Turn a Raw Financial Model Into an AI-Native Analyst Brief: A Prompt Workflow

The task

FP&A and controllership teams routinely get handed a raw model dump — a trial balance extract, a driver sheet, or a quarterly forecast — and are expected to produce a tight analyst brief for the CFO or audit committee within hours. This workflow turns that dump into a structured, defensible narrative using a chain of prompts you can run in any capable LLM. It's aimed at the person who owns the close deck, the treasury variance memo, or the pre-audit walkthrough.

Before AI

Today this usually means an analyst opens the workbook, hand-copies the numbers into a Word doc, calculates variances in a scratch tab, writes bullet commentary, then loops with a manager for tone and materiality thresholds. A clean brief takes 90 minutes to half a day depending on how messy the source is. Most of that time is transcription and re-formatting, not analysis — which is exactly the wedge the OpenAI finance function playbook argues AI should collapse.

The workflow

Step 1 — Normalize the dump into a clean fact table

Paste the raw export (TB rows, driver table, whatever you have) as your sample input. This prompt forces the model to reconcile totals before it writes a single word of commentary — catching bad copy-paste before it poisons the brief.

You are a senior FP&A analyst preparing an internal brief. Below is a raw financial data extract. Your job in THIS step only is to normalize it.

Do exactly this:
1. Parse the data into a clean markdown table with columns: Line Item | Current Period | Prior Period | Variance ($) | Variance (%).
2. Group rows under headers: Revenue, COGS, Opex, Below-the-line, Balance Sheet, Cash/Treasury (skip any group that has no rows).
3. Recompute every variance yourself — do not trust variances that appear in the source.
4. At the bottom, add a "Reconciliation check" section listing: total revenue, total opex, net income, and whether subtotals tie to the sum of their components. Flag any tie-out that is off by more than $1.
5. Add a "Data quality flags" section noting anything suspicious: sign flips, blank prior-period cells, line items with >100% variance, or duplicated account names.

Do NOT write narrative commentary yet. Output only the table plus the two check sections.

RAW EXTRACT:

Northwind Retail Co. — Q3 2026 Management P&L Extract (unaudited, from FP&A cube export)
Currency: USD thousands. Prior period = Q3 2025.

Account, Q3_2026, Q3_2025
Product Revenue - Retail, 48200, 44100
Product Revenue - Wholesale, 12750, 15900
Services Revenue, 6300, 3100
Total Revenue, 67250, 63100
COGS - Product, 31400, 28900
COGS - Services, 2100, 1050
Total COGS, 33500, 29950
Gross Profit, 33750, 33150
Salaries & Wages, 11200, 10400
Marketing, 4850, 6700
Technology & Software, 2300, 1400
Occupancy, 1900, 1850
Other Opex, 1450, 900
Total Opex, 21700, 21250
Operating Income, 12050, 11900
Interest Expense, 780, 620
Tax Provision, 2650, 2710
Net Income, 8620, 8570

Treasury snapshot (period end):
Cash & equivalents, 18400, 14200
Short-term investments, 6500, 9100
Revolver drawn, 5000, 0
DSO (days), 47, 39
DPO (days), 52, 58

Notes from controller: Services revenue includes a one-time implementation fee of $1,800 booked in September. Revolver was drawn in August to fund inventory build for holiday. Marketing underspend reflects delayed Q4 campaign launch.

Step 2 — Extract the story: drivers, risks, and audit exposure

Now the model has a clean, reconciled table. This prompt makes it think like an analyst — separating what moved, why it moved, and what the auditor or treasurer will actually ask about.

Using the normalized table and flags from the previous step, produce a structured analysis. Do not restate the table. Output these four sections as markdown:

## Top drivers of variance
List the 3-5 line items with the largest dollar impact on operating income or cash. For each: one sentence on direction and magnitude, one sentence on the most likely business driver (use the controller notes if provided; otherwise say "driver not stated in source").

## One-time vs. run-rate
Separate items that look non-recurring (one-time fees, timing shifts, unusual draws) from items that appear to be a change in run-rate. Be explicit about which is which and why.

## Treasury & liquidity read
Comment specifically on cash position, working capital signals (DSO, DPO, inventory if present), and any debt activity. Flag anything a treasurer would want to discuss this week.

## Audit / close risk flags
List items an external auditor or internal controls reviewer would probe: revenue cutoff risk, unusual reclassifications, large accruals, sign flips, or anything from the Step 1 data quality flags. If there are none, say so.

Keep each bullet to two sentences maximum. Do not speculate beyond what the numbers and notes support.

Step 3 — Assemble the CFO-ready brief

Final step. The prompt hands the model a fixed template so the output slots straight into your close deck or memo. This is where you'd swap in your own house style — tone, materiality thresholds, whether the CFO wants bullets or prose.

Assemble a final one-page analyst brief using the analysis from the previous step. Use this exact structure and headings:

**TO:** CFO
**FROM:** FP&A
**RE:** [Infer entity and period from the source data]
**DATE:** [Use today's date]

**Bottom line (3 sentences max):** The single most important takeaway a CFO needs before walking into a board prep meeting. Lead with the punchline, not the setup.

**What moved and why:** 4-6 bullets. Each bullet = line item, dollar/percent change, one-clause driver. Prioritize by materiality, not by P&L order.

**Recurring vs. one-time:** Two short paragraphs. Quantify the one-time impact on operating income so the reader can back into normalized earnings.

**Cash & treasury:** 2-3 bullets on liquidity, working capital, and any financing actions taken in-period.

**Watch items for next period:** 3 bullets. Concrete, testable — "monitor DSO trend against 45-day covenant" beats "keep an eye on receivables."

**Open questions for the controller:** 2-4 questions the analyst should get answered before this brief is finalized.

Rules: No hedging language ("it appears," "seems to"). No numbers not present in the source data or derivable from it. If a section has nothing material to say, write "Nothing material this period" — do not pad.

Gotchas

• Reconciliation is the whole ballgame. If Step 1's tie-out check fails, stop and fix the source before running Steps 2 and 3 — the model will otherwise write confident narrative on top of broken numbers.
• Controller notes are load-bearing. The "why" bullets are only as good as the free-text context you paste in. Without notes, the model will hedge or invent plausible-sounding drivers. Force it to say "driver not stated" when it doesn't know — the prompt above does this, but check the output.
• Materiality thresholds are yours to set. The default prompt treats large-dollar and large-percent items as notable. If you have a formal materiality policy (e.g., >5% of segment revenue), add it to Step 2 explicitly.
• Don't paste anything you wouldn't email. Real vendor names, employee comp lines, or unreleased M&A figures should be redacted before they leave your workbook. Use a code name for the entity if in doubt.
• This is a first draft, not a filing. The brief needs a human read for tone, forward-looking-statement hygiene, and anything that touches disclosure. Treat the output like a junior analyst's draft — fast, useful, and requiring review.

Time saved

Manually producing this kind of brief runs 90 minutes to 3 hours depending on how clean the source model is and how many stakeholders review it. The prompt chain collapses the transcription, variance math, and first-draft narrative into roughly 10-15 minutes of prompt-and-review, leaving the analyst's time for the judgment calls that actually need a human. Conservative estimate: 45-90 minutes saved per brief, higher on messier source data. The savings compound if you're running this weekly during close.

What we got when we ran it

We executed every prompt above, in order, against the sample input (model: claude-sonnet-4-6). Outputs below are real and unedited apart from length truncation.

Step 1 output

# Northwind Retail Co. — Q3 2026 Normalized Financial Data Currency: USD thousands | Current = Q3 2026 | Prior = Q3 2025 --- ## Revenue | Line Item | Current Period | Prior Period | Variance ($) | Variance (%) | |---|---:|---:|---:|---:| | Product Revenue – Retail | 48,200 | 44,100 | +4,100 | +9.3% | | Product Revenue – Wholesale | 12,750 | 15,900 | −3,150 | −19.8% | | Services Revenue | 6,300 | 3,100 | +3,200 | +103.2% | | Total Revenue | 67,250 | 63,100 | +4,150 | +6.6% | --- ## COGS | Line Item | Current Period | Prior Period | Variance ($) | Variance (%) | |---|---:|---:|---:|---:| | COGS – Product | 31,400 | 28,900 | +2,500 | +8.7% | | COGS – Services | 2,100 | 1,050 | +1,050 | +100.0% | | Total COGS | 33,500 | 29,950 | +3,550 | +11.9% | | Gross Profit (memo) | 33,750 | 33,150 | +600 | +1.8% | --- ## Opex | Line Item | Current Period | Prior Period | Variance ($) | Variance (%) | |---|---:|---:|---:|---:| | Salaries & Wages | 11,200 | 10,400 | +800 | +7.7% | | Marketing | 4,850 | 6,700 | −1,850 | −27.6% | | Technology & Software | 2,300 | 1,400 | +900 | +64.3% | | Occupancy | 1,900 | 1,850 | +50 | +2.7% | | Other Opex | 1,450 | 900 | +550 | +61.1% | | Total Opex |

…output truncated at 250 words.

Step 2 output

Top Drivers of Variance - **Services Revenue +$3,200K (+103.2%):** Largest single revenue driver, accounting for 77% of total revenue growth. Controller confirms $1,800K is a one-time implementation fee; the remaining $1,400K may represent organic services expansion. - **Product Revenue – Wholesale −$3,150K (−19.8%):** Largest revenue headwind, nearly fully offsetting the organic Services gain and partially eroding Retail momentum. Driver not stated in source; channel loss, pricing, or deliberate mix shift cannot be distinguished without further detail. - **Marketing −$1,850K underspend (−27.6%):** Single largest Opex favorable variance and the primary reason Operating Income held despite revenue mix pressure. Controller attributes this to a delayed Q4 campaign launch, making it a timing item that will reverse into Q4 cost. - **COGS – Services +$1,050K (+100.0%):** Directly offsets the gross profit benefit of Services Revenue growth; services gross margin is materially diluted when the one-time fee is excluded. Doubling of cost alongside a one-time revenue item warrants confirmation that cost recognition is appropriately matched. - **Salaries & Wages +$800K (+7.7%):** Largest structural cost increase and the only Opex line that is clearly run-rate in nature. Driver not stated in source; headcount growth or merit cycle is the most likely cause. --- ## One-Time vs. Run-Rate **One-time / non-recurring:** - **Services Revenue — $1,800K implementation fee:** Explicitly identified by the controller as a one-time September booking. Excluding it, Services Revenue grows ~$1,400K (+45%), and Total Revenue growth falls from $4,150K to $2,350K; run-rate operating income is correspondingly lower. - **Marketing underspend — −$1,850K:**

…output truncated at 250 words.

Step 3 output

TO: CFO FROM: FP&A RE: Northwind Retail Co. — Q3 2026 Operating Performance vs. Q3 2025 DATE: July 14, 2025 --- Bottom line: Reported net income of $8,620K (+$50K YoY) is flattering; strip out the $1,800K one-time services implementation fee and the $1,850K marketing underspend that shifts into Q4, and normalized operating income is approximately $2,000K lower than the headline figure suggests. The wholesale channel lost $3,150K of revenue with no explanation on record, which is the most consequential unresolved business question entering Q4. A $5,000K revolver draw funds the holiday inventory build, but net liquidity improved only ~$800K once the investment portfolio drawdown is accounted for — the balance sheet is more levered than the cash line implies. --- What moved and why: - Services Revenue +$3,200K (+103.2%): Driven by a one-time $1,800K September implementation fee per controller; organic growth is $1,400K. - Wholesale Revenue −$3,150K (−19.8%): Largest single headwind to operating income; driver not stated in source. - Marketing −$1,850K (−27.6%): Favorable variance is entirely a timing shift — Q4 campaign launch delayed; spend will reverse next quarter. - Salaries & Wages +$800K (+7.7%): Structural cost increase with no one-time component identified; treat as permanent step-up in run rate. - Technology & Software +$900K (+64.3%): No controller explanation provided; magnitude indicates a new contract rather than a one-time item. - COGS – Services +$1,050K (+100.0%): Cost doubled in line with the one-time implementation revenue; matching and cutoff require confirmation before close. --- Recurring vs. one-time: Two items are

…output truncated at 250 words.

---

This content is for informational purposes only and is not financial, investment, or accounting advice. Verify outputs against authoritative sources before use.