The boundary is the architecture
The first mistake is expecting a small local model to behave like a frontier system with fewer parameters. Frontier models are useful because they are wide. They can absorb weak prompts, incomplete intent, and strange questions because their training and runtime infrastructure cover a large search space.
Local AI needs a narrower job. A laptop model is not useless because it fails at random general chat. It is being misused when it is asked to solve work that has no boundary. The real question is whether the task can be reduced to a defined sequence with controlled input, visible output, and review points.
That distinction matters for privacy and cost. Government work, NDA code, internal data, personal assistants, local voice tools, and sensitive workflows often cannot be pushed casually to hosted APIs. Privacy alone does not make the local model useful. Privacy plus a defined operating model is where the value appears.
SOPs make fuzzy work operable
A standard operating procedure is not paperwork in this context. It is the interface between messy human material and a system that can do something repeatable. The workshop used the simple example of brushing teeth because even an ordinary action contains primitives: grab the brush, apply toothpaste, move through a sequence, finish with a known result.
Business and software work can be decomposed the same way. Once the primitives are visible, model placement becomes less mystical. If a step has clean input and a predictable transformation, it should probably be a script. If the step receives natural language, partial context, ambiguous phrasing, rough notes, voice, images, or other fuzzy input, then an LLM can be useful.
The model should not own the whole process. It should own the fuzzy step that actually requires language or perception, then pass a bounded output into the next deterministic part of the workflow.
The model is only one part of the product
A local model should be biased toward its task. That sentence sounds wrong only if the goal is a perfectly general assistant. For local AI, the useful outcome is usually the opposite: a model that knows the local domain, follows the local pattern, and behaves consistently inside a known workflow.
LoRA-style adapters are useful because they keep the footprint small. A base model can remain compact while the adapter pushes it toward a specific style of code, a translation domain, a company process, or a tool-use pattern.
The hardware ceiling is real. A seven-billion-parameter model at full precision can occupy roughly fourteen gigabytes of memory before overhead. On a sixteen-gigabyte laptop, that leaves little room for the operating system, browser, and the rest of the working environment. Swap memory is not a strategy.
Chatting with a local model is the shallowest version of the idea. The useful question is what the model can do once it is connected to controlled tools. A harness gives the model access to files, search, commands, calendars, MCP servers, and local APIs. That is what turns a text box into a working system.