The Four Risk Thresholds That Trigger Human Review

Deciding escalation by policy, not by mood

Some outputs must never go out as a Claude draft alone, no matter how good they look. This is the operational heart of Diligence, one of the two competencies from Anthropic’s AI Fluency Framework that anchor output evaluation: where Discernment is the skill of critically evaluating output against requirements, sources, and standards, Diligence is deciding when verification is required and taking responsibility for the result. The Claude Certified Associate - Foundations (CCAO-F) exam frames that Diligence as knowing which outputs must never ship unreviewed in advance, so that the decision to escalate is made by policy rather than in the moment after something has already gone wrong. The four risk thresholds are that policy: four questions you ask of every output to decide whether it needs a human before it ships.

The reason there are four, and not one, is that risk arrives from different directions. An output can be low-stakes but irreversible, or internal but regulated. A single question would miss whichever dimension it does not cover. Asking all four catches the risk wherever it comes from.

Stakes

Stakes asks the cost of being wrong. When an error would be expensive, damaging, or hard to recover from, the output demands human review regardless of how confident it appears. This is the threshold most people already feel, but the exam's emphasis is on the "regardless of how confident" clause: a high-stakes output that reads flawlessly still crosses the stakes threshold, because the appearance of correctness is not correctness. Stakes are judged by consequence, not by how the output looks.

Reversibility

Reversibility asks whether the action can be undone. A draft you can revise before it matters clears a low bar; a step that cannot be taken back, a sent client deliverable, a filed report, a public statement, clears a much higher one. This is a distinct question from stakes, because an action can be moderate in cost yet irreversible, and irreversibility removes the safety net that would otherwise let a mistake be caught and corrected downstream. When there is no undo, the review has to happen before the action, not after.

Audience

Audience asks who sees the output. An internal working draft is the low end; external, executive, and regulatory audiences raise the requirement sharply. The wider or more consequential the readership, the more a mistake costs in trust and the less room there is to quietly fix it. Importantly, an internal audience lowers this threshold but does not switch off the others: internal work can still be high-stakes, irreversible, or regulated, and any of those can require review on its own.

Regulatory exposure

Regulatory exposure asks whether a law, rule, or contract governs the content. Regulated material carries obligations that AI assistance does not remove; using Claude to draft something does not transfer or dissolve the compliance duties attached to it. When a rule governs the output, the regulatory threshold is crossed by that fact alone, independent of stakes or audience, because the obligation exists regardless of how the work was produced.

What the CCAO-F exam trips candidates on

The first trap is judging only stakes and ignoring that the same output might also be irreversible once sent. A scenario will foreground the cost of an error and hope you stop there, missing that the action cannot be undone. The credited answer runs all four thresholds, catching reversibility as its own trigger.

The second trap is assuming an output aimed at an internal audience never needs escalation, regardless of stakes. Internal lowers the audience threshold but does not clear the others. The exam rewards recognising that an internal but high-stakes, irreversible, or regulated output still crosses a threshold and still needs review.

Common misreadings to avoid

How this shows up on the exam

Domain 2 questions on this knowledge point describe an output and ask whether it needs human review. The dependable approach is to run all four thresholds independently, stakes, reversibility, audience, regulatory exposure, and escalate if any single one is crossed, treating a confident or internal-looking output as no exemption.

These thresholds are the general form of the stakes calibration idea, and they underpin the fixed do-not-ship-without-review categories, the iteration versus escalation signal, and their combined application in applying the escalation thresholds to a scenario. They also connect to the accountability ownership principle, which is why crossing any threshold matters.