World Model Readiness
Engraved functions instrument

For your company

AI in the Functions

Module · Numbers that reach the board unchecked

The AI-in-Finance Check

A model will produce a variance analysis that looks flawless and cites a figure that does not exist. In finance that is not an embarrassment, it is a misstatement with your name under it. This module checks the five disciplines that decide whether AI speeds the close or quietly corrupts it: verification of the figures, traceability to a source, a clean line between spreadsheet and model, an audit trail of adjustments, and close-process control that does not bend for a fast answer.

Question 1 of 5 · Figures are verified

Before an AI-produced number reaches a report, does someone verify it?

A model states a wrong figure with the same confidence as a right one, and finished-looking output invites you to skip the check. A required verification step, sized to the stakes, is the line between a faster analyst and a faster path to a misstatement.

Question 2 of 5 · Sources are traceable

For any AI-produced figure, can you trace it back to a source?

A number you cannot trace is a number you cannot defend to an auditor, a regulator or a nervous CFO. When AI assembles figures from mixed inputs, the chain from result back to source data is what turns output into evidence rather than assertion.

Question 3 of 5 · Spreadsheet-AI line is clean

Do you know where AI stops and your spreadsheets take over, and vice versa?

The dangerous mix is the invisible one: a model output pasted into a working sheet, then treated as if a human calculated it. When the boundary between AI-generated and human-built is undocumented, nobody knows which numbers were checked and which were assumed.

Question 4 of 5 · Adjustments are logged

When someone changes an AI-produced figure, is the change recorded?

Analysts override AI output constantly, and rightly so. But an adjustment with no record of who, what and why is a hole in the audit trail exactly where judgement entered. The log of overrides is often more important than the original number.

Question 5 of 5 · Close discipline holds

Does your close process survive the pressure to just use the AI number?

The month-end crunch is where controls quietly get skipped for speed. If a fast AI figure can bypass review, reconciliation or sign-off when the deadline bites, then your close discipline exists only on the calm days, which are not the ones that matter.

For the statistics · one click each

Three questions for the public picture

These do not affect your score. They feed the anonymised, aggregated statistics; groups under 8 respondents are never shown.

Where does AI sit in your finance work today?

Not used
Analysis and drafting only
Feeds numbers into reports
Part of the close process
We are not sure

How are AI-produced figures checked before they are used?

Verified against source
Reviewed if material
Glanced at
Taken on trust
No AI figures yet

Has an AI-produced figure ever reached a report with an error?

Not that we know of
Caught before it mattered
Yes, corrected quietly
Yes, with consequences
We could not tell

Your context

Used to calibrate the report. Company size and sector remain in the anonymized dataset; your email does not.

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.

Figures are verified

  1. 1Taken on trust
  2. 2Eyeballed
  3. 3Checked when material
  4. 4Verified before use
  5. 5Verified against source

At the low end: Trusting AI figures unread is how a hallucinated number ends up in a board pack. Require that every figure headed for a report is checked before it gets there, starting today. What good looks like: Figures verified against their source are the only ones that belong in a report you sign. Keep the bar tied to materiality; the check should scale with the cost of being wrong.

Sources are traceable

  1. 1No traceability
  2. 2Source is the prompt
  3. 3Partial trail
  4. 4Traceable to data
  5. 5Full documented lineage

At the low end: A figure with no traceable source is an assertion, and assertions do not survive an audit. Require that AI-produced numbers carry a pointer back to the underlying data before they enter any report. What good looks like: Full documented lineage is what lets you defend an AI-assisted number under scrutiny. Keep it attached to the figure, not filed separately; a trail nobody can find is not a trail.

Spreadsheet-AI line is clean

  1. 1Mixed invisibly
  2. 2In people's heads
  3. 3Loosely marked
  4. 4Clearly separated
  5. 5Labelled and controlled

At the low end: AI output blended invisibly into spreadsheets means you no longer know which figures a human stands behind. Mark what came from the model so the checked and the assumed are never confused. What good looks like: A labelled, controlled boundary between model and spreadsheet is what makes the whole workbook auditable. Keep the convention enforced; one unlabelled paste reopens the ambiguity you closed.

Adjustments are logged

  1. 1No record
  2. 2Value only
  3. 3Change noted
  4. 4Who and what
  5. 5Who, what and why

At the low end: Unrecorded adjustments erase the moment human judgement changed the numbers, which is the moment an auditor most wants to see. Start logging every override with who made it and why. What good looks like: A full adjustment trail of who, what and why is what separates defensible judgement from unexplained edits. Keep it immutable; a log that can be quietly rewritten is not evidence.

Close discipline holds

  1. 1Controls skipped
  2. 2Bends under deadline
  3. 3Mostly held
  4. 4Controls hold
  5. 5Controls hold, tested

At the low end: A close that skips controls for an AI number under deadline pressure has no controls, only intentions. Make review and sign-off non-negotiable regardless of how finished the output looks. What good looks like: Close controls that hold under deadline pressure are what let you move fast without moving loose. Test them deliberately in a real crunch; a control you have never stressed is one you have never verified.