Practical guide
Verification and risk control: check before release
Use a fictional event budget to choose consequential checks, record corrections and decide what is ready to release.
Verification and risk control means testing material claims and calculations before somebody relies on them, then deciding what to correct, qualify or hold. This original event-budget exercise separates evidence gathering from checking an actual output. It focuses on consequences and observable checks, not a blanket claim that every error can be eliminated or that a checklist makes a decision safe.
Identify what the reader may act on
Prioritize statements that could change an approval, payment or operational commitment.
The fictional event proposal contains a total budget, a catering estimate and a statement that cancellation is free. The budget has an arithmetic error, the catering figure comes from last year's quote and the cancellation statement has no supporting passage. These are three different problems; a single instruction to fact-check the document does not specify how to resolve them.
List the actionable claims and the consequence of each being wrong. Keep presentation defects separate from substantive uncertainty. A misspelled venue name may need correction, but it should not distract the reviewer from an unsupported condition that could affect a financial commitment.
Choose an independent check for each failure
A useful check has a way to disagree with the draft rather than merely repeat it.
Recalculate the total from the line items without using the draft's subtotal as an input. Compare the catering assumption with the current quote or label it as awaiting confirmation. For cancellation, locate the applicable terms and ask the responsible qualified owner when interpretation is needed. Do not ask the same assistant to reassure you that its earlier assertion is correct.
Write the expected evidence for each check before performing it. This makes a missing quote visible as an unresolved input, rather than allowing a plausible number to pass because nobody specified what would count as support.
Record corrections and remaining uncertainty
Keep a compact check record that distinguishes performed work from an intended review.
For the exercise, record the original amount, recalculated amount, source used and location of the corrected figure. If the current quote is unavailable, mark that item unresolved. Do not mark the budget checked merely because a review task was created or a source file was opened.
An activity log may show access to a spreadsheet without showing that the arithmetic was independently tested. Look for the calculation or correction itself where available. If evidence of a check is missing, state that limitation. Avoid converting missing visibility into certainty that the person either did or did not perform the check.
Define release conditions
Some defects can be corrected immediately; others change whether the document is ready to support action.
In the fictional budget, the arithmetic can be repaired now. The outdated catering price may require a conditional estimate. The unsupported cancellation statement should be removed or resolved before it is used to justify a booking. The release decision depends on the action requested, not just the number of unresolved items.
Separate a discussion draft from an approval-ready proposal. A discussion draft can expose open questions; an approval request must make its dependencies unmistakable. Name who owns each unresolved item and what evidence would close it. Do not use a polished layout to imply that those conditions have already been satisfied.
Test whether the check process survives a change
Introduce one new fact and see whether the relevant checks are repeated.
Replace the fictional guest count after the first review. The reviewer should revisit quantity-dependent costs, totals and any assumptions tied to attendance. They do not need to repeat an unrelated spelling review merely to show activity. This tests whether the check process follows dependencies rather than a memorized sequence.
Finish with the corrected proposal and a short release note listing what changed, what remains conditional and who must act next. Interpret the performance as evidence from this bounded task. The exercise does not certify professional competence, establish compliance or predict outcomes in other settings.
Sources and scope
NIST addresses AI risk management. Skills England describes workplace AI foundations.
These sources provide background, not endorsement of this exercise. The worked example and suggested review method are original illustrative guidance. They are not customer results, validated benchmarks or evidence of a particular product capability. Adapt the exercise to the task and use qualified review where consequences require it.
Sources: [1] [2]
Sources
- 1.AI RMF Core · NIST
- 2.AI foundation skills for work benchmark · Skills England
Explore this topic
- How to develop and assess assumption testing in AI-enabled work
Assumption testing is the practice of turning an uncertain belief that supports a work product into a checkable statement, then seeking evidence that could confirm, narrow, or break it. In AI-enabled work, it prevents a plausible model completion from becoming an invisible premise. The goal is not to remove uncertainty, but to expose which beliefs matter and test them in proportion to consequence.
- How to develop and assess calculation checking in AI-enabled work
Calculation checking is the ability to reconstruct a numerical result from its inputs, units, rules, and transformations, then test whether the result is both computed correctly and fit for its intended use. In AI-enabled work, it prevents a plausible number or polished table from being accepted without a reproducible path. The evidence is the check itself, not confidence in the tool that.
- How to develop and assess citation verification in AI-enabled work
Citation verification is the ability to confirm that a cited item exists, is identified accurately, has the relevant version and status, and actually supports the nearby claim in context. In AI-enabled work, it prevents fabricated, misattributed, outdated, corrected, or merely adjacent sources from creating false authority. A working link is only the start of the check.
- How to develop and assess control escalation in AI-enabled work
Control escalation is the ability to stop, contain, and route an AI-assisted work item when the person handling it no longer has enough authority, evidence, or control to proceed safely. It is not a vague instruction to “ask someone senior.” A useful escalation identifies the trigger, preserves the work already done, names the decision that is needed, and gives the next owner a bounded set of.
- How to develop and assess fact checking in AI-enabled work
Fact checking is the disciplined testing of a consequential statement against the best available evidence for its exact meaning, time, scope, and intended use. In AI-enabled work, the object of review is not “the answer” as a whole. It is each checkable claim that a reader may rely on. This distinction prevents a common failure: a draft may cite real sources yet combine them into a false.
- How to assess hallucination detection in AI-enabled work
Hallucination detection finds generated content without evidence or at odds with task context. It flags unsupported claims for review before they enter important work products.
- How to assess privacy-aware handling in AI-enabled work
Privacy-aware handling limits personal or sensitive data to what a task needs. It uses approved tools and checks AI outputs for needless disclosure before release.
- How to assess uncertainty management in AI-enabled work
Uncertainty management makes limits visible before a reader acts. It separates known facts, bounded estimates, assumptions, and unresolved questions, then assigns each one an appropriate next step.