The accuracy crux: AI is not a calculator

This is the most important lesson in the course, and the discipline that makes AI safe for finance: an LLM is not a calculator, and in finance, a confidently-wrong number is more dangerous than no number — because it flows straight into statements, decisions, and filings.

Why AI gets numbers wrong: an LLM generates text by predicting the next token from patterns — it has no arithmetic engine. It "simulates the syntax of calculation without preserving its mathematical invariants." So when you ask it to compute a variance, a margin, a growth rate, or a total in prose, it's pattern-matching what a plausible number looks like, not calculating. Accuracy is high on simple lookups but degrades sharply on multi-step, multivariate financial calculations — and it states fabricated numbers (a wrong margin, an invented headcount, a miscalculated growth rate) with the same confidence as correct ones. In finance, that confident fabrication is uniquely dangerous: a wrong number in a board deck, a forecast, or a filing has real consequences.

**The rule: never trust an AI-narrated number.** If AI types a figure into a sentence with no computation behind it, treat it as a guess — unusable for finance until verified.

The technique — have AI WRITE what a real system COMPUTES: instead of AI doing math "in its head," have it write the artifact that a deterministic system executes, and let the real system produce the number:

• Excel runs the formula AI wrote → the cell result is Excel's arithmetic.
• A database runs the SQL query AI drafted → the number is the database's real count/sum.
• Python (in a code-executing tool) runs AI's code → the calculation is real code, not a prediction.
• Your ERP/reporting system runs the report definition → the figures come from the system of record.

This is the "deterministic AI" / hybrid pattern: **the LLM interprets your question and hands the computation to fixed, correct code; a real system computes; and the AI explains the verified result.** The mental model (from the data-analysis course, but higher-stakes in finance): AI is the analyst who writes the recipe; the spreadsheet, database, or ERP is the oven that actually cooks it. Trust numbers from the oven; never trust numbers the AI merely narrates.

A specific finance benefit — "formula/query provenance": because AI writes the formula or query, you get an auditable trail — the exact logic behind every number is visible and reproducible. This isn't just accuracy; it's auditability (Module 3's controls), which finance and auditors require. "Here's the SQL/formula that produced this figure" is exactly the evidence finance needs.

The verification and audit-trail discipline to bake into every task:

1. Ground on verified data (the actual ledger, the source system) — never free-generation of numbers.
2. Show the formula/query, not just the answer — tie every number to its source and logic.
3. Reconcile AI output to a system of record before using it.
4. Keep the audit trail — the prompt, inputs, sources, the model, the output, and the human reviewer (Module 3).

A subtle trap to watch: AI tools sometimes narrate a summary number in prose that doesn't match the number their own code produced. Always trust the executed output (the computed cell, the query result) over the AI's sentence about it — read the real result, not the AI's description.

The mindset: in finance, exactness is everything, and AI's fundamental weakness is that it's a plausible-text predictor, not a calculator — so it will confidently state wrong numbers that, in finance, become material errors. The discipline that makes AI safe: AI writes the formula, query, or code; a real system computes the number; you verify and keep the audit trail. Trust only executed results; never a narrated figure. This single principle — write-don't-narrate, verify, audit — is what separates AI that accelerates finance from AI that corrupts your financials with confident fabrications. Every application in this course rests on it. Internalize it, and you can use AI powerfully in finance. Ignore it, and you're one hallucinated number away from a material misstatement.