How Much Autonomy AI Agents Get After Go-Live
AI agent autonomy stayed limited after go-live, with 10 of 39 deployments (25.6%) running unattended. The other 29 routed some decisions to a person, and support and IT operations ran 8 of the 10 unattended agents.
Study methodology: In August 2026, 40 practitioners responded to a public journalist query, each describing 1 live AI agent deployment. The study placed every deployment on a 5-level autonomy scale by mapping the workflow its respondent described. One respondent described a market-wide pattern with no single deployment to place, so the level counts use 39. Function cells hold 7 to 13 deployments, and a single coder assigned every level. Most respondents were founders or function heads at small companies, so the figures are not enterprise rates.
Essential statistics: AI agent autonomy in production
- 10 of 39 deployments (25.6%) ran unattended, at level 4 or 5 of the study’s scale.
- 15 deployments (38.5%) ran at level 3, where the agent acts in scope and a person reviews flagged cases.
- 14 deployments (35.9%) ran at level 2, where a person signs off on every proposed action.
- 9 deployments (23.1%) ran 1 workflow end-to-end unattended, and 1 (2.6%) ran across several.
- No deployment ran at level 1, where the agent drafts and a person performs every action.
- Support and CX ran 5 of 9 agents unattended, the highest share of any function. IT operations followed with 3 of 7, ahead of sales at 1 of 10 and engineering at 1 of 13.
- 9 of the 16 respondents who named their own level landed elsewhere once the study coded their workflow.
Key takeaways
- Roughly 1 in 4 live agents ran with no person at the gate.
- Approval by exception was the most common operating mode across 15 deployments, just ahead of approval on every action at 14 deployments.
- IT operations scored the highest mean AI agent autonomy, at 3.57 out of 5, while engineering scored the lowest, at 2.31. The report contrasts this with McKinsey’s 2025 State of AI survey, in which software engineering and IT led in the use of scaled agents in technology, media and telecom. It explains the gap by separating adoption from unattended operation.
- The study treats authoring and release as separate decisions, and agents authored work far more freely than they were allowed to release it.
- Self-reported levels drifted in both directions, so a team’s own label says little about how much its agent runs on its own.
Actionable insights
- Place each agent on a written autonomy scale from its workflow, since 9 of 16 self-reported levels changed once coded.
- Launch new agents on approval by exception, the level at which 15 of 39 teams ran in production.
- Grant authoring and release as separate permissions, since agents in the study authored freely while people held release.
- Set AI agent autonomy targets per function, since support ran 5 of 9 agents unattended and engineering ran 1 of 13.
“Unattended authoring is near-total. Unattended release is close to zero, deliberately.” — Beck Wang, Founder, IfShe









