Local LLMs hit workflow reliability limits

Local LLMs can work in production workflows, but the winning pattern is routing and triage, not free-form autonomy. The moment the model’s answer drives side effects, the real problem becomes orchestration, validation, and fallback design.

• –Several commenters describe real deployments for message triage, escalation, and media processing where the model chooses among a small set of actions.
• –JSON-only outputs, few-shot prompts, and custom parsers help more than simply moving to a larger quantized model.
• –For anything risk-bearing, a hybrid path still looks strongest: low-risk stays local, medium-risk escalates to a stronger API model, and high-risk goes to a human queue.
• –If the business logic needs repeatable structure, code should own the hard rules and the model should only handle ambiguity.
• –The thread reinforces a broader pattern: local models are already useful as decision helpers, but brittle as the final authority.