How to Invest in AI When the Next Model Update Could Change Everything

There's a question that keeps coming up in conversations with the companies and investors we work with. The phrasing changes, but the worry underneath doesn't: How do we invest in AI when the next model release could wipe out everything we just built?

It's a fair concern. OpenAI, Anthropic, Google, and a growing roster of players are shipping new capabilities at a pace that makes quarterly planning feel quaint. According to Stanford's 2025 AI Index Report, inference costs have dropped dramatically over the last 18 months. Features that required custom engineering six months ago, now come bundled into off-the-shelf APIs. The instinct to wait for the dust to settle makes sense on the surface.

But waiting is its own risk. In our experience, the companies gaining ground aren't the ones who picked the "right" model. They're the ones who built their AI investments so that they don't depend on any single model staying on top.

The model race is real, but it's the wrong place to focus

The numbers speak for themselves. Total corporate AI investment reached $252.3 billion in 2024, with private investment jumping 44.5% year over year, per Stanford's AI Index. Enterprise adoption rose from 55% to 78% in a single year. Nobody in a boardroom needs convincing that this is a big deal.

What we keep telling clients, though, is that the models themselves are converging. Performance gaps between leading LLMs narrow with each release cycle. MIT Sloan researchers noted earlier this year that competition is shifting away from models and toward the systems built on top of them. In plain terms: if you're choosing between two or three leading models, the difference in raw capability gets smaller every quarter. What actually separates one company's AI from another is the layer above the model, the data flowing into it, the business rules governing its outputs, the integrations connecting it to real workflows, and how easily you can swap one model for the next when something better shows up.

So the real exposure for a mid-market company (or a PE-backed portco) isn't that a model goes stale. It's that everything was wired so tightly to one provider, one API version, or one specific configuration that upgrading becomes a re-architecture project. That's where budgets get burned.

And it's worth paying attention to this insight from MIT Sloan Management Review: tech debt tied to AI compounds faster than traditional software debt because the underlying technology moves at a pace that enterprise planning cycles simply weren't designed for.

The cost problem nobody expected

Here's something that doesn't get discussed enough: the ongoing cost of running AI can change dramatically between one model version and the next.

This is becoming a real conversation among engineering teams. A company builds a workflow on a specific model, the per-token costs are manageable, leadership signs off on the budget. Then the provider ships an update. The new version is more capable, but it also consumes more tokens per interaction, or the pricing structure changes, or the older version gets deprecated and the replacement costs more to run at the same volume. The workflow still works, but the economics underneath it have shifted. With the same budget that used to cover full AI-assisted operations, the team now has to be more selective about where they apply it, handling some tasks manually again to stay within spend. Developers who were leaning heavily on AI coding assistants, for instance, are now rationing their usage, picking which tasks justify the token spend and doing the rest by hand.

The takeaway is straightforward: when you build on AI, you need to plan for cost volatility, not just technical obsolescence. That means building in the ability to route between models depending on the task (using a lighter, cheaper model for routine operations and reserving the most capable model for where it truly matters), and it means keeping your architecture flexible enough to switch providers when the economics shift.

How this shows up when you look under the hood

A year ago, a 2025 MIT study reported that 95% of enterprise AI pilots showed no measurable P&L impact. The landscape has improved since then, with a 2026 Gartner survey showing that roughly 35% of organizations have scaled AI to deliver real business value. What separates that 35% from the rest, though, tends to come down to how they built, not what model they chose.

The patterns above (tight coupling, cost drift) are easy to describe in the abstract. What makes them harder to catch is that they often look fine on the surface. It's only when something changes, a model update, a pricing shift, a new compliance requirement, that the fragility becomes visible.

In our work at Making Sense, the clearest signal is what happens when you ask a team: "What would it take to swap the model behind this workflow for a different one?" If the answer involves weeks of re-engineering, touching application code, or rebuilding integrations from scratch, the architecture wasn't built for a world where models change. And models change. We've written about how to surface these kinds of structural dependencies during diligence, and in a world where models update every few months, it's worth asking early.

What it actually means to build AI that survives the next update

"Futureproof" is an overused word, so let's be specific. What we mean is: you can swap the underlying model without re-engineering the business logic, the data connections, or the workflows your team actually uses. Think of it like a building designed to accommodate different tenants without a gut renovation every time one leaves.

MIT's Center for Information Systems Research published a maturity framework that maps this well. Companies achieving above-average financial performance from AI tend to build on modular platforms with strong data governance and a tight link between AI initiatives and business strategy. Instead of organizing around a specific model, they organize around the problem they're solving.

A few principles that make this real:

Separate the model from the workflow. API wrappers, middleware, orchestration layers. This is a relatively small amount of engineering upfront that saves you from a forced migration when (not if) a better model arrives or your current provider changes terms. When there's an abstraction layer between your application and the model, a deprecation notice becomes a configuration change instead of a re-architecture project.

Invest in data before you invest in models. Not glamorous, but extremely important. Your data pipelines, quality controls, and access governance will outlast every model you ever deploy. A company with clean, well-structured data can switch models in weeks. Without that foundation, it doesn't matter which model you're running. We saw this clearly in our work with Auto Approve, where building a solid data pipeline and validating business hypotheses through analysis came before any automation effort. That groundwork is what made everything that followed actionable.

Build in cost flexibility from day one. Design your system so you can route different tasks to different models based on complexity and cost. Not every customer inquiry needs the most powerful (and most expensive) model. This kind of tiered routing is what keeps AI economics sustainable as pricing shifts.

Start with a focused discovery, but keep it fast. Every AI engagement we run at Making Sense begins with understanding the actual business problem before choosing tools. But discovery doesn't need to be a long, drawn-out phase; a focused sprint can give you the clarity you need to build something that holds up. The companies that skip this step are disproportionately the ones rebuilding a year later.

For anyone exploring where to start with AI adoption more broadly (which model type, which use cases first), we've covered that in detail in a recent post focused on PE-backed companies and in our breakdown of AI Jumpstart Kits.

Why this matters to PE firms evaluating a portco

If you're a PE professional running diligence or building a 100-day plan, this question about model resilience feeds directly into valuation. A portco whose AI is built on modular architecture, with clean data and the ability to swap models, is a very different asset than one that's locked into a single vendor with no abstraction layer and climbing token costs.

We've written extensively about how tech due diligence creates value across the hold period and about the questions that reveal whether a portco's AI strategy is real or performative. The model-update risk is one more lens to add to that evaluation.

Speed still matters, but so does the foundation

None of this argues for going slow. MIT Sloan faculty have been clear that 2026 is the year AI moves from experimentation to operational scale. Companies still running disconnected pilots while competitors embed AI into core workflows will lose ground.

But moving fast on a weak base has its own cost. "Zombie AI" (systems that drain budget without delivering value and have no viable upgrade path) is a growing category of technical risk. These are implementations built quickly on shaky foundations, and now the organization can't retire them, upgrade them, or explain their ROI.

The play is rapid implementation on a modular foundation. Weeks, not quarters. But on architecture designed to absorb the next wave, not get swept aside by it.

The bottom line

The fear that a new model release will make your AI investment obsolete is understandable. In most cases, it's also misplaced. Models will keep evolving, costs will keep shifting, and new capabilities will keep arriving. For a well-built system, that's fuel.

If you're evaluating AI for your company or assessing a portfolio company's tech stack, the question worth asking isn't "Which model should we use?" It's "Are we building something that gets better every time the technology improves, or something we'll have to tear apart, or pay dearly to adapt?"

We help companies answer that question every day.

Let's talk about building AI that holds up, no matter what ships next →