
For your company
Foundations
Module · Where does AI stop and you start?
The Boundary Layer Assessment
Every AI deployment draws a line between what the machine decides and what a human still decides. Drawn well, per decision class and by stakes, that line is your main safety control; drawn by accident, it is your main risk. This module checks where the line actually sits, how explicit it is, and whether the human side is real.
What the five levels look like
Every dimension in this assessment is scored 1 to 5. This is what the levels mean, dimension by dimension. The graded report diagnoses where your own answers land and what to do about it.
The line is explicit
- 1Nowhere stated
- 2Implicit only
- 3Some classes written
- 4Most classes written
- 5Explicit per class
At the low end: An unwritten line is drawn by whoever ships next, not by you. Write it for your three highest-stakes decision classes this week; it is a paragraph each, not a project. What good looks like: An explicit line per decision class is the foundation of deliberate autonomy. Keep it versioned, so you can see when and why the line moved.
Placement matches stakes
- 1Convenience decides
- 2No logic
- 3Loosely by risk
- 4Mostly by stakes
- 5Stakes and reversibility
At the low end: Drawing the line for convenience puts automation exactly where it hurts most when it fails. Re-place your riskiest decisions by asking what an error costs and whether you can undo it. What good looks like: Placing the line by stakes and reversibility is how mature programmes decide what to automate. Revisit it as reversibility changes; a new refund policy can move the line.
The boundary is staffed
- 1Unstaffed
- 2Token reviewer
- 3Understaffed
- 4Adequately staffed
- 5Staffed and skilled
At the low end: A human side nobody has time to staff is automation in disguise. Count how many decisions your reviewers actually face per hour; the number usually exposes the fiction. What good looks like: A staffed, skilled boundary is what makes human oversight more than a slogan. Watch volume growth; the boundary silently becomes a rubber stamp as throughput rises.
Humans really override
- 1Never overridden
- 2Rubber-stamped
- 3Rare overrides
- 4Real overrides
- 5Overrides tracked and used
At the low end: An override that never happens means the human side of the line is fictional. Measure how often reviewers actually change the AI's call; near zero tells you the truth. What good looks like: Tracked, meaningful overrides prove the boundary is alive, not ceremonial. Feed the override reasons back into where the line should sit next.
The line moves on evidence
- 1Frozen or drifting
- 2Moves by accident
- 3Moved once
- 4Reviewed periodically
- 5Evidence-based moves
At the low end: A frozen line ages into folklore while an accidental one moves without anyone deciding. Put one scheduled review on the calendar of whoever owns the riskiest line. What good looks like: Evidence-based boundary moves are how autonomy expands safely. Publish the moves and their evidence; it builds the trust that lets the next move happen.