Most organizations selected their AI model the same way they select an ERP. RFP, evaluation, procurement, done. A committee assembled, a shortlist was built, a contract was signed, and the decision was filed away as resolved.

That playbook has served enterprise software buyers well for decades, and the instinct to apply it to AI is completely understandable. It's what procurement knows how to do. It's what legal knows how to review. It's what the board knows how to approve.

The problem is that playbook was built for a world where software doesn't change quarterly and switching costs are measured in years, not sprints. ERP implementations take eighteen months and cost eight figures. You evaluate carefully because reversing the decision is brutal. AI model selection works nothing like that, and organizations that treat it like it does are making a category error that will compound quietly until it doesn't.

Enterprise software pricing is predictable. Seats, tiers, renewals. A CFO can model it in a spreadsheet and trust the output for three years. The tools built around that predictability, ROI projections, total cost of ownership analyses, cost-per-seat calculations, have become standard practice because they work. The numbers are stable enough to be useful.

AI model costs don't behave like anything in that toolkit. They vary by task, by token count, by model version, by use case. A support workflow running GPT-4o costs something different than a document classification pipeline running the same model, and both cost something different next quarter when the provider reprices. It's a bit like buying a farm where the price of the land changes every time you plant a different crop. The CFO who applies a three-year TCO model to AI spend isn't doing bad analysis. They're using the right tool on the wrong material.

I've sat in enough of these planning conversations to know what happens next. The projections get built, the board approves the budget, and six months later nobody can explain why the actual spend looks nothing like the model. That isn't a forecasting failure. It's a signal that the underlying economics were never predictable to begin with, and the measurement framework needs to change before the decisions can improve.

Every other major capital allocation a business makes has a feedback loop attached to it. Marketing spend has attribution. Headcount has output metrics. Real estate has utilization rates. AI is the only significant line item on the P&L where the money goes in and the outcome comes back as a feeling.

That isn't a technology problem. The models are fully instrumented. The APIs return latency, token counts, error rates. The data exists at the infrastructure level. What doesn't exist yet for most companies is the layer that connects infrastructure data to business outcomes: cost per task, per workflow, per team, per decision made. Without that layer, model selection will keep feeling like a vendor decision because there is no other frame available to make it in.

Standardizing on one vendor made sense when the switching cost was migration. Moving a database, retraining a team, rebuilding an integration layer. Those are real costs that justified locking in early and staying put. The calculus was sound because the friction was real.

With AI, you're not migrating a database. You're rerouting an API call. The technical switching cost has collapsed, and that changes everything about how lock-in should be evaluated. The organizations treating model selection like a vendor consolidation decision are paying a strategic tax that doesn't show up on any invoice.

The future isn't one model for everything. It's Claude for reasoning, GPT-4o mini for speed, Gemini for document scale. Each task matched to the model built for it, each workflow running on the economics that actually make sense for that use case. The organizations that get there first won't just have better unit economics. They'll have a fundamentally different relationship with AI, one where the infrastructure serves the outcome rather than the contract serving the vendor.

Model selection isn't a decision you make once in a boardroom. It's a practice. The same way a finance team doesn't set a budget in January and ignore the P&L until December, an organization running AI at scale can't evaluate models once and assume the answer holds. The landscape changes too fast. The economics shift too often. The use cases multiply in ways nobody planned for at procurement time.

The organizations that will pull ahead are the ones that treat AI like a utility they actively manage rather than a contract they signed and filed. That means reviewing model performance by task, by workflow, by team on a cadence that matches how quickly the underlying technology moves. It means having the instrumentation to make that review meaningful rather than anecdotal. It means the people making model decisions are operators working from data, not procurement teams working from proposals.

I was talking to a founder recently who described his company's AI setup as a "set it and forget it" infrastructure decision made eighteen months ago. Smart team, real usage, genuine scale. When I asked him which model was handling his highest volume workflow and what it was costing per task, he didn't know. That's not an edge case. That's where most organizations are right now, and it's exactly the gap that separates a managed AI operation from an unmanaged one.

Every major technology transition has the same pattern hiding inside it. The companies that win aren't always the ones with the best technology. They're the ones that recognized earliest that their old mental models were the constraint, not the technology itself. The retailers that lost to e-commerce weren't beaten by a faster checkout. They were beaten by companies that stopped thinking about retail as real estate and started thinking about it as logistics. The mental model was the moat, and then it became the trap.

The contract you signed is not a strategy.

AI is following the same pattern. The organizations still asking "which model did we choose?" are asking a procurement question in an operations environment. The right question is whether you actually know which model is earning its cost across every task, every workflow, every team running it today. That question doesn't get answered in a boardroom. It gets answered by instrumentation, by measurement, by treating model selection as a living practice rather than a resolved decision.

The companies that make that shift early will compound the advantage every quarter. The ones that don't will keep signing better contracts while their competitors build better operations. Those are different games, and only one of them produces durable leverage.