One Model Is Not a Strategy

If a client walked into your office and told you they’d put their entire portfolio into a single stock, you’d know exactly what to say. You’d talk about concentration risk. You’d talk about what happens when that one position has a bad quarter, a bad CEO, a bad year. You’d talk about diversification — not as a hedge against opportunity, but as the foundation that lets you pursue opportunity without betting the whole relationship on one outcome.

Now look at how most advisory firms are building their AI workflows.

Over the last two years, firms have quietly wired their operations around a single foundation model. One vendor, one API, one set of prompts, one model’s behavior baked into everything from meeting notes to investment commentary. It happened the same way most tech stacks happen not through a deliberate decision, but through a hundred small ones, each reasonable on its own.

And just like that single stock position, it works fine — until it doesn’t.

Consider what happened on the evening of June 12th. The US government issued an export control directive to one of the leading AI providers, ordering the suspension of access to its two most advanced models — effective immediately, for every customer worldwide. Not because the technology failed. Not because of a data breach. A regulatory action, issued at 5:21 PM on a Friday, disabled production workflows overnight across every firm that had built on those models. Three days after the models had publicly launched.

This wasn’t a hypothetical. Firms that had integrated those models into client-facing workflows — drafting communications, synthesizing research, generating commentary — had that infrastructure pulled with no warning and no recourse. The models were simply gone.

Share

And this kind of disruption doesn’t require a government directive to happen. Providers update model versions, sometimes substantially, with little notice. A prompt that worked reliably last month produces different output this month; same words, different model underneath, different result. Pricing structures shift. Usage policies shift. OpenAI deprecated several of its models earlier this year, shutting down access to tools that firms had built workflows around. A capability your practice depends on today can be deprecated, rate-limited, repriced, or pulled from a geopolitical directive tomorrow.

If your firm’s client communications, research synthesis, and compliance review all run through one model’s API, you haven’t built an AI strategy. You’ve built a single point of failure and given it a seat at every client meeting.

This is the part of the build-versus-buy conversation almost nobody is having. Earlier in this series, we talked about the hybrid model: buy standardized tools, build proprietary intelligence on top. We talked about why the governance layer for that intelligence can’t realistically be built in-house. But there’s a third dimension underneath both of those conversations — and it’s one most firms haven’t identified as a risk yet.

Model diversification isn’t about hedging your bets on which AI company “wins.” It’s about architecture. It’s about whether your firm’s intelligence infrastructure is portable and can draw on multiple models, route tasks to whichever model performs best for that job and absorb a provider’s changes without your client-facing workflows breaking overnight.

Think about how this maps to what you already know. You wouldn’t build a financial plan around a single fund manager’s continued outperformance. You build in flexibility — the ability to reallocate, rebalance, and adapt as conditions change, without unwinding the entire plan. The plan is durable because it isn’t dependent on any single decision holding up forever.

Your firm’s AI infrastructure deserves the same design principle. Different models have different strengths, some excel at long-form synthesis, some at structured reasoning, some at speed and cost-efficiency for high-volume tasks. An intelligence layer built correctly doesn’t ask you to pick a winner. It routes work to the right model for the job today and adjusts as the landscape shifts without you rewriting your prompts, retraining your team, or rebuilding your workflows every time a provider makes a change. Or every time a government issues a directive at 5:21 on a Friday afternoon.

This is also where the governance conversation from the last article and the diversification conversation in this one converge. A firm that built its AI workflows directly on top of a single model’s API now carries two unmanaged risks at once: the compliance exposure of operating without embedded governance, and the operational exposure of being entirely dependent on one provider’s roadmap and one geopolitical environment’s stability. Neither risk announces itself until something changes. And in this category, something is always changing.

This is precisely why the context layer we’ve described throughout this series has to sit above the model layer, not be welded to it. An intelligence infrastructure that’s model-agnostic by design — that treats foundation models as interchangeable components rather than as the foundation itself — gives firms the same thing diversification gives a portfolio: resilience that doesn’t come at the cost of performance.

The June 12th shutdown wasn’t an anomaly. It was a preview. The firms that were insulated weren’t the ones who predicted it. They were the ones who had already built for it.

Humans lead. Agents scale.