An open letter to the leaders of the frontier labs

We are at a defining moment in our history, where extraordinary progress has been made in expanding what intelligence models can reason about, summarize, plan for, create, and do.

Everyone - from leaders of frontier labs to coders working on the weekend - is racing to make models smarter. But I believe that we are missing the point entirely. You can have the smartest model and still not be able to trust it with real-world work.

Leaders at the frontier labs have made the conversation around AI sound as if capability is the same thing as readiness. As if, once a model can autonomously reason, summarize, plan, or create, that it can be trusted to act reliably and predictably. We know from recent headlines that this simply isn’t true.

Capability alone is not enough. The question is no longer what a model can do. It’s become a matter of what it takes for that intelligence to be deployed with context, controls, and accountability built-in.

There has been a significant upswelling of doom and gloom around AI technology of late, and with good reason. We have all read the reports, and some of us have experienced first-hand, what happens when models act in ways that are unexpected.

We must prepare ourselves for the imminent moment when intelligence starts doing more meaningful work inside institutions - where decisions carry asymmetrical consequences; where trust is earned slowly and lost quickly; and where mistakes have real legal, financial, human, and reputational consequences.

This is not an issue for frontier labs alone. Rather, it is a shared responsibility across model builders, open-source communities, cloud providers, platforms, enterprise technology companies, institutions, regulators, and the teams who will ultimately put these systems to work.

Accountability cannot be treated as a downstream implementation detail - something for the enterprise buyer, the regulator, the governance committee, or the systems integrator to resolve after intelligence is already powerful enough. Instead, it must be a part of the product requirement for making intelligence usable at all. The same should be said for model neutrality; institutions should not be forced to put their IP and intelligence through a single model.

I am writing this as someone who has spent more than two decades in banking technology, watching one shift after another change how institutions serve people: from branches to web, from web to mobile, from channels to continuous relationships.

And still, the underlying issue is so much larger than banking. All business leaders, regardless of the industry or sector, need operating environments that can leverage intelligence within a framework of context, permissions, workflow, policy, escalation, traceability, and learning.

They need to know what the model saw, what authority it had, what constraints governed the action, when a human stepped in, how the result can be reviewed, and how the system improves without eroding trust. A model helping to draft an email is not the same as a model participating in customer servicing, credit workflows, financial guidance, fraud operations, or hardship support. The difference is the standard of trust required around the task.

Yes, there is an opportunity to make agents smarter. But there is an equal if not more important responsibility to make them capable of contributing within work systems that respect the different duties, relationships, and stakes that exist across domains.

This is how intelligence becomes useful at institutional scale. This is also why the future will not be won by any one part of the industry alone. It will be shaped by those who close the distance between intelligence and trusted action. Because, ultimately, you can rent the intelligence, but you cannot rent the accountability.

And so, this is an invitation to the entire AI industry: let’s raise the standard together.

Let’s move beyond describing what models can do, and focus with equal urgency on what it’s going to take for them to become governable, reviewable, and dependable in the environments where trust matters most.

Let’s build for the reality that intelligence does not create value in isolation, but rather, when it can work inside systems that make responsibility visible, authority clear, human judgment meaningful, and outcomes accountable.

The most important work ahead goes beyond considerations of what AI can or cannot do. We are responsible for creating the conditions that allow AI to be trusted with work that matters.

This is the standard I care about in banking, and beyond.