"What's our AI strategy" usually comes out as a question about models. Should we be on OpenAI or Anthropic. Are we ready to swap to whichever provider released the best benchmark this month. Did we miss a moat by picking wrong.
We get this question often enough that it's worth saying directly: the model isn't where the durable advantage lives. There are perhaps three companies in your industry with a real shot at a model-level moat, and you aren't one of them. Everyone else gets the same models, on roughly the same timeline, at prices that converge fast. Your competitors are already running the model you wish you had.
The advantage is everywhere except the model.
What's left is the work around the model. The data you feed it. The prompts you've actually tuned for your operations. The agents wired into your real systems. The evaluations that tell you what's good enough to ship. None of that is portable. None of it arrives in a release note. None of it gets cloned by a competitor swapping providers.
We've watched this play out in client engagements. Two companies in the same niche, both running the same frontier model. One has a structured intake form, a domain-specific eval suite, and a clear escalation path when the agent isn't confident. The other has a chat window. Same model. Wildly different business outcomes. The first one wins on a workflow advantage that took months to build — and another six months for a competitor to plausibly catch up.
Two practical implications.
This reframes the strategy question entirely. The right question isn't which model are we standardizing on. The right question is: where, specifically, in our business does AI need to plug in — and what would it take, in our data, our tooling, our processes, to make that integration excellent. The answer to that is durable. The answer to the model question changes every quarter.
First: assume you'll switch models. Build accordingly. Keep your prompts and tools provider-portable. Treat any vendor lock-in as a cost, not a feature. The models will keep changing under you for years.
Second: spend your budget on the workflow, not the subscription. A more expensive model rarely matters compared to a small amount of well-placed integration work. We've replaced "we need a smarter model" engagements with two weeks of plumbing more times than we can count.
The moat is real. It just isn't where most teams are looking for it.