How a company's switch to local AI models quietly broke its dev workflow

A software firm of around twenty developers initially found success using AI tools collaboratively, with developers retaining control over architecture and using AI for targeted assistance. Leadership then shifted to a one-developer-per-project model after the rise of vibe coding, expecting dramatic speed gains, but results fell short of expectations while AI subscription costs rose sharply. To cut costs, management mandated a switch to a single in-house local AI model running on one dedicated machine shared across all twenty developers. The setup proved inadequate for agentic workflows — where AI plans, edits multiple files, runs tests, and iterates — which the team had come to rely on, as one machine could not realistically serve twenty concurrent users at that capacity. Developers quietly resorted to using external AI tools for planning and reasoning, then feeding outputs into the approved local model, effectively circumventing the policy to get work done.