9Q / AI autonomy levels
AI autonomy levels: how much should an AI agent do on its own?
There is no single right level of AI autonomy: it depends on the work. In Suhit Anantula's 9Q Grid, AI can Assist (the human leads, at the helm), Augment (human and AI iterate, the human in the loop) or Adapt (the AI acts and humans govern boundaries and exceptions, on call). The right level is different for structured, dynamic and complex work.
The 9Q Grid is Suhit Anantula's map of AI agency and type of work. Read the 9Q Grid.
Assist / Augment / Adapt × structured / dynamic / complex
Why one autonomy scale isn't enough
The same autonomy level means different risk on structured, dynamic and complex work. A five-level list that ignores the type of work will mis-set the oversight you need.
Three levels of AI agency
| Level | What the AI does | Human role | Typical oversight |
|---|---|---|---|
| Assist | AI helps; the human does the work | At the helm | Human runs every step; AI supports |
| Augment | Human and AI iterate together | In the loop | Human approves or shapes each step |
| Adapt | AI acts; humans set boundaries, check outcomes and govern exceptions | On call | Human governs aims and exceptions, not each transaction |
The same autonomy on different work
Adapt plays out differently across the three types of work — and each has a book failure mode:
- Q3 Automator (structured): silent drift — automation running on last quarter's rules.
- Q6 Optimiser (dynamic): optimising the wrong thing at scale.
- Q9 Conductor (complex): orchestra drift, where governance turns ceremonial.
More human involvement has risks too. On the Augment row: Q2 recommendation capture, Q5 dyad dependency, Q8 simulation capture. More involvement is not automatically safer.
Human in the loop, on the loop, or on call?
| Phrase | 9Q reading |
|---|---|
| Human in the loop (HITL) | ≈ Augment — the human approves or iterates each step |
| Human on the loop (HOTL) | ≈ Adapt with active monitoring |
| On call | ≈ Adapt where humans govern exceptions |
| Out of the loop | Humans never fully leave; they govern aims and exceptions |
For the organisation and agent-autonomy angle, stay on this page. For how Assist, Augment and Adapt show up when placing initiatives, see how to prioritise AI use cases.
How to choose the level for an AI agent
- What type of work is it — structured, dynamic or complex?
- What trust and governance have we already earned? (trust gradient; see the AAA Licence chapter)
- What's the failure mode, and who sees it first?
The 9Q Grid is not a ladder. Q9 is the most demanding quadrant, not the best one.
Read more on Position is not permission (AAA Licence).
Agentic AI governance by quadrant
Governance must grow with agency. On the Adapt row, watch for silent drift (Q3), optimising the wrong objective (Q6), and orchestra drift (Q9). Boards and risk committees should ask which quadrant each agent sits in before they approve more autonomy — see the 9Q Grid for boards.
Helix Lab work on AI governance for boards and public-sector programmes lives at government and enterprise.
A note on other autonomy scales
Research frameworks such as the Levels of Autonomy for AI Agents paper (arXiv 2506.12469) often use five user-role levels. Not to be confused with the "operator" user role in that paper: in the 9Q Grid, Q1 Operator means AI assisting a person on structured work. Those scales grade autonomy on one axis; 9Q crosses agency with the type of work.
FAQ
What are the levels of autonomy for AI agents?
Research frameworks often use five user-role levels. The 9Q Grid uses three levels of AI agency (Assist, Augment, Adapt) crossed with three types of work.
What is the difference between human in the loop and human on the loop?
In the loop: a person approves or shapes each step. On the loop: the system acts and a person monitors and can intervene.
What does autonomy mean when managing an agentic AI system?
The scope of decisions the agent can take without a person, and who governs the boundaries.
What is autonomy vs automation?
Automation follows fixed rules. Autonomy adapts toward a goal. In 9Q terms, Q3 automates structured work, while Q6 and Q9 adapt.
Is ChatGPT human-in-the-loop?
Mostly used as Assist or Augment: a person prompts, reads and decides.
Helix work in this area
With the City of Mitcham, Helix narrowed 51 use cases to 3 pilots and ran a six-week AI Pod. Helix case studies ↗
Related questions
Continue the work
Turn your map into a next move.
Use the 5Ps to explore responses to what the map reveals, test a move and carry the learning into the next loop.
For AI software options, explore Agent Watch ↗.