Budgeting and forecast prompts

A budget is an approved plan or allocation, usually tied to organizational decisions and accountability. A forecast is a current estimate of likely outcomes based on available information. A scenario asks what would happen under an explicit set of assumptions.

Do not use the words interchangeably. If actual revenue is below budget, that does not automatically mean the latest forecast is inaccurate. If a scenario shows a cash shortfall under a severe assumption, it does not assert that the shortfall will occur.

Require every AI-produced narrative to label the type, as-of date, horizon, version, owner, and source model.

Before drafting numbers, ask AI to organize—not approve—an assumption register:

Turn the supplied planning notes into an assumption register.

Return columns for: assumption ID, metric, description, value, unit,
effective period, source, owner, status [approved/proposed/unknown],
dependencies, and sensitivity direction.

Do not invent values or convert proposed assumptions into approved ones.
Mark conflicts and missing units. Preserve source wording in a notes field.

Review the register with owners. Common assumption groups include volume, price, conversion, retention, headcount, compensation, supplier cost, payment timing, tax, exchange rates, capital spending, and financing. The exact set depends on the organization.

An assumption without an owner becomes folklore. An assumption without an effective period may be applied to the wrong month. An assumption without a unit can turn 5 into five dollars, five percent, or five people.

AI can propose a driver tree:

revenue = active customers × average units × average net price

payroll = filled roles × average compensation + planned hires × start-date-weighted cost

But the finance owner must decide whether the structure reflects reality. A subscription business, marketplace, manufacturer, and nonprofit have different drivers.

Once approved, implement formulas deterministically. Use checks such as:

• totals equal component sums;
• opening balance plus movement equals closing balance;
• headcount cannot be negative;
• dates fall inside the forecast horizon;
• currency conversions use the approved rate table;
• scenario inputs are separated from formulas;
• no hard-coded values are hidden inside formula cells.

Ask AI to generate test cases for these checks, then run them in the actual model.

Avoid vague labels such as “best,” “base,” and “worst” without definitions. Prefer scenario names tied to drivers: “renewal rate down five points,” “supplier cost up eight percent,” or “hiring delayed one month.”

A scenario prompt can be:

Using the verified baseline outputs and approved assumption register below,
draft three scenario specifications. Change only the named drivers. For each,
list changed assumption IDs, direction, rationale from supplied evidence,
expected model relationships, and questions for the owner.

Do not calculate results, assign probabilities, recommend a decision, or call
any scenario likely. Keep all unchanged assumptions explicit.

The planning model computes results. AI can then draft a comparison using those verified outputs. Require it to state that scenarios are conditional, preserve units, and avoid probability language unless probabilities were independently developed and approved.

Variance analysis compares actuals with a reference such as budget, forecast, or prior period. State which reference is used. Keep volume, price, mix, timing, foreign exchange, and classification effects separate when the underlying model supports that decomposition.

Use a two-stage pattern. First, provide verified values and ask for questions:

List the five variances that meet these supplied materiality rules.
For each, show metric, actual, reference, absolute variance, percentage
variance where meaningful, unit, and source. Ask what evidence is needed to
explain it. Do not infer causes.

Second, after owners provide evidence, ask for a narrative that labels confirmed explanations and unresolved items. This prevents a convenient story from appearing before investigation.

Point estimates look more certain than the process usually is. Where appropriate, show ranges, scenarios, key sensitivities, and what would cause an update. Never ask a model to manufacture confidence intervals from narrative context.

A forecast memo should answer:

• What is being estimated, over what horizon, and as of when?
• Which model and source data produced the figures?
• Which assumptions most affect the result?
• What changed since the previous version?
• Which inputs are approved, proposed, stale, or missing?
• Which scenarios were tested?
• What decision, if any, does an accountable person need to make?
• When will actuals be compared with the forecast?

Generated prose should not say “will” when the evidence supports “could,” “is forecast to,” or “under this scenario.” Calibrated language is part of calculation integrity because it preserves the status of the number.

Budget and forecast outputs may influence staffing, pricing, financing, or investment. AI should not make personal financial recommendations or represent itself as a qualified adviser. It should not approve a budget, choose a financing instrument, set a customer’s credit terms, or direct a person to buy or sell an asset.

Use it to structure analysis and communication. Qualified, authorized people own assumptions, policy, judgment, and decisions.