The most expensive decision in an agentic rollout is rarely the model. It is the standardization meeting that follows it.
A team picks a provider. Prompts get tuned to that provider's quirks. Output formats get shaped around its habits. Six months of institutional knowledge accumulates inside one vendor's interface. Then the pricing changes, or a stronger model ships, or procurement asks a question nobody can answer. The switching cost is not the contract. It is everything the team built on top of it.
Open or closed is the wrong first question.
The debate usually arrives as a values choice. Open weights mean control and inspectability. Closed frontier models mean capability and support. Both framings assume you are picking one.
You are not picking one. You are deciding where the choice lives.
In a well-built brand system, the model is a component. It sits behind a routing layer that decides which task goes to which model, and that decision is a setting, not a rewrite. Drafting a long-form argument, classifying inbound signal, and checking a claim against a fact set are three different jobs with three different cost and quality profiles. Routing them identically is how model spend climbs while quality stays flat.
Our own system runs model-agnostic routing across thirteen agents. In sixty-three days that produced more than 44,000 runs at under $50 in model spend. That is not a boast about frugality. It is what happens when cheap work goes to cheap models and expensive judgment goes to expensive ones, automatically, on every run.
Portability is a design decision, not a clause.
Portability gets written into contracts and then quietly abandoned in architecture. The clause says you can leave. The system says you cannot.
What actually travels between models is the layer above them: the brief, the brand rules, the approval gate, the logs. If your voice standards are enforceable rules rather than a mood board, they hold when the model underneath changes. If your connectors are read-only, swapping a model does not reopen a security review — that is the point of a defensible posture. If every run carries audit-grade logs, you can test a new model against your own work instead of a public benchmark. That is the whole evaluation.
Companies built this way hold an option most do not. They can adopt the next capable model in an afternoon. They can decline the next expensive one without drama.
Three questions to ask before you standardize.
Can you name the task types running through your system, and does each one have a routing rule? Can you move a workload to a different model without rewriting the brief that governs it? If your provider doubled its price tomorrow, how many days until you are running somewhere else?
If that last answer is measured in quarters, you did not buy a component. You bought a dependency.
The strategic asset was never the model. It is the accumulated brand memory, the enforced standards, and the operating discipline around them. That layer compounds. The model layer keeps getting cheaper and better underneath it, which is only an advantage if you built to swap. The ninety-day sequence for getting there is laid out in the journey.
If you want a read on where your own stack is dependent rather than portable, book the strategy blueprint call.
