Answer
Read the grain of the existing operating system first. Then build the smallest version of each of the five pillars (data, tools, agents, governance, cadence) that lets one real workflow run end to end. Then add the second workflow, and the third. AI OS work that starts with infrastructure and ends with use cases almost always stalls. The reverse compounds.
Related answers
Data that is reachable and trustworthy, tools the models can call, agents that handle multi-step work, governance that says what is allowed, and an operating cadence that keeps the whole system maintained. Skip a pillar and the system rots in that exact place.
Pilots prove a model can do a task. Production proves an organisation can absorb the consequences. Most pilots stall because the data pipeline is manual, the tool integration is a hack, governance is a Slack thread, and nobody owns maintenance. The model was the easy part.
Five tiers of AI operating maturity: ad-hoc (individuals using tools), assisted (tools embedded in workflows), augmented (agents extending teams), autonomous (agents owning outcomes), and self-operating (rare). Each rung needs a different shape of AI operating system underneath.