gshc2020.com

I get why this feels like a defining decision. After two decades in telecom and software infrastructure, I've watched companies make platform bets that shaped their next 10 years. However, most companies are putting strategic weight on a decision that's becoming less important by the month and missing the one that actually matters.

AI models are becoming commodity infrastructure.​

I've seen this pattern before. In telecom, companies regularly bet big on platforms that looked dominant, only to wake up a few years later to find the market had moved on without them. By the time leadership recognized the shift, switching had become expensive, painful and bogged down by internal resistance.​

AI is moving even faster. METR research found that the length of tasks AI agents can autonomously complete has roughly doubled every four months for advanced models released since 2023. A model that can reliably handle an hourlong task today will be handling tasks eight times longer a year from now.

Also, the lead moves between providers, not just within one. Whoever's leading today probably isn't 12 months from now. Any business architected around a single provider is locked into something already on its way out. Meanwhile, pricing keeps collapsing, and performance gaps that seemed enormous a year ago close in a single release cycle.​

The model is becoming the least unique part of the AI stack. Treating it as a long-term strategic commitment is like committing to a single cloud region in 2010. I've watched companies make that exact bet for two decades. It looks like a safe play today, but it ages badly.​

The real advantage lives in the workflow.​

If the model isn't to your advantage, what is? Everything around it.

Take a regional health system using AI to handle patient communications across dozens of clinics. The differentiation has almost nothing to do with the underlying model. What matters is whether the system recognizes when a routine appointment request actually describes symptoms that require urgent care routing, verifies insurance eligibility across half a dozen payers in real time and handles a patient whose preferred provider has just left the network.

That operational intelligence lives in the workflows, prompts and EHR integrations built around the model, not in the model itself. Swap the model tomorrow, and it all comes with you. A competitor can license the same underlying model and still be years behind.

We built Vida on the principle of a model-agnostic architecture. Build systems so that tasks can be routed across providers based on what works best for a given workflow, not based on which vendor cut the best deal last quarter. As the underlying models keep improving, deployment automatically benefits. No rebuilds. No reevaluation cycles. No regrets.

How does lock-in actually happen?​

Companies rarely choose vendor lock-in. They drift into it.

It starts rationally. A provider offers strong pricing, dedicated support, early access to new capabilities and integration help that accelerates your road map. Six months in, your evaluation pipeline, observability tooling, prompt libraries and internal workflows have all formed around that provider's architecture. The decision you made in week one is now embedded across your stack. When pricing changes or a competitor leapfrogs, the cost of switching is measured in rebuilt operations.

Keep enough separation between your application layer and your model provider that the underlying dependency may evolve without forcing changes across your entire system. That principle is decades old in software engineering. AI's pace just makes it urgent.​

Why do so many organizations still focus on model choice?

Model choice is still the most visible part of the AI stack. Model providers constantly promote new capabilities, benchmark results and product releases, so executives assume that picking the "right" model is their biggest strategic decision.

Model-agnostic architecture hasn't gone mainstream yet because it requires more upfront engineering discipline, while committing to a single provider often feels easier when vendors offer bundled credits, dedicated support and proprietary developer tools that accelerate deployment. In my conversations with enterprise customers, that upfront investment starts to make sense the moment they see how quickly AI capabilities evolve, pricing changes and leading models shift—making flexibility a long-term advantage.​

Ask these questions instead.​

Stop asking which model to pick. Start asking three questions.​

1. If your current provider raised prices 30% tomorrow, how long would migration actually take? If the answer is more than a quarter, your architecture is the problem, not the pricing.​

2. If a new model started outperforming yours on the workflows that drive your business, could your team take advantage without a major rebuild? If not, you've optimized for today at the expense of every tomorrow.​

3. Where is your team's time actually going: improving the model layer or improving the workflows, data and customer experience around it? Only one of those compounds.​

Five years from now, the winners in AI won't be the companies that picked the right model. They'll be the ones that built, so they never had to.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?