AI·June 18, 2026·7 min
Where the data sleeps.
Local models are not a religion. They are an address. Copilot, Claude, and a box in the closet can all be the right call — once you say where the file is allowed to live.
A lot of AI shopping starts with the model name. The useful question is duller: where is this file allowed to sleep tonight? If the answer is “in the plant,” you are in Ollama, llama.cpp, LM Studio, vLLM territory. If the answer is “in the tenant we already pay for,” Copilot is the door. If the answer is “a vendor we trust with a contract,” Claude or GPT can come to work.
We have watched companies buy a cloud copilot and then paste customer drawings into it, because nobody wrote the rule. We have also watched companies refuse every cloud tool and then fail to run a local model that anyone will actually open. Both are a desk problem, not a model problem.
Pick the brain. Keep the throttle.
The pattern that holds: the model proposes, the workflow carries, a person posts. Agents earn their keep on triage and routing. They do not hold the pen on a PO, a credit, or a change to a recipe. That is the human gate — the same idea as a NOC that calls before you notice, applied to language.
If you cannot say, in one sentence, where the document goes when the model is done with it, you are not ready to buy the model. You are ready for the audit of the work.
AI
The agent should not write the PO.
Multi-agent demos look like magic until someone lets them post. Workflow-first means the model never holds the pen.
The packing slip that trained your first real model.
Document intelligence is not a moonshot. It is the unglamorous work of teaching a system what your receiving dock already knows.