Finance use cases and limits

Finance work mixes language, tables, formulas, policies, and judgment. AI is strongest when the task involves variable language or an early draft that a person can inspect. Bounded uses include:

• explaining a financial term in plain language;
• proposing questions about an income statement, balance sheet, or cash-flow statement;
• extracting labeled fields from a consistent, approved document into a review table;
• summarizing management commentary while preserving citations;
• drafting variance commentary from verified figures and supplied assumptions;
• turning an approved scenario into an executive-friendly narrative;
• checking a report for inconsistent labels, missing caveats, or unexplained changes;
• generating test cases for a budgeting model.

These are assistance tasks. The output remains a proposal. A person verifies it against records, calculations, definitions, and policy before it informs a decision.

Many finance operations are exact. A spreadsheet formula, database query, accounting system, or tested program should calculate totals, compare thresholds, convert currencies under an approved rate, apply a chart of accounts, or reconcile transactions. Those tools are easier to repeat and audit.

Do not use a language model as a calculator merely because it can print arithmetic. It may copy a value incorrectly, confuse thousands with millions, use the wrong sign, or silently change a denominator. Let deterministic software compute; let AI explain the verified result.

The same separation applies to authority. AI may draft a proposed explanation for an unusual expense. It should not approve the expense, post a journal entry, release a payment, change a credit limit, file a return, or send a material disclosure. Consequential actions need authenticated systems, explicit permissions, policy controls, and appropriate approval.

Before prompting, define four boundaries.

Purpose: Name the exact review question. “Analyze finances” is too broad. “Identify the three largest month-over-month operating-expense changes in the supplied management report” is inspectable.

Evidence: State which files and periods are authoritative. Distinguish audited statements, management accounts, exports, estimates, and commentary. Tell the model to use only supplied evidence and to mark missing information.

Allowed transformation: Specify whether the model may extract, compare, explain, draft, or generate questions. If calculations are provided, say they are inputs to explain, not values to recompute or “correct.”

Decision boundary: State what the output cannot decide. For example: “Do not recommend securities, approve spending, determine tax treatment, assess legal compliance, or initiate transactions.”

A useful prompt might say:

Using only the supplied, approved monthly report, identify notable changes
that deserve human investigation. Quote the row label, current value, prior
value, unit, and source location. Do not infer causes. Separate observations,
questions, and management-provided explanations. Do not make decisions or
financial recommendations.

Finance language contains traps. “Revenue increased 20%” is incomplete without a baseline and period. “Margin rose five points” differs from “rose five percent.” A favorable expense variance may mean spending was below budget, while a favorable revenue variance may mean revenue was above budget. Parentheses can mean negative amounts. Cash, profit, bookings, billings, and recurring revenue are not interchangeable.

Models may also combine sources that use different accounting bases, fiscal calendars, currencies, consolidation scopes, or definitions. They can invent causal explanations from correlation: lower marketing spend and lower sales in the same month do not prove one caused the other.

Require the output to preserve labels, signs, units, dates, scope, and uncertainty. If the source does not explain a change, the correct output is a question, not a confident story.

Classify the proposed use before deployment:

1. Low consequence: learning terminology or generating generic review questions with no sensitive data.
2. Review support: drafting commentary from approved internal figures, with source checks.
3. Decision support: influencing budgets, pricing, credit, hiring, forecasts, or external reporting.
4. Execution: changing records, moving money, filing, disclosing, or communicating a binding decision.

Controls should increase at every level. Decision support may require subject-matter review, documented assumptions, access controls, and independent validation. Execution should stay outside a general-purpose model unless a formally governed system constrains every action—and even then, many actions require human authorization.