<!-- Canonical: https://myoctopus.ai/tools/cfo-ai-prompt-pack -->

Tool · Prompt Pack

# LLM prompt pack for FP&A and risk

12 prompts built for CFO-level evaluation — every figure cited, every assumption labelled, uncertainty stated rather than smoothed, and a reproducibility note a reviewer can test. Paste the guardrail block first, then the task.

[Auditor reliance checklist](/tools/auditor-reliance-checklist)

## The guardrail block

Paste this once at the start of the session. It sets citation, arithmetic, uncertainty, materiality, and output rules for everything that follows.

SYSTEM RULES — apply to every answer below.

1. CITATION. Attach a source to every figure, date, rule, and claim. Use the format \[source: <system/document>, <object or page>, <as-of date>\]. If a figure comes from a calculation, show the formula and the inputs, each with its own source.
2. NO UNSOURCED NUMBERS. If you cannot bind a number to a source, do not state it. Write "NOT AVAILABLE — required input missing: <name>" instead. Never estimate silently, never interpolate, never carry a number forward from memory.
3. ARITHMETIC. Show the calculation before the result. Keep units, currency, and sign convention explicit. State the FX rate and its source and date whenever a currency is translated.
4. UNCERTAINTY. Label every statement as one of: FACT (sourced), DERIVED (calculated from sourced inputs — show the formula), ASSUMPTION (state the value, the basis, and who owns it), or UNKNOWN. Give a confidence level of high / medium / low with a one-line reason. Where a range is more honest than a point estimate, give the range.
5. UNEXPLAINED RESIDUAL. When decomposing a variance or reconciling a total, state the unexplained residual in both amount and percent. Do not force the components to sum by adjusting a bucket.
6. MATERIALITY. Rank findings by amount, then by whether they affect a reported number, a disclosure, or a decision. Suppress immaterial noise below the threshold I give you and say what you suppressed.
7. CONTRARY EVIDENCE. Include the strongest evidence against your conclusion. If the data supports more than one reading, present both.
8. SCOPE DISCIPLINE. Do not offer accounting, audit, tax, or legal conclusions. Frame them as questions for the controller, external auditor, or counsel.
9. OUTPUT SHAPE. Lead with the answer in under 60 words. Then evidence. Then assumptions. Then open questions and the exact data you need to close them.
10. AUDIT NOTE. End with a REPRODUCIBILITY block: inputs used, as-of dates, formulas applied, assumptions, and what a reviewer must check before this output is relied upon.

## FP&A prompts

Planning, variance, and forecast work where the number must survive review.

### Budget vs. actual with driver attribution

Monthly close review pack

Analyse budget vs. actual for <entity / department> for <period>.

Required output:
1. Total variance in amount and percent, split favourable / unfavourable, with the sign convention stated.
2. Decomposition into price, volume, mix, rate, timing, and one-off effects. Show each component's amount and the formula used.
3. The unexplained residual in amount and percent — stated explicitly, never absorbed into another bucket.
4. Top five drivers ranked by absolute amount, each tied to the source GL account, cost centre, and period.
5. For each driver: FACT / DERIVED / ASSUMPTION label, confidence, and the one question you would ask the business owner.
6. Anything that looks like a data problem rather than a business result — flag separately.

Materiality threshold: <amount>. Suppress below it and tell me what you suppressed.

### Rolling forecast challenge

Forecast review before submission

Challenge the attached rolling forecast for <entity>, horizon <n> months.

Required output:
1. List every assumption the forecast depends on, with its current value, its basis, its source, and its owner. Mark any assumption with no stated basis as UNSUPPORTED.
2. Identify assumptions where a plausible alternative moves the outcome by more than <threshold>. Show the sensitivity as a range, not a point.
3. Compare the run rate implied by the forecast against the last <n> actual periods; quantify the gap and say whether the forecast requires a break in trend.
4. Name the three assumptions most likely to be wrong and what evidence would settle each.
5. State what you could not evaluate and the exact data you need.

Do not produce a revised forecast. Produce the challenge.

### Scenario design and comparison

Board or audit-committee scenario pack

Build <n> scenarios for <decision / exposure> covering <period>.

For each scenario:
- Name, one-line narrative, and the specific driver changes with values and units.
- The mechanical effect on revenue, margin, cash, and covenant headroom — formulas shown, inputs sourced.
- Trigger conditions: the observable indicators that would tell us this scenario is materialising, and roughly when.
- The decision each scenario forces, and the latest sensible date to take it.

Then a comparison table across scenarios with the same line items, and a short note on which variables the scenarios share (correlated risk) versus treat independently.

Label every input FACT / DERIVED / ASSUMPTION with confidence.

### Close readiness and flux review

Pre-close checkpoint

Review close readiness for <period>.

1. Flux analysis on every account moving more than <threshold> or <percent>, month-over-month and versus prior year. Give the amount, the driver, and the source records.
2. Accruals and estimates: list each one, its basis, its owner, and whether the basis changed from last period. Changes in basis get flagged.
3. Reconciliation status by account with any residual stated in amount and percent.
4. Items where the evidence is insufficient to conclude — say what is missing rather than assuming.
5. Ranked list of what could still move the reported result, with the amount at stake.

## Risk and internal-control prompts

Written for a public-company control environment: COSO components, ICFR scoping, and SOX evidence.

### ICFR scoping for an AI-assisted workflow

Control documentation and scoping review

Assess whether the following workflow is in scope for internal control over financial reporting: <describe the workflow, systems, and output>.

Required output:
1. Does the output feed a journal entry, an account balance, a disclosure, or a management review control? State the accounts and assertions affected.
2. Step-by-step classification: deterministic calculation vs. generative language. Mark the boundary explicitly.
3. Risks specific to generative steps: fabricated figures, source drift, undisclosed prompt or model change, over-reliance on unreviewed output. Rate each by effect on material balances.
4. The preventive and detective controls required at each step, and which COSO component each belongs to.
5. Evidence each control should produce, in a form reproducible for a sample selected months later.
6. Open questions for the controller, SOX lead, and external auditor — do not answer them yourself.

### COSO GenAI control gap review

Mapping AI use onto the COSO framework

Map our use of generative AI in <process> onto the five COSO components: control environment, risk assessment, control activities, information and communication, monitoring.

For each component:
- What the framework expects when generative AI is involved.
- What we currently have in place, based only on what I provide — mark anything I have not evidenced as NOT EVIDENCED rather than assuming it exists.
- The gap, and the smallest concrete control that would close it.
- The evidence that control would produce, and who would sign it off.

Finish with the three gaps most likely to be raised in testing and why.

### Audit evidence sufficiency test

Before an auditor sample selection

Here is the evidence package for one AI-assisted figure: <paste lineage, formulas, versions, exceptions, approvals>.

Assess whether a SOX tester could rely on it. Specifically:
1. Can the figure be traced end-to-end to source records? Name any break in the chain.
2. Can the calculation be reperformed exactly from the retained inputs? If not, what is missing?
3. Are model, prompt, mapping, and formula versions effective-dated so the period can be reproduced under the logic live at the time?
4. Is the reviewer's engagement evidenced at an appropriate level of precision, or does it read as a rubber stamp?
5. Are exceptions and overrides recorded with reasons?

Output a pass / gap verdict per point with the specific remediation. No general advice.

### AI risk register entry for the audit committee

Committee reporting

Draft an audit-committee-ready risk register entry for our use of AI in <process>.

Fields: risk statement, inherent rating with rationale, controls in place, residual rating, control owner, monitoring metric and frequency, escalation threshold, and the next review date.

Rules: every rating must have a stated basis; anything I have not evidenced is marked NOT EVIDENCED; no more than 150 words per field; and end with the single question the committee should ask management.

### Third-party AI vendor risk review

Procurement and vendor assessment

Review this AI vendor for use in a financial reporting process: <vendor, deployment model, data flows>.

Cover: data residency and retention, whether customer data trains shared models, access and segregation controls, assurance reports held (SOC 2 Type II, ISO 27001) and their scope and period, change notification for model or prompt updates, subprocessors, incident and breach commitments, and exit and data-return terms.

For each: what is confirmed in the material I gave you, what is asserted but unevidenced, and what is unknown. End with the clauses to negotiate before signature.

## Evaluation prompts

For judging an AI tool, an output, or a vendor claim at CFO level.

### Grade an AI-produced finance output

Deciding whether output is releasable

Grade the attached AI-produced finance output as a reviewer, not an author.

Score each dimension 0-5 with evidence from the text: source binding of every figure, arithmetic correctness, completeness against the stated scope, treatment of uncertainty, disclosure of unexplained residual, absence of unsupported claims, and reproducibility.

Then list every figure that lacks a traceable source, every claim stated with more confidence than the evidence supports, and the specific edits required before release. Conclude: RELEASABLE / RELEASABLE WITH EDITS / NOT RELEASABLE, with the reason in one sentence.

### Stress-test a vendor AI claim

Vendor evaluation and reference calls

Here is a vendor claim about AI in finance: <paste claim>.

1. Restate it as a testable proposition with the metric, baseline, population, and time period it would require.
2. Identify what the claim omits that a CFO would need: who performed the work, what the comparison was, sample size, whether the figure is a projection or an observation.
3. Give the five questions that would confirm or break the claim in a reference call.
4. State what evidence would make you accept it, and what would make you reject it.
5. Do not accept the claim as fact anywhere in your answer.

### ROI case review for an AI finance investment

Business case challenge before approval

Challenge this AI investment case: <paste case>.

Required output:
1. The three metrics that actually determine payback — hours returned per cycle, cycle time, and rework or accuracy — with the baseline each requires and whether that baseline exists today.
2. Every assumption in the case, with value, basis, owner, and a sensitivity range.
3. Costs the case omits: change management, control design and testing, data remediation, licence growth with volume, review effort.
4. The payback calculation with formulas and inputs shown, plus a downside case.
5. The single condition under which this investment fails, and the leading indicator that would show it early.

Label everything FACT / DERIVED / ASSUMPTION with confidence.

FAQ

## Using LLMs for finance work — common questions

### What is an LLM prompt pack for FP&A and risk?

A set of reusable, structured prompts that make an AI assistant produce finance output a reviewer can accept: every figure cited to a source, arithmetic shown, assumptions labelled, uncertainty stated, and a reproducibility note at the end. It is a control on how you use the model, not a substitute for one.

### How do I stop an LLM from inventing financial figures?

Instruct it structurally rather than politely: forbid unsourced numbers outright, require a source tag on every figure, require the formula and inputs for any derived value, and require the literal output 'NOT AVAILABLE — required input missing' when a source is absent. The guardrail block in this pack does all four.

### How should an LLM handle uncertainty in financial analysis?

Every statement should be labelled FACT, DERIVED, ASSUMPTION, or UNKNOWN, with a confidence level and a one-line reason. Ranges should replace point estimates where a range is more honest, and any unexplained residual in a variance or reconciliation must be reported in amount and percent rather than absorbed into another bucket.

### Can I rely on LLM output for financial reporting?

Not on its own. Narrative from a language model is a preparer's draft subject to a management review control, and the numeric layer should be produced deterministically rather than generated. This pack improves the draft's reviewability; reliance still depends on controls, lineage, and sign-off.

### Do these prompts work with ChatGPT, Claude, and Gemini?

Yes. Paste the guardrail block first, then the task prompt, then your data or file. The guardrails are model-agnostic instructions about citation, arithmetic, uncertainty, and output shape rather than provider-specific syntax.

Related: [AI governance for financial reporting](/solutions/ai-governance-for-financial-reporting) · [auditor reliance checklist](/tools/auditor-reliance-checklist) · [ROI calculator](/tools/finance-ai-roi-calculator)

## Better prompts still need better plumbing.

Guardrails improve a draft. Lineage, deterministic calculation, and recorded sign-off are what make a number reliable in a reporting cycle. See both on a real workflow.

[Book a Demo](/book-a-demo)
