KPI 06 of 07 · Oversight Phase · Consumer-side · Optimal range metric
Intervention Frequency Ratio
The percentage of agent actions requiring human intervention, correction, or override
Abbreviation: IFR · Optimal range: 5-25%IFR is the oversight calibration metric of Agentic Experience Design. It measures the balance between human control and agent autonomy - the fundamental tension at the heart of every agentic system. Too much intervention means the human has not actually delegated. Too little intervention may mean the human has abdicated responsibility for outcomes they should be monitoring.
Unlike the other metrics in the AXD Metrics Standard, IFR does not have a simple directional target. It has an optimal range that varies by domain, stakes, agent maturity, and relationship history. A financial agent managing high-value investments should have a higher IFR than a shopping agent purchasing routine household items. The design challenge is not to minimise intervention but to calibrate it appropriately.
IFR captures the quality of the interrupt design - the system of triggers, notifications, and escalation pathways that determine when the human is brought back into the loop. Well-designed interrupt systems produce calibrated IFR: the human intervenes on decisions that warrant human judgment and allows routine operations to proceed autonomously. Poorly designed interrupt systems produce either excessive intervention (notification overload) or insufficient intervention (missing triggers).
The metric draws from the Trust Architecture pillar and the Interrupt Design Framework. It operationalises the founding principle that absence is the primary use state of agentic systems - and that the design of re-engagement is as important as the design of delegation.
Interrupt design and oversight architecture
IFR is governed by the interrupt design of the agentic system. Well-designed interrupts bring the human back into the loop at the right moments - when stakes exceed the delegation threshold, when the agent encounters uncertainty beyond its confidence bounds, or when the situation has changed in ways the original delegation did not anticipate.
Commerce protocols influence IFR by providing structured transaction data that enables better interrupt design. When an agent can access real-time pricing, inventory, and product specifications through ACP or UCP, it can make more confident decisions autonomously - reducing unnecessary interrupts while maintaining appropriate escalation for genuine edge cases.
Numerator
Agent actions requiring human intervention (correction, override, approval)
Denominator
Total agent actions in the measurement period
× 100 = IFR %
Optimal range metric. Target 5-25% for most agentic commerce scenarios. Report alongside domain stakes and agent maturity context.
Measurement protocol
Log every agent action and classify it as autonomous (no human involvement) or intervened (human corrected, overrode, approved, or modified the action). Calculate the ratio of intervened actions to total actions. Segment by action type, domain, and stakes level.
Distinguish between designed interventions (the system correctly triggered human re-engagement) and undesigned interventions (the human proactively checked and corrected). Designed interventions indicate healthy interrupt architecture. Undesigned interventions indicate missing triggers.
Track IFR trends over time alongside TEI. A healthy pattern shows IFR gradually declining as trust builds and the agent demonstrates competence. An unhealthy pattern shows IFR declining while TEI rises - indicating the human has stopped monitoring rather than genuinely trusting the agent.
Four levels of oversight calibration
>50%
Human is intervening in more than half of all agent actions. The agent is not operating autonomously - it is operating as a tool that requires constant supervision. Delegation design has failed; the human has not actually delegated.
25-50%
Frequent intervention. The human is checking and correcting a significant portion of agent actions. Trust is insufficient for autonomous operation. The agent may be capable but the human has not calibrated trust appropriately.
5-25%
Intervention is focused on high-stakes decisions and edge cases. The human intervenes when appropriate but allows routine operations to proceed autonomously. This is the target range for most agentic systems.
<5%
Near-full autonomy. The human rarely intervenes. This is appropriate for mature, well-calibrated relationships with proven agents in low-stakes domains. In high-stakes domains, this may indicate under-monitoring.
What moves IFR up, down, and sideways
Well-designed human re-engagement triggers, clear escalation criteria, progressive autonomy that expands with demonstrated competence, domain-appropriate intervention thresholds, transparent agent confidence signals.
IFR has a sweet spot, not a simple direction. Too high means the human hasn't delegated. Too low may mean the human is under-monitoring. The optimal IFR depends on domain stakes, agent maturity, and relationship history. Report IFR alongside domain context.
Missing human re-engagement triggers (IFR artificially low), excessive notification fatigue causing intervention avoidance (IFR artificially low), anxiety-driven over-monitoring (IFR artificially high), no progressive autonomy design (IFR stuck at initial level).
Why IFR matters commercially
IFR determines the operational efficiency of agentic commerce. An agent system with excessive IFR is not saving the human time - it is creating a new form of work (monitoring and correcting the agent). The commercial promise of agentic commerce - that agents handle routine tasks so humans can focus on higher-value activities - depends on achieving calibrated IFR.
For agent platform providers, IFR is a competitive differentiator. Platforms that achieve calibrated IFR - appropriate intervention without excessive monitoring burden - will retain users. Platforms with excessive IFR will lose users who find the monitoring burden outweighs the delegation benefit.