Result
Highest autonomy the risk supports
Enter tasks to see the result.
Calculator Library / AI Operations
Agent Autonomy and Oversight Planner
Score each task you might hand to an AI system or agent, and get the highest autonomy level the risk supports, from assist to act with monitoring, with the oversight controls each level needs.
Tool
Plan autonomy and oversight for each task
Risk index = 0.4 impact + 0.3 irreversibility + 0.2 data + 0.1 compliance
List the tasks you want an AI system or agent to do. Score each from 1 to 5 on how bad a wrong result would be, how hard it is to undo, how sensitive the data is, how much compliance exposure there is, and how easy it is for a person to check the output. The tool gives the highest autonomy level the risk supports and the controls that level needs. Load the example to see six tasks.
| Task | Impact of an error | Irreversibility | Data sensitivity | Compliance exposure | Ease of checking |
|---|
Impact, irreversibility, data sensitivity, and compliance: 1 is low and 5 is high. Ease of checking: 1 means a person can hardly verify the output, 5 means checking is quick and reliable.
Entries stay in your browser and are saved there between visits on this device. Nothing is sent to a server. The thresholds are examples: set your own in your risk assessment. The tool supports a decision and does not replace one.
Controls
What each level needs
Instructions
How to use this planner
- List tasks, not whole jobs. One row for each action you might hand to an AI system.
- Score the five factors with the people who know the work and the risks. Score the worst credible outcome, not the usual one.
- Read the highest level the risk supports. It is a ceiling, not a target: start lower, and move up only with evidence.
- Check that the controls listed for that level are in place before you start. If you cannot provide them, choose a lower level.
- Record the result in your agent charter and review it when the task, the data, or the system changes.
What This Agent Autonomy and Oversight Planner Does
The planner helps you decide how much freedom to give an AI system or agent for each task, and what oversight it needs. You score five factors from 1 to 5: the impact of an error, how hard the action is to undo, the sensitivity of the data, the compliance exposure, and how easy the output is to check. It combines the first four into a risk index, finds the highest of five autonomy levels the risk supports, lowers it by one level when the output is hard to check, and lists the controls that level needs.
It matches the levels and examples in the Human-in-the-Loop Process Design guide and the AI Agents in Operations guide.
How the Level Is Chosen
| Step | Rule (editable thresholds) |
|---|---|
| Risk index | 0.4 × impact + 0.3 × irreversibility + 0.2 × data sensitivity + 0.1 × compliance exposure |
| Level the risk supports | Risk index of 4 or more: level 1. At least 3: level 2. At least 2: level 3. Below 2: level 4 |
| Check for hard-to-check output | If ease of checking is 1 or 2, lower the level by one (not below 0) |
| Levels | 0 Assist, 1 Draft, 2 Act with approval, 3 Act with humans sampling, 4 Act with monitoring |
Worked Example
The pre-loaded example scores six tasks at an illustrative plant. Classifying and routing complaints (risk index 2.6) supports level 3. Closing low-priority work orders (1.7) supports level 4. Sending replies to customers (3.7) is capped at level 2. Changing the production schedule (3.1) supports level 2, but hard-to-check output lowers it to level 1. Updating a supplier’s quality status (4.6) supports level 1, lowered to level 0 because it is hard to check.
Limits and Good Practice
- The weights and thresholds are examples. Set your own in a risk assessment with the people who know the work, and have qualified people review high-risk uses.
- The level is a ceiling. Start lower and move up only with evidence from testing, shadow runs, and monitoring.
- Scoring the worst credible outcome matters more than scoring the usual one.
- The tool does not assess whether a use is allowed. Legal, privacy, and safety requirements apply whatever the level.
- Entries are saved in your browser on this device. Nothing is sent to a server.
Agent Autonomy and Oversight Planner Frequently Asked Questions
What are the five autonomy levels?
Level 0 Assist: AI suggests and a person does the work. Level 1 Draft: AI drafts and a person checks and sends everything. Level 2 Act with approval: AI prepares each action and a person approves it. Level 3 Act with humans sampling: AI acts and people review a sample and all exceptions. Level 4 Act with monitoring: AI acts and automatic checks and alerts watch it.
Why does hard-to-check output lower the level?
Because oversight only works if a person can tell whether the output is right. If checking is as hard as doing the task, the review step gives little protection, so the task needs more human involvement, not less.
Can I choose a higher level than the tool gives?
The result is a ceiling based on the scores. A higher level would need lower risk scores, which means changing the task: for example, limiting what the system can do, adding checks that make errors easier to catch, or making actions reversible. Rescore after such changes.
Who should do the scoring?
The process owner, the people who do the work, and someone who understands the risks, together. Differences between scorers are useful: they show where people see the risk differently.
Is my data stored anywhere?
Your entries are saved in your own browser on this device so they are there when you return. Nothing is sent to a server.