IDD concept: Govern execution
Autonomy Boundary
The Autonomy Boundary is the line between the decisions and actions an organisation delegates to autonomous systems and those that remain with people. Risk Dials move it step by step as evidence accumulates, while a small set of decisions stays permanently on the human side.
A boundary that has always moved
The history of software development can be read as the story of a boundary moving: the line between human intent and machine execution. Each generation of technology, from high-level languages to frameworks, cloud platforms, DevOps pipelines and now AI, has moved execution to the machine side and left people closer to intent. Organisations adopting agentic AI are not facing a new kind of decision so much as a faster and more consequential version of an old one.
There is a natural limit to how far that boundary can move. Machines can execute, optimise and automate, but they do not decide what should exist, which trade-offs are acceptable or which outcomes matter. That remains human work, and the purpose of naming the Autonomy Boundary is to make sure it stays that way by design rather than by accident.
How the boundary moves
In practice the boundary runs through every step of every workflow rather than being drawn once for the whole organisation, and its position at each step is set by a Risk Dial. A step at Red sits firmly on the human side; a step at Green has been delegated, with people notified and able to intervene. Moving a dial moves the boundary, and IDD allows that only when Intent Fidelity shows sustained alignment.
The boundary moves in both directions. A change of model, a new kind of work or a fall in measured fidelity are all reasons to move it back, and regression under uncertainty is a sign of disciplined risk management rather than failure. It can also sit in different places for different agents, since specialist agents earn trust at different rates.
What stays on the human side
Some decisions never cross the boundary, however mature the organisation becomes or however good the measurements look:
- Strategic intent: which problems to solve, what success means and what the organisation is optimising for.
- Ethical boundaries: whether something should be built, who could be harmed and what the implications are.
- Security and compliance: production system security, the handling of customer data and regulatory obligations.
- Final production approval: whether a change is ready for customers, and what happens if it is not.
Other areas, such as regulatory decision-making, high-impact ethical trade-offs and strategic domain modelling, may also reasonably remain permanently human-controlled. Delegating execution does not delegate responsibility.
Designing for delegation
Work can only cross the boundary safely if it was designed to. That depends less on the capability of the agent than on the quality of the intent it receives, and in particular on four things: a bounded scope, explicit success criteria, a transparent statement of constraints, and clear identification of the boundaries that must not move. When intent is underspecified, both human and machine execution become fragile; when it is well defined, delegation can expand within controlled limits.
Maximum autonomy is not the goal
It is easy to treat a boundary as something to be pushed as far as possible. IDD does not. The objective is sustained alignment, not maximal automation, and for many organisations the right position leaves substantial work delegated while high-risk decisions remain firmly human. The Maturity Model reflects this directly: reaching a plateau can represent optimal calibration rather than incomplete progress.
Where it comes from
The Autonomy Boundary draws on the prologue of the ebook, on the decisions that never leave human control set out in Human Gates in Agentic Flows, and on the design for delegated autonomy in Designing Intent at Scale. It runs through Chapters 5 and 9 of the ebook.