Insights
The autonomy threshold: how people and AI agents share the work
Mark Tack, Founder of RealMandates and author of Marketing Rebuilt · October 31, 2026 · 6 minutes readSharing the work is a setting, not a switch
Most conversations about AI agents treat the work as all or nothing: either a person does the work or the agent does. In practice, the most useful question is narrower. For this kind of work, how far should this agent go before a person steps in?
The autonomy threshold answers that question with four levels.
The four levels
- Observe. The agent watches and reports. It notices patterns, flags changes, and tells a person what it sees. Example: an agent that monitors pipeline and alerts the team when deal velocity drops.
- Recommend. The agent suggests and a person decides. Example: an agent that proposes moving budget from one channel to another, with its reasoning.
- Draft. The agent prepares the work and a person approves it. Example: an agent that writes a customer email that a manager reviews before it is sent.
- Act. The agent executes within its mandate. Example: an agent that adjusts bids inside a range a person has already approved.
How to set the level
Four factors decide where a kind of work belongs:
- Reach: does the action touch customers, partners, or the public?
- Money: does it spend, commit, or change revenue?
- Reversibility: can it be undone quickly and cheaply?
- Record: does it change a system of record or speak for the company?
The more of these an action involves, the lower its starting level should be.
| Kind of work | Reach | Money | Hard to reverse | Starting level |
|---|---|---|---|---|
| Monitoring signals | No | No | No | Act |
| Summarizing research | No | No | No | Act |
| Drafting a customer email | Yes | No | Yes, once sent | Draft |
| Moving budget between channels | No | Yes | Partly | Recommend |
| Changing a price or offer | Yes | Yes | Yes | Recommend |
How agents earn more responsibility
Just like a new hire, an agent earns more responsibility by showing good judgment. Responsibility should be earned the way people earn it: through a track record. Start an agent at a conservative level, review its work on a set schedule, and raise the level when its decisions consistently match what the accountable person would have done. Lower it just as readily when they do not.
When people frequently override an agent's recommendations, that is a signal. Either the agent is not ready for more responsibility, or its mandate needs to be clearer.
The rule that never changes
Some actions should always get a human decision, no matter how good the agent becomes: actions with large financial impact, sensitive customer situations, public statements, and anything a regulator or board would expect a person to decide. Confidence is not the same as accountability.
The bottom line
The goal of the autonomy threshold is not to keep agents on a short leash. It is to give them as much freedom as they have earned, and no more, so companies can scale AI with confidence.
Frequently asked questions
Written by Mark Tack. Examples labeled illustrative are composites created to explain an idea; they do not describe a specific company.