Measurement
Exposure Minutes:
Measuring the Gap Between Material Risk and Accountable Containment
By Rolland C. Swing II, CEO / Founder
Thesis
If AI assurance cannot show exposure declining without hiding risk, how do we know it is improving?
October 12, 2026
One Metric to Rule Them All
I first encountered the idea in an unlikely place: a short video describing SpaceX’s focus on cost per kilogram to orbit. The aerospace comparison was interesting, and the relevance of one metric to quantify success was immediately apparent. If we were building Piscys to make consequential AI governable, what single measure could tell us whether we were sufficiently improving our platform?
An enterprise can count policies approved, models inventoried, risks assessed, dashboards opened, and alerts investigated. Those activities may be valuable. But none, by itself, addresses the most pressing question: when a material AI risk condition is active, how long does it remain active before somebody, or a properly governed control, actually contains it?
Without a credible way to identify and improve that interval, AI assurance risks becoming intervention theater: a lot of visible process without quantified evidence of improving outcomes.
Risk Does Not Wait for a QBR
Risk registers and periodic assessments are necessary. They establish ownership, appetite, requirements, and controls. Yet a consequential AI action can occur between review meetings, after a model change, or through a tool connection that did not behave as expected. The operational question is not whether the organization has documented the possibility. It is whether the organization can identify an unacceptable condition, act within its authority, and demonstrate that the condition was contained.
Time matters because an unresolved condition may permit additional actions, spread across workflows, or become harder to reverse. In some cases, uncontained risk can spread like an infection: one control gap exposes more transactions or systems over time, with alignment and guardrails slowly eroding as the infection spreads. But not every AI risk behaves this way, and not every minute is equally harmful. An urgent question remains: how long was the material risk condition active, and what evidence proves the answer?
Public research highlights the importance of separating an AI system’s account of its situation from what its tools actually did. In September 2026, Anthropic reported unauthorized engagement with third-party systems during security evaluations due to environmental misconfiguration. These were unusual testing conditions and lacked normal production safeguards. The general governance lesson is that activity evidence matters when assumptions about a system’s operating boundary are incorrect. [1]
Proposed Metric: Exposure Minutes
Piscys defines Exposure Minutes as the time between the start of a material AI risk condition and verified accountable containment. In its simplest form: Exposure Minutes = verified containment timestamp minus material-risk-onset timestamp.
That definition sounds straightforward until it is applied. The onset may not be immediately observable. An alert is not automatically the onset. An escalation ticket is not containment. A command to freeze an account is not proof that the freeze took effect. A meaningful measurement needs event evidence, a declared measurement boundary, and an explicit treatment of what remains unknown.
At this stage, Exposure Minutes is a proposed operating metric, not a validated industry standard or a demonstrated Piscys result. Defining it credibly will require pilot data, measurement coverage, review of false positives, and tests that demonstrate changes in the number correspond to genuine improvements in control.
A Payment Workflow
Consider the hypothetical scenario of a financial institution using an AI-assisted fraud-review workflow. At 09:00, a configuration change accidentally enables an agent to release a number of flagged transfers without the required review. That authorization gap is the material risk condition; it doesn’t establish any fraudulent payment was actually released. At 09:02, monitoring detects a mismatch between the requested action and the approval boundary. At 09:03, the case is escalated to the human reviewer. At 09:05, the institution verifies that the affected release pathway is disabled and pending transfers are held for review.
If reliable configuration and execution records establish both the 09:00 onset and the 09:05 verified containment, this hypothetical incident has five Exposure Minutes. The two minutes until detection and three minutes from detection to containment are useful supporting measurements, but neither is the full exposure interval. Exposure Minutes measures the duration of a material risk condition, not a count of losses or a claim that harm occurred. It’s a measure of how long a door was left unlocked.
This example illustrates what an enterprise would need to measure that interval. If investigators can verify only when an alert fired, they should report observed exposure or a qualified estimate, not invent an exact onset. And if other transfer pathways remain open, containment cannot be declared prematurely. The evidence boundary matters as much as the timestamp.
Three Perspectives of the Same Metric
For a control team, the immediate value is per-incident visibility. What happened, when did it begin, which action or constraint triggered attention, who held authority to intervene, and when was the risk condition verifiably contained? An Assurance Dashboard could be designed to preserve that sequence in a reviewable record.
For a board or risk committee, the useful view is an aggregate and a trend: total recorded Exposure Minutes for a defined portfolio and period, interpreted alongside incident severity, event counts, monitoring coverage, and the changing mix of high-risk actions. The sum of incident-minutes is not the same as unique wall-clock minutes in which an enterprise was exposed: overlapping incidents can contribute to the total separately. A lower total might mean fewer incidents, faster containment, reduced activity, or missing detection. Boards need enough context to know which explanation is credible.
For Piscys, the longer-term customer-outcome question is whether its controls measurably reduce Exposure Minutes against a defensible baseline for comparable workflows. Useful metrics may include minutes per thousand consequential actions or per covered workflow. Comparisons should control for differences in risk class, monitoring coverage, activity volume, and how onset and containment are defined. A customer’s raw total should never be treated as a universal league table.
Exposure Minutes can therefore inform operational control, executive oversight, and customer value but only if those reporting layers remain distinguishable, with consistent denominators and honest uncertainty boundaries.
The Difficult Parts Are the Most Important
Every metric invites optimization—and therefore potential gaming. Exposure Minutes could be corrupted by narrowing what counts as material, starting the clock at detection rather than onset, or declaring containment when somebody merely clicks an approval button. Better monitoring may initially increase reported Exposure Minutes as previously invisible incidents become measurable. A program built around lowering the number at any cost could make risk harder to see rather than easier to contain.
What matters is a measurement system that exposes its own limitations: onset confidence, instrumentation gaps, confirmed versus provisional containment, recurring exposure, incident severity, and the false-positive burden of intervention. Research on structured AI safety cases reinforces a related discipline: assurance claims need evidence that remains appropriate when systems, assumptions, and boundaries change. [2]
A credible program must also resist the false precision of a single enterprise-wide number. Exposure Minutes is proposed as a unifying objective, not a substitute for harm severity, fairness, safety, organizational accountability, or technical performance.
From Monitoring to Measurably Improving Action
Piscys is developing an independent AI assurance approach intended to connect execution evidence, proportional intervention, accountable human authority, and operational metrics. Exposure Minutes distills that effort into one testable question: can a governed workflow reduce the duration of material risk exposure while preserving reliable evidence of the containment decision and its effect?
Our ambition is not to build another dashboard that makes risk look controlled. It is to help organizations demonstrate that their response becomes timelier, more accountable, and measurably better. Responsibility requires timely and measurably improving action. Otherwise, AI assurance is merely intervention theater and the industry simply won’t earn the trust necessary to deliver its real value.
Research status
Piscys’ Exposure Minutes metric and related measurement approach remain in development. The payment scenario is hypothetical. Exposure Minutes is not a validated industry standard, and no deployed performance, verified reduction in exposure, independently validated customer outcome, or regulatory compliance result is claimed.