Six levels of operations autonomy

Why a model

"We use AI in operations" can mean a chat window or an agent that restarts services at night. The useful question is not which model you run but who approves a change in production. The levels below describe that, one step at a time.

The levels

• Level 0. People only. Every investigation and every change is done by hand.
• Level 1. Agents answer. Agents read production and answer questions in chat. They change nothing.
• Level 2. Agents investigate and propose. Every alert arrives already investigated, with cause and evidence. Fixes are proposed as pull requests that people review. The agents have learned your system and passed an examination on your past incidents.
• Level 3. Agents act on approved plans. A named person approves a bounded plan in chat; the agent carries it out with access that exists only for that plan.
• Level 4. Agents earn autonomy. The first kinds of action, proven by a record of approved, harmless runs, run without waiting for a person, within limits, and fall back to approval on the first mistake.
• Level 5. Production only agents touch. Routine operations are autonomous under policy people wrote. People have no standing access to production; they set policy, approve what policy does not cover, and read the audit.

Moving up

Each step is earned with evidence, not declared. Autonomy is granted per kind of action and per part of the system, never for "the agent" as a whole, and the riskiest kinds of change stay with people at every level. Most teams that start today can reach Level 2 quickly; Level 3 needs the access model in place; Levels 4 and 5 follow from the record.

Where we start

An assessment places your operations on this scale and lays out what each next step needs: knowledge, access, approvals and the evidence to earn the next level.