Keywordsai for finance teamsBlogai for workai at workworkplace aiai for professionalshow to use ai at workai productivity
Related searchesai for finance teamschatgpt for financial commentaryai excel for financechatgpt finance month endai financial commentary at workhow to use ai for finance teams
Commentary structure and formula drafts. The ledger still wins
Finance teams do not need another place that can invent a number. They need faster structure around numbers that already exist. ChatGPT can outline month-end commentary, turn a variance list into a readable note, and draft a formula you will paste into a copy of the workbook. Using AI in finance means the figure always comes from the ledger, the subledger, or a controlled extract. If a director asks where a total came from, the answer is a cell, a report ID, or a named accountant. Chat is not an acceptable origin. Fluency in a commentary is how an EBITDA figure that was rounded wrong still sounds like control. Control is the workbook and the reconciliation, not the paragraph.
The typical miss is a commentary that uses last year's growth rate from training data, or an FX rate the model remembered, or a rounding that does not match the pack. Another miss is a formula that runs and includes intercompany twice. Both will survive a skim because finance prose is supposed to sound sure. Your method is old-fashioned on purpose. Structure the note with AI if that saves time. Fill every figure from the pack in the same hour. Run every formula in a copy. Reconcile to the signed total. Keep hypotheses in comments: maybe returns are in gross. If your organisation has an Excel copilot, the same split applies. The copilot proposes. The control owner accepts.
This essay is for people who already own a close, a forecast pack, or a management account, not for a demo that analyses a fake P and L. Fake packs teach the wrong lesson: that a fluent narrative is the deliverable. The deliverable is a number someone will sign and a note that does not contradict it. A named person still signs. Internal audit will not interview the model. They will interview you. Work as if that interview is this month, because one day it will be. Prepare for that interview.
If you cannot point to the report or the cell that produced the figure, it does not go in the commentary. Structure the sentences with AI. Source the numbers from the ledger.
What finance can draft, and what the control still owns
Useful drafts: a commentary skeleton from a variance list you paste, a formula shape from column names, a list of questions a sceptical FD will ask, a shorter version of a note you already wrote. Not useful, and not allowed: a market multiple from nowhere, a tax treatment, a going-concern view, or a restated prior period the model inferred. Paste only what the data class allows. Unpublished results do not go into consumer ChatGPT. If the only approved tool is an in-tenant copilot, stay there. Ask for [missing] on any figure not in the paste. You round according to the reporting manual. Completing a decimal is how 12.1 becomes 12.4 and then becomes a board slip.
Segregation still matters. The person who generates a commentary draft should not be the only person who can release the pack if that would break your existing control. AI does not create a new approver. It creates a faster first page. Tie the draft to a versioned extract. When the extract refreshes, the commentary is stale until you refill the figures. Do not reuse last month's chat with new adjectives. Reuse the brief, not the numbers. Last month's numbers in this month's story is a classic miss, and the model will not know the period changed unless you say so. Say so. Then check the headers in the extract anyway.
| Finance job | Let AI draft | Control owner keeps |
|---|---|---|
| Month-end commentary | Structure from a variance list | Every figure from the pack |
| Excel formula | A shape from column names | Run in a copy, reconcile to signed total |
| Question list for the FD | What a sceptic will attack | The answers from the ledger |
| Board slip number | No draft of the number itself | Cell, report ID, named signer |
A commentary brief that cannot supply the total
Write the brief with the period, the entity, and a ban on new figures. Paste the variance list from the extract, not a screenshot of a dashboard you cannot tie. Ask for a commentary skeleton with placeholders, not with invented percentages. Generate once. Type the figures from the pack. Run any proposed formula in a copy of the workbook. Reconcile to the signed total and to last period's signed total. Read the verbs: fell, rose, one-off, underlying. If the pack does not support the verb, change the verb. Save the brief in the team folder with the close checklist. The checklist already exists in good finance teams. Add one line: no figure from chat.
For formula work, describe the columns and the decision. Ban named ranges that were not pasted. After the formula calculates, pick one row that should be excluded and prove it is excluded. If you cannot, the formula is not ready for the close file. Do not fix it by asking the model to try a different function until the total looks like your memory of last month. If the total disagrees with the extract, the formula is wrong or the extract is wrong. Resolve that in the systems, not in the chat. Then delete the chat that contained unpublished figures if your policy requires it. Retention of close data in a consumer history is a control failure of its own.
Role: You are a financial commentator who refuses to invent figures.
Task: Structure a commentary from the variance list I paste. Draft formulas only from the columns I name.
Context: Period, entity, and the decision the pack supports.
Constraints:
- Do not supply a number that is not in the paste. Write [missing] instead.
- Do not round. Do not restate a prior period. Do not add an FX rate.
- Separate hypothesis from figure. Hypotheses go in a comment list, not in the total.
Output: Commentary skeleton with placeholders, any formula, and the checks I must run in Excel.
Quality checks: Where this draft could smuggle in a figure, a verb, or a period the pack does not support.Type every placeholder from the extract in the same hour. Run the formula in a copy, not in the live close file. Reconcile to the signed total. Change any verb the pack does not support. Keep the brief. Delete or retain the chat according to policy, especially if unpublished results were discussed. The pack that goes upstairs should be rebuildable by a colleague who never saw the chat. If it is not rebuildable, you do not yet have commentary. You have a private narrative.
Training-data FX, double-counted close, and chat as origin
Training-data FX and growth rates wander in when the brief says explain the movement and the paste is thin. Double-counted close happens when a formula includes parent and child, or actuals and a forecast version, and the total still looks familiar. Chat as origin happens when someone quotes the commentary in a meeting without opening the pack. Stop all three with the same discipline: thin pastes get [missing], formulas get a row-level proof, and spoken numbers get a cell in the same hour. If you manage a finance team, ask where the number lives, not whether AI was used. Using AI is allowed when the origin is still the ledger. Using AI as the origin is not a modern close.
Watch unpublished results in the wrong tool. A consumer chat with a pre-announcement pack is an incident waiting for a screenshot. Use the approved environment or do the commentary without a model. Watch over-automation of the narrative. A stack of generated commentaries that nobody can explain is worse than a slower note a senior accountant can defend. Internal audit will ask who accepted the control. The answer must be a role, not a product name. Write that role on the checklist. Then keep doing the reconciliation you already knew how to do. AI did not retire it. It only made it more tempting to skip.
| Mistake | What it looks like | What to do instead |
|---|---|---|
| Figure from chat | A total with no report ID | Cell or extract, same hour |
| Thin paste | Explain the movement, no list | Variance list, [missing] on gaps |
| Familiar total | Formula looks like last month | Prove one excluded row |
| Wrong tool | Close pack in consumer ChatGPT | Approved environment or no model |
If a board number cannot be tied to a cell in the same hour, it is not ready. Do not let a commentary paragraph become the origin. The ledger is the origin.
Related reading on StudyGrid
Read next: AI for Data Analysis AI for Excel and Spreadsheets How to Fact-Check ChatGPT. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.
What to do this week
On this close, pick one commentary note. Generate only a skeleton from a pasted variance list in the approved tool. Type every figure from the pack. Run one formula draft in a copy and reconcile. Add no figure from chat to the checklist. Speak only numbers you opened in the sheet that morning. Keep the brief if the note still matched the ledger after review. Do this on the live close, not on a dummy pack.