Goldman Sachs CEO’s view on AI applications to banking risk management

Pancreatic cancer is one of the most lethal adversaries in medicine — rarely detected until it is too late. But in April of this year, the Mayo Clinic showed how AI can change that prognosis. Its AI model identified 73 per cent of pre-diagnostic pancreatic cancers up to three years before diagnosis — turning a near-certain terminal sentence into a window for treatable intervention.

I’ve made no secret of my optimism about what AI can do for society. Disease is one frontier. There will be thousands more — some incremental, some transformative — over the coming decades.

The question is not whether AI will advance. It will. The question is whether we advance it well. The potential dividends are too vast to ignore, from drug discovery and food production to supply chains, transportation and creative industries. We also need the very best frontier models to protect our national security interests and defend against nefarious actors of every kind.

We should not throttle this progress. We should focus on getting it right.

Unfortunately, public debate on AI seems trapped in a false binary: race forward without guardrails or freeze innovation through heavy-handed restrictions. But there are three critical layers of responsibility to consider: governments setting the rules of the road, developers building safe models, and — the layer that gets the least attention — companies deploying the AI models. They are the ones who decide where an AI model sits in a workflow, what authority it is given, and who is accountable for failure.

I’ve seen suggestions for grafting bank regulatory structure on to AI oversight. This is the wrong approach. AI presents unique technical challenges, distinct patterns of deployment and a much faster pace of development. I’m not arguing that AI should face less scrutiny than banks do. I’m arguing that scrutiny should be built for this technology, not replicated from another one.

I talk to hundreds of CEOs every year, and it’s clear to me that every one of them is now running an AI company. Like Goldman Sachs, they are deploying the technology across their operations and creating agents at scale. The benefits will be enormous. But there is corresponding operational risk — a risk that many of them have not had to manage before, but one that financial firms have managed for decades.

At Goldman Sachs, we manage risk based on principles that could be helpful not only for AI developers but for AI deployers — that is, companies large and small that are integrating AI into their operations.

First, people must be held accountable. Financial companies have been running automated trading platforms using machine learning for decades. The concept that we cannot blame the technology when a trade goes wrong is well established. Individual decision makers must be responsible for the risks their systems accept and the outcomes they produce.

Second, counterparty risk must be understood. Companies must evaluate and understand the risks inherited from the model developers and avoid becoming dangerously dependent on a single provider. AI systems must be independently reviewed, well documented, continuously monitored and regularly reassessed as their capabilities evolve.

Third, stress testing. AI deployment decisions should incorporate comprehensive risk reviews that pressure-test models against realistic, severe stress scenarios. This includes deliberately probing for failures, misalignment and unintended consequences in controlled environments. It’s well known that banks perform stress tests for regulators, but less well known is that they routinely stress test their own processes and models. The best companies will adopt routine stress testing for AI.

Finally, risk escalation should be rewarded. Vulnerabilities can appear anywhere in an organisation. Raising concerns should be celebrated with clear escalation channels available to employees of all levels. This must become a core component of the culture of any company deploying AI.

The companies that win the AI era will not be the ones that moved fastest or waited longest. They will be the ones that built the discipline to deploy at scale with trust and confidence. The reward for getting this right could be measured in diseases caught early, discoveries that occur years ahead of expectations and compounded productivity gains across every sector of the economy.

The choice is not between innovation and safety. It’s between poorly governed experimentation and progress built on accountability, trust and confidence. We can be impatient for that future and still be rigorous about how we reach it.