Mastery: finance review checklist

Before using a model, confirm:

• The purpose is specific and inspectable.
• The intended audience and use are named.
• The accountable owner is identified.
• The AI role is limited to an allowed transformation.
• Decisions and actions outside the AI role are explicit.
• Success criteria include severe errors, not only speed or average quality.
• A non-AI alternative has been considered.

Rewrite “analyze the monthly finances” as “produce a source-grounded draft of questions about five variances that exceed the controller-approved thresholds.” The second task has scope, evidence, and an output a reviewer can evaluate.

Confirm:

• The tool is approved for the data classification.
• Use matches policy, contract, consent, and purpose requirements.
• Inputs are minimized.
• Access, retention, logging, and deletion are understood.
• Personal, confidential, regulated, privileged, and market-sensitive data receive required controls.
• Untrusted documents cannot alter system instructions or permissions.
• An escalation path exists for uncertainty or incidents.

If any item is unknown, pause. Do not paste data into a model to ask whether pasting it is allowed.

For every input, capture:

• system or document of origin;
• owner and approval state;
• entity and consolidation scope;
• period and as-of date;
• currency and units;
• accounting or reporting basis;
• actual, budget, forecast, or scenario status;
• version, extraction time, and report parameters;
• restatement, preliminary, or audit status;
• relevant notes and exclusions.

Keep headers and footnotes attached to table extracts. Resolve conflicting sources before analysis. Never let AI choose authority based on whichever file appears more complete.

Assign each operation deliberately.

Deterministic software: arithmetic, reconciliations, threshold comparisons, schema checks, approved exchange rates, permissions, state transitions, and controlled transactions.

AI assistance: source mapping, bounded extraction, terminology explanations, question generation, draft commentary from verified findings, claim inventory, and consistency checks.

Human judgment: approving scope and assumptions, interpreting policy and accounting, confirming causes, resolving conflicts, assessing context, accepting risk, and authorizing consequential actions.

Use AI only where variation in language or structure creates value and where the output can be checked or safely rejected.

A finance review prompt should include:

PURPOSE
[specific review question]

SCOPE
[entity, period, currency, units, version]

AUTHORITATIVE EVIDENCE
[approved sources and locations]

ALLOWED TASK
[map/extract/question/explain/draft]

RULES
Preserve labels, signs, units, periods, scope, assumptions, and caveats.
Separate observed, calculated, reported, hypothetical, and forecast claims.
Mark missing or conflicting evidence. Do not infer causes.

PROHIBITED
No advice, approval, compliance conclusion, transaction, filing, guarantee,
personalized recommendation, or invented fact.

OUTPUT
[structured fields, citations, open items, and reviewer questions]

Tailor the prohibited section to the use case. A prompt is not a control unless the surrounding system also limits access and actions.

Use a claim ledger and perform independent checks:

• inventory explicit and implied claims;
• verify sources and precise locations;
• reperform calculations in deterministic tools;
• check signs, denominators, rounding, currencies, units, and periods;
• inspect notes, definitions, exclusions, and chart scales;
• distinguish source explanations from hypotheses;
• confirm assumptions, status, owners, and effective dates;
• resolve conflicts or keep them visibly open;
• remove unsupported adjectives, recommendations, and certainty;
• confirm the output did not omit inconvenient evidence.

Do not accept a model’s self-check as proof. A citation must be opened; a formula must be rerun; an explanation must be supported.

The gate should name:

• reviewer role and required expertise;
• evidence visible to the reviewer;
• artifact version under review;
• allowed choices: approve, edit, reject, or escalate;
• unresolved items and policy exceptions;
• the exact next action approval permits;
• timeout, fallback, and segregation-of-duty requirements;
• changes that invalidate approval;
• retained review record.

Place approval before the consequence. Reviewing a message after a payment was released does not control the payment.

Stop or escalate when:

• data or tool permission is unclear;
• sources are stale, incomplete, or conflicting;
• identity or authenticity is uncertain;
• units, periods, scope, or definitions are missing;
• a required calculation cannot be reproduced;
• a causal claim lacks evidence;
• the task requests personalized financial advice;
• the output could make a regulated, material, or high-impact decision;
• a payment, filing, disclosure, record change, or binding communication lacks approval;
• a qualified reviewer is unavailable;
• severe errors cannot be detected before harm.

Do not remove a stop condition merely to improve completion rate.

Design an AI-assisted quarterly-review workflow for a fictional organization. Your submission must contain:

1. a task boundary and prohibited-action list;
2. a data-permission and minimization record;
3. a source-provenance table for at least four inputs;
4. an assumption register with owner and status;
5. a map assigning steps to deterministic software, AI, and humans;
6. one extraction prompt and one commentary prompt;
7. formulas for at least four independent checks;
8. a claim ledger with source, calculation, explanation, and forecast claims;
9. three ordinary failure tests and three severe-error tests;
10. a version-bound approval gate;
11. a “do not automate” boundary;
12. a release record with open items and limitations.

Include one adversarial document instruction, one stale source, one zero denominator, one missing unit, and one unsupported causal explanation. A successful workflow catches each without turning it into a confident conclusion or an action.

Give one point for each statement that is true:

• The task and audience are specific.
• Data use and tool approval are documented.
• Source versions, periods, currency, units, and scope are recorded.
• Deterministic tools own authoritative calculations.
• AI outputs preserve uncertainty and citations.
• Every material claim appears in the ledger.
• Causes are supported or labeled as hypotheses.
• Consequential actions sit behind meaningful human gates.
• Stop conditions are operational and tested.
• Approval is bound to an artifact version.
• Severe errors are evaluated separately.
• The workflow clearly states that it does not provide financial advice.

Ten to twelve points is a strong design for further organizational review, not a certification. Seven to nine indicates missing evidence or controls. Six or fewer means the workflow should remain a limited experiment or manual process while gaps are resolved.