Where the evenings actually go
You closed the month. The numbers are reconciled and right. Then you spend three hours writing them up - the variance commentary, the exec summary, the board bullets - saying almost the same things you said last month with different figures. That write-up is exactly what a grounded LLM is good at: it is a language task over data you supply, not silent arithmetic. This is the highest-leverage, lowest-risk AI use in the whole function, as long as you keep the two rails from b1 and edit the draft yourself.
The killer use-case: numbers to narrative 6 min live
Survey after survey, the reporting write-up is the task finance teams most want to hand to AI - and the one where the payoff is fastest. The numbers are already done and checked; the model is only turning them into sentences. That framing matters: you are not asking it to know anything, you are asking it to phrase what you already know. The human owns the final word, always.
LiveFour write-ups, one move3 min▶
MD&A, variance commentary, the exec summary, and board-deck bullets look like four jobs. They are one move done at four lengths: take reconciled numbers, explain what changed, why, and what it means, in the right voice for the reader. AI drafts all four fast once you feed it the numbers and the structure. Your job shifts from typing to judging.
- Variance commentary - the line-by-line story of actual vs plan/prior. Most repetitive, biggest time win.
- MD&A - the management discussion around the statements. Longer, more standardised, style-sensitive.
- Exec summary - the one-screen "so what" for leadership. Ruthless compression is the whole skill.
- Board bullets - the deck version. Decision-relevant, no throat-clearing.
Techniques that make it board-grade 5 min live
A generic prompt gives you generic prose - the vague, padded paragraphs everyone recognises as "written by AI". Board-grade output comes from five deliberate moves. Each one trades a bit of prompt effort for a draft that needs far less editing and sounds like your firm, not a chatbot.
LiveThe five moves, and what each buys you3 min▶
| Technique | Effect on the draft |
|---|---|
| Feed reconciled numbers and last period's commentary as a style example (few-shot) | Draft copies your house voice, structure, and length - not the model's default register |
| Enforce structure: what changed / why / outlook | Every point becomes decision-relevant; no wandering paragraphs |
| Apply a materiality filter (e.g. "only variances over $50k or 5%") | The board sees the three things that matter, not fifteen that do not |
| Fix terminology up front (gross margin, run-rate, opex - your definitions) | Consistent language across sections and across months |
| Strip AI "tells" - demand specificity from the actual numbers | Kills generic filler ("various factors", "market dynamics"); forces "$240k, driven by the freight rate step-up" |
The last one matters most for credibility. AI filler is vague because the model is guessing at meaning. When you require every claim to cite a supplied figure, there is nothing left to pad with - the draft either ties to your numbers or it does not, and you can see which instantly.
Self-studySpotting AI "tells" before the board does2 min read▶
Reviewers smell AI prose instantly, and it costs you trust. Learn the tells so you can strip them on the edit pass:
- Vague causation. "Driven by a combination of factors" - if it does not name the figure and the driver, cut or replace it.
- Hedging everywhere. "May have contributed to a potential shift" - the board wants a view, not a weather forecast.
- Symmetry padding. A tidy sentence for every line even when only three lines matter. Delete the filler; keep the material.
- No "so what". It describes the movement but never says what leadership should do or watch. That missing sentence is your job.
The commentary that got sent back. An FP&A analyst shipped an AI first draft nearly unedited. The CFO returned it with one comment: "This says a lot and tells me nothing." The numbers were right; the narrative had no view. The fix took ten minutes of human editing - and became the standing rule: AI writes the draft, a human puts the spine in.
Audience: write for the board 3 min live
The same numbers get written up completely differently for the board than for your finance peers. The board is time-poor, non-specialist on the detail, and there to decide. Write for them and the whole draft tightens.
LiveFour rules for board-facing narrative2 min▶
- Concise. One screen, not three. If it does not change a decision, it does not make the cut.
- Decision-relevant. Lead with what the board should notice, approve, or worry about - not a tour of every line.
- Always the "so what". Every material point ends with the implication: what it means for the forecast, the covenant, the plan.
- No jargon. "Run-rate opex" and "phasing" are fine internally; spell out the meaning for a mixed board. Clarity reads as competence.
Bake these into the prompt as instructions, and again into your edit pass. The model can follow "write for a time-poor board, lead with the so-what, no unexplained jargon" - but only you can judge whether it actually landed.
Draft board-grade variance commentary from a bridge ★ 12 min · everyone builds
You are given a synthetic variance bridge - actual vs budget by line. You ground the model with it, enforce the what-changed / why / outlook structure and a materiality filter, get a draft, then edit it into something you would actually send. Numbers are made up and safe for any consumer tool (rail one).
Open any chat AI. Copy the variance bridge below into your prompt exactly - these are the only numbers the model may use.
Ask for commentary structured as what changed / why / outlook, with a materiality filter so only the big movers get a paragraph.
Read the draft against the numbers. Does every "why" cite a figure? Is anything invented or hedged? Flag the tells.
Edit it. Cut the filler, add the real driver where the model guessed, and write the one "so what" sentence the board needs on each material line.
Re-run with last month's commentary pasted as a style example and compare - the few-shot version should need far less editing.
Exec summary from a management pack - then edit the filler out ★ 8 min · build your own
Now the compression skill. You get a small synthetic management pack, ask for a one-screen exec summary, then do the human edit pass that removes filler and fixes the "so what" - the 30% that turns a draft into something a CEO reads and acts on.
Paste the mini management pack below. Ask for an exec summary of no more than six bullets, each decision-relevant, no jargon.
Read the draft as the CEO would. Which bullets describe rather than decide? Which hedge? Which bury the "so what"?
Edit pass: delete any bullet that does not change a decision, rewrite vague causation with the actual figure, and add the implication to each survivor.
Sanity check the tone: would you say this out loud to the board in the meeting? If not, plainer words.
The two open roles that were the real story. A raw AI summary listed all five metrics evenly. The human edit cut three, and led with the buried one: two sales roles unfilled for three months while the order book grew 12% - a revenue risk the board needed to fund now. Same numbers, completely different summary. The judgement about what matters is the human's, every time.
Try it yourself - this week ◐ 30-45 min total
- Take last month's real variance commentary (use a synthetic or already-published version if the real one is not cleared for a consumer tool - rail one) and rebuild it with the grounded prompt. Compare the AI draft to what you actually shipped.
- Save a reusable "house-style" prompt: your structure, your materiality threshold, your terminology, and one pasted example of your best past commentary.
- Run the exec-summary prompt on a real pack (synthetic numbers if needed) and time your edit pass. Note which tells you keep having to strip - those go into your house prompt as rules.
- Share one AI-drafted, human-edited write-up with a colleague and ask: "can you tell which parts I edited?" If yes, your edit pass is working.
- Optional: draft the board-bullet version of the same numbers and see how much further the compression can go.
Official sources covered
This session teaches from public finance-AI curricula. Certificates, graded assessments, and paid tooling stay with the official programs - noted honestly. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Why feed the model the reconciled numbers and last period's commentary?
The numbers ground the draft so it ties to reality; the example commentary teaches your voice, length, and structure far faster than instructions. Together they cut editing dramatically.
2 · What is the point of a materiality filter in the prompt (e.g. "only variances over $50k or 5%")?
Board narrative should be decision-relevant. A materiality filter forces the draft to focus on the handful of movements that change a decision and drop the symmetry padding.
3 · The AI produced a clean, well-formatted variance commentary. Who owns the final word before it goes to the board?
Rail two. AI drafts; a human reviews, puts the view in, and owns every word that reaches the board. "The AI wrote it" is no defence to a director or an auditor.