
For your team
Teams by Function
Module · When AI touches the numbers
The Finance Team AI Check
Finance runs on numbers that have to be right, and AI produces numbers that are only usually right. A model that fabricates a plausible figure with total confidence is a specific hazard when the output becomes a forecast, a reconciliation, or a line in the close. This module checks the five habits that keep AI a help rather than a hidden error: whether your team verifies before trusting, whether AI outputs get reconciled to a source, whether you can audit how a number was produced, whether the same person builds and approves, and how far AI has crept into the close itself.
What the five levels look like
Every dimension in this assessment is scored 1 to 5. This is what the levels mean, dimension by dimension. The graded report diagnoses where your own answers land and what to do about it.
Verify before trusting
- 1Trusted blindly
- 2Spot-checked rarely
- 3Checked when it matters
- 4Verified before use
- 5Verified and documented
At the low end: Numbers taken from an AI without a check will eventually put a fabricated figure into a decision. Make verification against a source the default before any AI output is used, not an afterthought when something looks wrong. What good looks like: Verifying before use and recording that you did is what lets finance use AI without inheriting its errors. Keep the discipline as the tools improve; trust built on a good month is how the unchecked error gets in.
Reconciled to a source
- 1Never reconciled
- 2Taken at face value
- 3Reconciled sometimes
- 4Reconciled to source
- 5Reconciled and signed off
At the low end: AI output that never meets the system of record can drift from reality without anyone noticing. Reconcile every AI-touched figure back to the ledger before it feeds a report. What good looks like: Reconciliation to a system of record with a sign-off is the control that lets AI accelerate the work without owning the truth. Keep the source authoritative; if the AI output starts feeding the record it was meant to check, the control is gone.
The number is auditable
- 1No trail
- 2In someone's head
- 3Partial notes
- 4Inputs and method saved
- 5Full reproducible trail
At the low end: A number with no record of how the AI produced it cannot be defended when someone asks. Start capturing the prompt, inputs and model version alongside any AI-assisted figure. What good looks like: A full, reproducible trail from prompt to figure is what makes AI-assisted finance auditable. Keep it version-stamped; the same prompt on a new model can return a different number, and you will need to show which one you used.
Duties stay separated
- 1One person throughout
- 2Separation on paper
- 3Separated for big items
- 4Enforced separation
- 5Enforced, AI-aware
At the low end: If one person plus an AI can now build and approve the same work, the check that catches errors and fraud is gone. Restore a second human approver on anything AI helped produce. What good looks like: Enforced separation that accounts for what AI now enables keeps the create-and-approve boundary real. Revisit it as the assistants get more capable; each capability jump is a chance for one person to quietly reclaim both roles.
Close-cycle AI is governed
- 1Ungoverned in the close
- 2Creeping in unmanaged
- 3Used, informally
- 4Used with controls
- 5Controlled and reviewed
At the low end: AI making its way into the close with no controls puts unverified figures into your most consequential process under the most pressure. Draw an explicit line now for what AI may and may not do in the close. What good looks like: Governed AI use in the close, with the high-risk steps kept under human control, lets you gain the speed without betting the numbers. Review the boundary each cycle; deadline pressure is a constant push to let the AI do just a little more.