Your most expensive work is not the work that needs the smartest model. It is the work that repeats. Weekly reporting. Competitor sweeps. First-pass copy across twelve markets. The quarterly deck that gets rebuilt from scratch. That work is high-volume and low-variance, and it consumes senior hours that should be spent on judgment.
So when the open-versus-closed model question reaches your desk, it usually arrives framed as an IT standard. Pick one. Sign the contract. Roll it out. That framing is wrong. It converts a task-level decision into a company-level commitment, and it hands your cost structure to a single vendor's roadmap.
The decision belongs to the work, not the company.
Different jobs have different tolerances. Turning a call transcript into a structured brief has a wide margin for error and a cheap correction path. Drafting a positioning line that will sit on a homepage for two years does not. Treating those as one procurement decision is how companies end up paying frontier prices for clerical work, or running strategic work on a model chosen for its invoice.
A brand system should assign each task to the cheapest model that clears the bar for that task, and the bar should be written down. That is what model-agnostic routing does. It makes the model a variable rather than an architecture. Our own system runs 13 agents this way, and across 44,000+ runs in 63 days total model spend stayed under $50. That figure is not a discount. It is a routing outcome.
Control is the part you cannot rent.
The open-weight case is usually argued on price. The stronger case is control. You can hold the weights, inspect behavior, and keep work inside your own boundary when regulation or IP exposure demands it. The closed case is usually argued on capability, and at the top of the difficulty curve it still holds.
Both arguments are about the model. Neither is about your brand. What actually protects the output is the layer above it: enforceable voice rules, read-only connectors, a human approval gate before anything ships, and audit-grade logs that let you replay any decision. Those controls do not change when the model does. That is what the agent layer is for.
Which is also the real answer to the procurement question. If your controls live in the model, every model change is a migration. If your controls live in your system, a model change is a configuration edit. That is the difference between no lock-in as a slogan and no lock-in as a property you can test before you sign.
What to do with this before your next renewal.
List your ten highest-volume recurring outputs. For each one, write the failure cost in a single sentence: what happens if it is wrong, who catches it, how long the fix takes. You will find that most of them are cheap to correct and expensive to produce. Those route down. A small number are expensive to correct and cheap to produce. Those route up, and they keep the approval gate.
Do that and the open-versus-closed debate stops being a board topic. It becomes a routing table your team can change on a Tuesday afternoon. The strategic question, which work deserves your senior people's attention, stays where it belongs. Your posture is set by what you refuse to automate, not by whose weights you licensed.
If you want that routing table built against your actual workload rather than a vendor's category page, book the strategy blueprint call.
