalising the intelligence, but there is still one piece missing because the money itself is now becoming global, digital and programmable. That changes the argument again, because technology is no longer simply changing how banks operate. It is beginning to challenge the national borders around which banking itself was built.