The black box problem
There is a familiar argument around local AI: frontier models are expensive, small local models are cheap, so run the routine work locally and keep the big model for the hard problems.
I agree with the logic, but I think the framing is wrong. It treats AI architecture as a cost issue when in fact it is a governance issue.
Most organisations are not just trying to save money. They are trying to deploy systems they can explain, control, and audit.
That is where the black box problem starts to matter. Send a prompt to a third-party frontier API and the reasoning leaves the building. The decomposition, the intermediate decisions, the tool calls, the validation steps—most of it happens inside someone else's infrastructure and is invisible to the people who own the outcome.
That is fine for a consumer app. It is not fine for a bank, a hospital, an insurer, or a regulated business. It is a decision process you cannot inspect, and that is a problem when the decision has consequences.
Most of the discipline never disappeared
Modern AI tools are sold as if we have invented a whole new way of working. In reality, a lot of the good practice is just the old stuff in a new wrapper.
Plan. Do. Check. Act. That is not a new idea. It is how we have managed risk in complex work for decades. We just got lazy when we started letting large language models run at machine speed and called the whole thing “agentic.”
We did not lose the need for governance. We just stopped carrying receipts.
So the solution is not to stop using the model. It is to put a local model at the head of the loop and keep the reasoning inside the organisation unless there is a very clear reason to escalate.
Local orchestrator, deliberate escalation
A local model at the head of the loop is not a smaller version of the big model. It is the orchestrator. It decides how to break the work into steps. It calls the tools the business controls. It enforces policy. It validates outputs. It logs decisions.
That matters because the hard problem is not “can the model think?” It is “can the system decide when the model should stop and when it should escalate?”
When that local loop is working correctly, a frontier model stops being the default and becomes a subcontractor. You send it a narrowed brief, not the raw context. It does the expensive reasoning only because the task is genuinely hard. The local system checks the answer before any action is taken.
That is a meaningful architecture. It creates a chain of evidence. It gives regulators and operators something to look at.
The engineering problems are real
There are two hard problems here, and anyone selling the architecture without saying so is doing a disservice.
First, context tends to degrade over a long agent loop. Memory fills up. The system starts to lose the thread. Without careful summarisation and state management, the workflow becomes less intelligent exactly when it ought to be most capable.
Second, small models are often poor at knowing when they are out of their depth. The real failure is not always a wrong answer. It is a confident answer delivered without enough awareness of its own limits.
That is why the escalation contract matters. Without it, the system is just a machine making a bad call with the wrong level of certainty.
The sovereign loop
The architecture I think matters is the sovereign loop: a reasoning system that runs on infrastructure the organisation controls, keeps the sensitive data in place, and only sends work out when the business has decided the value justifies the risk.
That is not a theoretical model. It is the only version of agentic AI that a regulated enterprise can really get behind without feeling like it is handing over its judgment to a black box.
The bigger story is not “cloud vs local.” It is “where is the boundary, and who owns the decision?”