Autonomy to do what?


With consequences for whom

The Boundary Framework decides how much autonomy an AI system should have by looking at where its consequences land: in the physical world, in law and organizations, or in a particular person's life.

The same autonomy can be right in one setting and wrong in another. Governance follows the consequences, not the technology.

See the framework in action


Short videos from the Bacon Center

Video coming soon

The first Boundary Framework video will appear here.

Three boundaries


Where consequences land

Every action an AI system takes meets the world at a boundary. Each boundary type shows errors differently, corrects them at a different speed, and calls for a different kind of oversight.

Implementation

Consequences land in the physical world: equipment, materials and physical processes.

Governance
Extend established safety engineering: certification, failure-mode analysis, redundancy and containment.
Autonomy
High autonomy within an approved safety regime.
Examples
Warehouse robots, precision sprayers, predictive maintenance.

Institutional

Consequences land in legal, contractual, regulatory, financial and organizational structures.

Governance
Conditional autonomy: defined parameters, accountability tracing, and human review for interpretive or high-stakes matters.
Autonomy
Bounded autonomy for routine, well-specified tasks.
Examples
Accounts payable, regulatory filings, procurement agents.

Normative

Consequences land on a particular person: access, opportunity, liberty, livelihood, health, dignity and fair treatment.

Governance
Preserved human agency: the AI system may assist, and a person takes the consequential action.
Autonomy
Assistive role. A person keeps the final say.
Examples
Hiring screens, benefit decisions, claim denials.

Where does it land?

How fast would we find out?

What keeps it in check?

When an agent crosses a boundary


Goals drift. Oversight has to follow.

AI agents break goals into sub-actions. A goal set at one boundary can produce sub-actions at another, while the system's autonomy setting stays the same. Each step can look routine. In its consequences for a particular person, it is not.

A scheduling agent

Its goal is institutional: cut staffing costs while meeting demand. Along the way it could:

  • Assign shifts (routine)
  • Deny accommodation requests (normative)
  • Cut hours for people nearing benefits eligibility (normative)

Routine scheduling can run on its own. The normative sub-actions go to a manager who decides.

62%

of recorded AI incidents involved boundary crossing, in a Bacon Center analysis of 1,491 incidents in the AI Incident Database.

Put it to work


The governance cycle

Organizations apply the framework at design, at procurement and at periodic review, and repeat it over the life of a deployment.

  1. MapIdentify every boundary the system's actions will touch, including foreseeable sub-actions.
  2. ClassifyAssign each action a boundary type and judge how serious its consequences could be.
  3. CalibrateMatch autonomy and oversight to each action's boundary type and risk class.
  4. ArchitectBuild boundary awareness in, so crossings go to a person by design.
  5. MonitorWatch for new crossings as goals, tools, data and permissions change. Then map again.

Read and use the framework


Free under CC BY 4.0

You may share and adapt this material for any purpose, including commercial use, if you give credit, link to the license and say whether you made changes.

Suggested citation

Howard, M. D., & Murray G. Bacon Center for AI Ethics in Business. (2026). The Boundary Framework: Specification, version 1.0 (Draft for comment). Iowa State University.