Measuring Trust

Trust is invisible. You cannot see it, weigh it, or count it. Yet it is the most consequential property of every human-agent relationship. The measurement problem in AXD is this: how do you quantify something that exists only in the relationship between a human and an agent, that changes moment to moment, and that the human themselves may not be able to articulate?

Traditional approaches to trust measurement rely on surveys - asking humans to rate their trust on a scale. These approaches are inadequate for agentic systems for three reasons. First, they are retrospective: they measure how the human felt about trust at the moment of the survey, not how trust actually operated during the agent's autonomous actions. Second, they are declarative: they capture what the human says about trust, not what they do - and there is often a significant gap between declared trust and behavioural trust. Third, they are static: they capture a snapshot, not a trajectory.

AXD requires a different measurement paradigm - one built on behavioural indicators rather than declared attitudes. Trust is measured not by what the human says but by what the human does: how much they delegate, how often they intervene, how quickly they re-delegate after a failure, and how their delegation patterns evolve over time.

Delegation breadth. The range of domains in which the human delegates to the agent. A human who delegates only grocery shopping trusts the agent less than a human who delegates grocery shopping, travel booking, and financial management. Increasing delegation breadth over time is a signal of growing trust.

Delegation depth. The consequence level at which the human delegates. A human who allows the agent to spend up to £20 autonomously trusts less than one who allows £500. Increasing delegation depth - higher spending limits, more complex decisions, higher-stakes negotiations - signals deepening trust.

Intervention frequency. How often the human overrides, corrects, or checks the agent's decisions. High intervention frequency signals low trust - the human does not believe the agent can operate independently. Declining intervention frequency over time signals growing trust. A sudden spike in intervention frequency signals trust erosion.

Recovery speed. How quickly the human re-delegates after a failure. If the agent makes a mistake and the human immediately re-delegates (perhaps with tighter constraints), trust is resilient. If the human withdraws delegation for weeks or months, trust has been deeply damaged. Recovery speed is the most sensitive indicator of trust architecture quality.

Absence tolerance. How long the human is comfortable leaving the agent to operate without checking in. A human who checks the agent's activity every hour trusts less than one who checks weekly. Increasing absence tolerance is the ultimate trust indicator - it means the human trusts the agent enough to stop watching.

Beyond measuring trust level, AXD must measure trust calibration - the alignment between the human's trust and the agent's actual competence. Trust calibration has three states:

Well-calibrated trust: The human's trust matches the agent's competence. The human delegates appropriately - neither too much nor too little. This is the target state.

Over-trust: The human trusts the agent more than its competence warrants. This leads to over-delegation - the human grants authority the agent cannot handle responsibly. Over-trust is dangerous because it leads to larger failures, which lead to deeper trust collapse. Over-trust is often the result of performative trust signals - the agent appearing more competent than it is.

Under-trust: The human trusts the agent less than its competence warrants. This leads to under-delegation - the human constrains the agent unnecessarily, reducing the value of the relationship. Under-trust is wasteful but not dangerous. It is often the result of poor trust signal design - the agent failing to communicate its competence effectively.

Measuring trust calibration accuracy requires comparing the human's delegation patterns (a proxy for their trust level) against the agent's actual performance (a proxy for its competence). The gap between these two measures is the calibration error - and minimising this error is a primary objective of trust architecture.

A single trust measurement is nearly useless. Trust is meaningful only as a trajectory - a pattern of change over time. The AXD designer must track trust trajectories to understand whether a human-agent relationship is healthy, stagnating, or declining.

Healthy trajectories show gradual trust growth with occasional dips (from minor failures) followed by recovery. The overall trend is upward, with increasing delegation breadth and depth over time.

Stagnating trajectories show trust plateauing at a fixed level. The human delegates within a narrow scope and never expands. This often indicates a competence plateau - the agent has reached the limit of its demonstrated capability, and the human has no reason to trust it further.

Declining trajectories show trust gradually eroding over time. Delegation breadth narrows, intervention frequency increases, and absence tolerance decreases. This is the signature of trust erosion - and it requires immediate design intervention.

Collapsed trajectories show a sudden, dramatic drop in trust - typically following a catastrophic failure. The human withdraws most or all delegation and may abandon the agent entirely. Recovery from collapse requires the most intensive trust recovery protocols.

By monitoring trust trajectories in real time, the system can detect erosion before it becomes visible to the human and trigger proactive interventions - alignment rituals, performance reports, or constraint adjustments - that arrest the decline before it becomes irreversible.

The logical culmination of trust measurement is the trust dashboard - a system-level view of trust health across all human-agent relationships. For organisations deploying agentic systems at scale, the trust dashboard is the equivalent of a financial dashboard: it shows the health of the organisation's most valuable asset.

A trust dashboard should surface: aggregate trust trajectories across the user base, identifying systemic trust trends; individual trust anomalies - users whose trust is declining faster than the baseline; trust calibration accuracy - the proportion of users who are well-calibrated versus over-trusting or under-trusting; recovery metrics - the average time to trust recovery after failures; and erosion alerts - early warnings of the seven erosion patterns identified in AXD.

The trust dashboard is not a product feature - it is an organisational capability. It allows the organisation to manage trust as a strategic asset, investing in trust formation where it is weak, protecting trust where it is strong, and recovering trust where it has been damaged. In the agentic economy, the organisations that measure trust most accurately will be the organisations that earn the deepest delegation.