Levels of AI autonomy in creative and account work

Tasks each level can suit, as examples:

• Level 1: creative for publication, client strategy, anything making a claim.
• Level 2: intake conversations, gathering materials from a client.
• Level 3: internal summaries, reformatting, routine reporting.

The same AI can sit at different levels for different tasks and different brands. Five questions usually settle it:

• Who sees it? Internal work can run with more freedom than anything a client or the public sees.
• Can it be undone? A reformatted file can be redone. A published claim or a sent promise is harder to take back.
• Does it bind anyone? Commitments, budgets and timelines belong with people.
• How regulated is the category? A brand with legal or regulatory exposure needs more supervision on the same task.
• Has this workflow earned it? Past approvals are the evidence for moving up.

The more public, binding or irreversible the task, the lower the level.

S77.AI's rule is that a workflow moves up a level only as its approval rate proves out: how often people approve the AI's work, how often they edit it, and how often they reject it. Each workflow has to earn this on its own, because a good record on one task says little about another.

S77.AI's Autonomous AI Agency is in supervised pilot with one AI staffer, and it uses two of these levels:

• Level 2, for intake conversations: its replies reach the client live, without approval of each reply. A scope lock written into the software limits it to brand intake: no commitments, pricing, product claims or legal advice. Anything beyond that gets flagged to a person. It confirms it is an AI when asked, and its supervisor can read the full transcripts.
• Level 1, for what it produces: the brand core it builds from cited evidence lands as a proposal that a senior human approves, edits or rejects.

Further AI roles are added only as each workflow's approval rate supports it.

1. Which tasks run at which level of supervision?
2. What is the agent not allowed to do, and is that enforced in the software or only in its instructions?
3. Does it say it is an AI when asked?
4. Who can read its conversations and logs?
5. What happens when a request falls outside its scope?
6. What record decides when a workflow gets more autonomy?

The EU AI Act's human oversight article applies to high-risk AI systems, which most marketing work isn't. It uses similar logic, though: oversight should match the risks, the system's level of autonomy and the context of use, and people must be able to override the output or stop the system.