Autonomy Boundary Calibration Score (ABCS)

GMI-AR-006Asset ReliabilityResearch-Forward

Definition

Autonomy Boundary Calibration Score (ABCS) measures whether the division of decision authority between human and AI is set correctly: tracking both restrictive failures (the system couldn't act when it should have) and permissive failures (the system acted when it shouldn't have).

The Measurement Gap

A metals processing plant deployed an AI-based condition monitoring system on their rolling mill. The system could detect roll surface degradation, bearing wear, and alignment drift with roughly 80% accuracy: solid performance for a first-generation deployment. During commissioning, the plant faced a question that nobody had a clean answer for: how much decision authority should the system have?

The conservative approach won out. The AI could generate alerts and recommend actions, but every work order required planner approval before anything happened. Six months in, the pattern was uncomfortable. The planner was approving about 60% of recommendations and rejecting 40%. The approvals included both the legitimate findings and a portion of the false positives: the planner couldn't reliably tell the difference on every call. The rejections included real degradation events that just didn't look urgent enough at the time of review.

The deeper problem surfaced gradually. The AI was flagging early-stage degradation conditions with 48-hour response windows. Planner review was taking 72 hours on average. Some events slipped past the window. Failures occurred inside a window where faster response would have enabled action. The system was technically right. The process was too slow to act on it.

At a sister plant, the same system had been configured with broader autonomy: it could automatically schedule work orders for non-critical assets without human approval. It generated 340 work orders in a quarter. Post-analysis showed that 38% of those work orders were unnecessary. The system was acting on sensor noise that should have been filtered. Planners at that plant spent 15 hours per week cancelling machine-generated work that should never have been created. Technicians were showing up to jobs that didn't need doing. Trust in the system had collapsed.

Two plants, same system, opposite failure modes. The first plant's boundary was too tight: the AI couldn't act when it should have. The second plant's boundary was too loose: the AI acted far too often relative to its actual reliability level.

Neither plant had a metric for it. Nobody was measuring whether the line between human authority and machine authority was drawn in the right place. The only way they found out there was a problem was through accumulated cost overruns, planner frustration, and backlog growth. By that point, the program credibility damage was significant.

This is the autonomy boundary calibration problem. Every AI system that takes or recommends action operates within an authority boundary, a set of rules defining what the system can decide on its own versus what requires human approval. That boundary gets set during implementation, usually based on organizational risk tolerance, vendor guidance, and gut feel. It rarely gets revisited systematically. And there is no standard metric for evaluating whether the boundary is correctly placed relative to the system's actual current performance.

Too tight, and you lose the value the AI was supposed to deliver. Too loose, and you generate waste and erode trust faster than the system can recover. The question of where the boundary belongs is not a one-time implementation decision: it needs to be tracked continuously as AI accuracy changes, as asset populations evolve, and as operational context shifts. Without a metric, you find out the boundary is wrong the same way the two rolling mill plants did: after the damage is done.

Formula

Autonomy Boundary Calibration Score (ABCS) measures whether the division of decision authority between human and AI is set correctly: tracking both restrictive failures (the system couldn't act when it should have) and permissive failures (the system acted when it shouldn't have).

In plain language:

Is the line between "the machine decides" and "a human decides" drawn in the right place?

The formula:

ABCS = 1 − (Rᵣ + Rₚ) / 2

Where:

Rᵣ = restrictive boundary failures / total events where AI lacked authority

Rₚ = permissive boundary failures / total autonomous actions taken

Rₚ captures every autonomous action the AI took that post-hoc analysis determined was unnecessary or harmful. "Unnecessary" means analysis showed no condition warranting intervention existed at the time the action was triggered. This requires a review process, not just system logs: someone with maintenance engineering knowledge has to evaluate whether each flagged action was warranted.

Worked example: the two rolling mill plants:

Restricted system (Plant A):

Over the evaluation period, the plant tracked every instance where the AI flagged a condition but lacked authority to act without planner approval. Result: 82 events where the AI identified a need but lacked authority. Of those, 14 resulted in worse outcomes because human delay exceeded the response window.

Rᵣ = 14/82 = 0.171

The system took no autonomous actions, so Rₚ = 0: you cannot have permissive failures when the system has no autonomy.

ABCS = 1 − (0.171 + 0) / 2 = 0.914

That score looks good. Read the breakdown. All miscalibration is restrictive. The zero permissive failure rate is not a sign of good governance: it is a sign of zero autonomy. Plant A's ABCS directs you to expand machine authority, not celebrate.

Permissive system (Plant B):

Same evaluation period, broad autonomy configuration. Only 12 events where the AI lacked authority (the system was authorized to act on almost everything). Of those 12, 2 resulted in worse outcomes from the brief approval delay.

Rᵣ = 2/12 = 0.167

The system took 340 autonomous actions. Of those, 129 were determined unnecessary or harmful in post-hoc review.

Rₚ = 129/340 = 0.379

ABCS = 1 − (0.167 + 0.379) / 2 = 0.727

Plant B's 0.727 reflects a system that has too much authority relative to its accuracy level. The directional guidance is clear: constrain Plant B's autonomous scope (reduce Rₚ) while expanding Plant A's (reduce Rᵣ). Both plants need adjustment in opposite directions, and the Rᵣ/Rₚ breakdown tells you which direction for each.

Edge cases to handle before you compute:

The AI had authority but chose not to act. This is a model quality issue (PCCI or MDDR), not a boundary calibration issue. Exclude it from ABCS. ABCS answers whether the boundary rules are correctly set, not whether the model's decisions within those rules are accurate.

New boundary settings require a 90-day stabilization period before computing ABCS on the revised configuration. Operators and the system both need time to adapt after a boundary change. Computing ABCS in the first 30 days produces noise, not signal.

Sign convention:

Higher is better. Both failure rates drive the score down. The composite score alone doesn't tell you which direction to adjust: always read the Rᵣ/Rₚ breakdown alongside the ABCS value.

Proposed formula, pending validation pilots.

At a Glance

Category
Autonomous Action and Strategy Integrity

Decision Supported

Should I expand or contract the AI system's decision authority for this asset class or decision type?

← Prediction Confidence Calibration Index (PCCI-AR)Degradation Projection Validity Rate (DGPVR) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment