The Most Advanced AI Model Can Also Be the Most Fragile: A Lesson for Investment Stewardship

The most capable large language model (LLM) available anywhere was switched off.¹ Anthropic had released Claude Fable 5 only days earlier, which was the first of a new and more powerful class of AI models. But a U.S. government directive forced its withdrawal.² Since user access could not be cleanly limited to foreign nationals, it was disabled for users and enterprises around the world.

The event will be mostly understood as a story about export controls, geopolitics, and model sovereignty. For those of us responsible for stewardship and investment workflows, there is a more useful perspective, which we’ve recently argued for.

In The Foundation Matters More Than the Model,³ we made the case that as LLMs grow more capable, durable advantage comes not from the AI model on top, but from the foundation beneath it: trusted, investor-grade data, domain expertise, and the fiduciary standards that makes an output reliable enough for an institutional decision. This event demonstrated in real time what the market has already shown; that firms anchored in proprietary data weathered a recent market correction better than those anchored in software.⁴ In this instance, the thing that vanished was the frontier model itself. The most advanced model in the world – the newest, most capable layer, and the one most often recognized as offering a competitive edge – turned out to be the most fragile element in the technology stack.

The Frontier Model as a Source of Risk

This is especially worth considering, because the prevailing instinct runs the other way. We assumed that access to the model at the frontier was the prize. The lesson here is that the frontier model is also the layer most exposed to disruption, whether from regulation, repricing, or a capability simply being pulled. If the value of a stewardship workflow is anchored in which model it runs on, it inherits that model's fragility.

A foundation built on governed data behaves differently. When an output rests on proprietary, investor-grade data that is architected to be traceable to source, consistent across jurisdictions, and validated by experts, the model on top becomes closer to interchangeable infrastructure. It can be upgraded, downgraded, repriced, or withdrawn, while the controls that make the output defensible remain unchanged.

What a Resilient and Robust Foundation Looks Like

I want to be precise about what this promises. When the technology foundation is sound, the substance of an AI-assisted output is reproducible across models. The facts retrieved, their provenance, the consistency rules applied, and the human sign-off do not depend on the AI-layer that sits in the middle, because the model doesn’t produce them. Swap a frontier model for a smaller, open-weight, or sovereign one, and the principle should hold. What a less capable model changes is the effort, not the result. It may take more oversight to reach the same governed output, but that output is reproducible. The path to it may not always be identical. A well-governed workflow degrades gracefully when it has to fall back to a lesser model. A model-dependent one breaks.

For stewardship, a robust foundation has real consequences. Our analysis informs voting, engagement, and client reporting, and those decisions have to be explainable and reproducible, regardless of which tool helped produce them. A workflow that becomes unavailable or unverifiable because a model was pulled is not fit for a fiduciary context. Resilience to that kind of disruption is not a convenience. It is part of what makes AI safe to use in institutional decision-making in the first place. And this foundation acts as a competitive moat in offering model-independent, exclusive, domain-specific value for investors, that AI models alone cannot replicate.

Retaining a Place for Trusted Data and Human Judgement

None of this argues against the frontier. The best models are remarkable, and they will keep getting better. But they will also keep being launched, repriced, restricted, and occasionally switched off, because that is the nature of a fast-moving technology that is only beginning to attract this and other kinds of government intervention.

The lasting advantage in our field was never going to come from being on the best model in any given week. It comes from the trusted data and human judgment that make any model's output worthy of an institutional decision. We made that case in principle. This is what it looks like in practice.

Read the full white paper: The Foundation Matters More Than the Model: Why Trusted Data and Human-Centric AI Will Define the Next Era of Investment Stewardship.‍