Using AI on finance work, safely

Print this and keep it at your desk.

A one-page reference you can take to your own team. It condenses the governance a controller, analyst, or accountant applies before and after AI touches real work. Your firm and client rules always win; this is the floor, not a substitute for them.

Before you start: the red-lines check

Thirty seconds, every time.

Right dataIs any client, company, or personal data involved? If so, it belongs only in an approved enterprise instance, never a consumer account or a self-published site.
Right toolIs this an approved instance for this data class? Consumer and enterprise versions are different products on training and retention.
Scoped accessHave you curated the folder to the minimum the task needs? An agent can see and act on everything you give it.
Approvals onDoes the work stay inside the normal preparer and reviewer sign-offs? The model is a fast preparer, not an approver.
Audit trailCan you show what went in, what came out, and who checked it?

Data tiers: what goes where

Public or synthetic dataAny tool, including a personal account.
Internal, non-sensitiveThe approved enterprise instance only.
Client, company, or personal dataThe approved instance, only after confirming the engagement allows it; when in doubt, ask before you upload.

Before it ships: the five-point review

Nothing ships without a human review.

1Verify every sourceOpen each cited source and confirm it says what the draft claims. Purpose-built professional tools still cite things that do not exist.
2Own every conclusionFor each recommendation, ask whether you would state it in your own voice to a decision-maker.
3Strip AI writing patternsRemove em dashes and formulaic framings. Rewrite until it reads the way you would explain it aloud.
4Trace every numberFollow each figure back to the input it came from.
5ReconcileConfirm totals foot, cross-references agree, and no conclusion overstates the evidence.

AI is good at

Drafting and rewriting narrative.
Extraction and summarization from documents.
Classification and tagging.
Writing formulas and code that run deterministically.
First-pass and adversarial review.

Do these yourself

The arithmetic itself; have AI build the formula instead.
Novel professional judgment.
Choosing the accounting or valuation conclusion.
Anything that ships without a human trace of sources and numbers.

The habit that matters: never ship without adversarial checks and a human review, because judgment remains the one input these tools cannot supply.

Back to all modules