Practical guide

AI executive business review verification checklist

Use this checklist after an executive business review has been drafted but before it is issued. It tests whether the review can support a real decision: whether its scope is clear, its figures can be reproduced, adverse evidence is visible, alternatives are credible, and approved actions can be traced. A polished deck can fail every one of those tests.

By Two Prune

Use this checklist after an executive business review has been drafted but before it is issued. It tests whether the review can support a real decision: whether its scope is clear, its figures can be reproduced, adverse evidence is visible, alternatives are credible, and approved actions can be traced. A polished deck can fail every one of those tests.

What this checklist does and does not verify

The UK Government Project Delivery Functional Standard connects governance, roles, planning, benefits, risk, issues, reporting, assurance, and change control. The Government Analysis Functional Standard requires proportionate, quality-assured analysis that supports informed decisions, while the Gate 2 review guidance asks for evidence behind the business case, benefits, controls, risks, and recorded decisions. These are authoritative public-sector sources for disciplined review. They do not prescribe one universal executive-review format or prove that any particular business result is accurate. UK Government: Project Delivery Functional Standard UK Government: Analysis Functional Standard UK Government: Gate 2 Review

Sources: [1] [2] [3]

The TRACE verification method

Record pass, fail, not applicable, or not evidenced for every item. A pass requires an evidence reference, not a reviewer’s confidence. Assign each failure an owner and due date. Do not average away a critical failure.

### T — Terms, time, and reporting population

Stop condition: hold the review if the reporting population cannot be reconciled or a headline comparison uses incompatible definitions without a visible bridge.

### R — Recalculation and reconciliation

Stop condition: hold any decision that depends on a calculation the reviewer cannot reproduce or a source version that cannot be identified.

### A — Adverse evidence and alternative explanations

Stop condition: return the review when contrary evidence is omitted, a causal claim rests only on before-and-after movement, or uncertainty is converted into certainty by wording.

### C — Choices, consequences, and control

Stop condition: do not present a recommendation as a choice when alternatives are artificial or the person asked to accept risk lacks that authority.

### E — Execution and evidence retention

Stop condition: an action is not closed because someone reports progress; require the stated acceptance evidence.

  • [ ] The review names the business unit, programme, product, customer group, geography, and period in scope.
  • [ ] Every headline measure has a definition, unit, owner, extraction date, source system, and evidence ID.
  • [ ] Numerators and denominators refer to the same eligible population and period.
  • [ ] Currency, tax treatment, exchange rate, time zone, and rounding rules are stated where material.
  • [ ] Changes in definitions, instrumentation, organizational boundaries, or coverage are marked on the affected trend.
  • [ ] Missing, late, estimated, or manually adjusted records are quantified rather than silently excluded.
  • [ ] A reviewer can reproduce each headline total, rate, variance, and forecast from retained inputs.
  • [ ] The review reconciles to the approved budget, plan, benefits register, risk register, incident log, and prior action log where applicable.
  • [ ] Material spreadsheet formulas, joins, filters, allocations, and AI-assisted calculations have been independently checked.
  • [ ] Charts agree with tables, narrative, axis scales, labels, and source extracts.
  • [ ] Forecasts separate actuals, committed values, assumptions, and scenarios.
  • [ ] Each benefit states whether it is forecast, observed, validated, displaced, delayed, or no longer expected.
  • [ ] Material incidents, complaints, control exceptions, rework, delivery debt, and user harm are visible on the decision page.
  • [ ] The narrative distinguishes observed facts from interpretation and recommendation.
  • [ ] At least one credible competing explanation is tested for each material movement.
  • [ ] Activity and output measures are not described as outcomes without supporting evidence.
  • [ ] Sampling coverage, response bias, missingness, and non-comparable periods are disclosed.
  • [ ] Positive and negative findings use the same materiality rule.
  • [ ] Each decision request names the accountable decision maker and deadline.
  • [ ] The current course is compared with at least one credible alternative, including delay, restriction, repair, or stop where relevant.
  • [ ] Options show cost, benefit, risk, dependency, affected parties, reversibility, and evidence still needed.
  • [ ] Risk statements include cause, event, consequence, owner, response, and escalation trigger.
  • [ ] Proposed controls have an operating owner, test method, evidence record, and exception route.
  • [ ] Residual uncertainty and any decision conditions are explicit.
  • [ ] Open actions from the prior review retain their original due dates, age, owners, and acceptance evidence.
  • [ ] New actions specify deliverable, owner, due date, acceptance test, dependency, and escalation route.
  • [ ] Meeting decisions, dissent, conditions, and changes to plan or risk are recorded.
  • [ ] The issued review is versioned and linked to the exact source packet, calculation file, and decision note.
  • [ ] Access, retention, and sensitive-data handling match the organization’s rules.
  • [ ] The next review date and evidence required for the next decision are set.

Verification record

| Test area | Result | Evidence IDs | Failure or uncertainty | Owner | Due date | | --- | --- | --- | --- | --- | --- | | terms and population | | | | | | | recalculation | | | | | | | adverse evidence | | | | | | | choices and controls | | | | | | | execution record | | | | | |

Overall disposition: issue, issue with visible limitations, return for repair, or hold the dependent decision. The disposition applies only to this version and evidence packet.

Worked check: synthetic service-operations review

A quarterly review recommends expanding an AI-assisted case-summary workflow. It reports a 16% fall in average handling time, stable quality, and no material incidents. A reviewer applies TRACE.

The terms check finds that the timing measure covers 2,240 assisted cases, while the quality result covers a sample of 70 cases from both assisted and unassisted queues. The prior quarter used a different five-item quality checklist. The quality trend is therefore marked not evidenced, not stable.

The reviewer recalculates handling time and reproduces the 16% difference, but the assisted queue contains fewer escalated cases. The deck’s “productivity gain” is replaced with a descriptive timing difference. The source packet also contains seven unsupported summaries caught before issue, two missing attachments, and one privacy exception. They move to the decision page.

The options check finds only “expand now” and “expand later.” The team adds a bounded option: retain the current population for six weeks while capturing case complexity, running a bridged quality sample, reconciling attachments, and retesting the privacy control. Expansion stops if a further sensitive-data incident occurs.

The final disposition is issue with visible limitations for the current bounded decision. The checklist does not approve expansion. It records owners, due dates, exact evidence needed, and the next decision gate.

Using AI during verification

An approved AI tool may compare narrative claims with tables, enumerate missing evidence IDs, or flag inconsistent terminology. It must not mark its own output as verified, decide materiality, accept residual risk, or close exceptions. Open decisive sources and reperform calculations outside the generated summary.

Two Prune provides realistic AI-enabled work assessments and evidence about observed participant performance. It can sample whether someone verifies an executive review using supplied records. It does not audit an organization, validate business performance, approve a decision, certify controls, or make employment decisions.

Sources

UK Government: Project Delivery Functional Standard

  • UK Government: Analysis Functional Standard
  • UK Government: Gate 2 Review

Sources: [1] [2] [3]