.entry-content pre, .wp-block-post-content pre { overflow-x: auto; white-space: pre-wrap; word-break: break-word; max-width: 100%; } .entry-content pre code, .wp-block-post-content pre code { white-space: pre-wrap; word-break: break-word; display: block; max-width: 100%; }

Most CFOs spend Friday afternoons on weekly review that could be a 30-minute prompt session. The problem isn’t the LLM. The problem is you’re asking one open-ended question (“summarize this week”) instead of running a structured 5-prompt sequence. Here’s the exact playbook I run every Friday at 2:00 PM.

This is not a “someday when the tools mature” essay. I run this today across an operating restaurant portfolio and a set of controllership engagements. Prompts are copy-paste ready; the privacy problem gets its own section below.

Prompt 1: Variance Narrative from the Weekly P&L

What it does. Turns a weekly P&L (actual vs. budget vs. prior week) into a written variance narrative you could send to a board member.

Copy this exactly:

You are a CFO writing a weekly variance narrative for the executive team.

Input data (CSV, redacted): [PASTE WEEKLY P&L WITH COLUMNS: Line Item, Actual, Budget, Prior Week, Variance $, Variance %]

Rules:
1. Lead with the 3 largest absolute-dollar variances (favorable or unfavorable).
2. For each, propose 2 plausible drivers based on line-item behavior (do not invent facts I did not give you).
3. Flag anything above 15% variance to budget OR above 10% swing versus prior week.
4. End with a one-line "so what" for the CEO.
5. Plain prose. No bullet lists over 5 items. No hedging language.

Business context: [ONE SENTENCE: "SaaS company, $[X]M ARR, gross margin ~[Y]%"]
Week being reviewed: [WEEK ENDING DATE]

Which LLM. Claude Sonnet 4.5 or newer. It’s my default for connected numerical reasoning across a table, because it doesn’t lose track of which row a number came from. GPT-5 handles it too; Sonnet is more disciplined about not inventing drivers you didn’t feed it.

Good output. Three to five sentences of prose, each anchored to a real line item, each ending in either a quantified variance or a flag. Not a bulleted list of “insights.”

Common failure mode. Pasting the raw P&L without column headers or without a “Prior Week” column. The model then hallucinates a comparison. Fix: always include the columns you want compared, and put a literal “n/a” in empty cells so the model doesn’t guess.

Prompt 2: Cash Position and 4-Week Forecast

What it does. Takes current cash, the last 8 weeks of net cash flow, and known upcoming AP and payroll, and produces a rolling 4-week direct cash forecast.

Copy this exactly:

Act as a CFO running a 13-week cash forecast in direct method. I need a 4-week outlook.

Current cash: $[X] as of [DATE]
Last 8 weeks of net cash flow (weekly, oldest first): [WEEK1, WEEK2, ..., WEEK8]
Known items in the next 4 weeks:
- Payroll runs on: [DATES, AMOUNTS]
- AP checkruns on: [DATES, AMOUNTS]
- Debt service: [DATE, AMOUNT]
- Known large customer receipts: [DATE, AMOUNT, PROBABILITY]

Produce:
1. Week-ending cash for W+1, W+2, W+3, W+4 in a table (base case only).
2. A one-paragraph confidence assessment. Where is the biggest uncertainty?
3. The single week where cash is tightest, and by how much versus my minimum operating cash of $[MIN].
4. Do not "smooth" the forecast. If payroll drops cash by 30% in one week, show it.

Return only the table and the assessment paragraph. No preamble.

Which LLM. GPT-5 or Claude Sonnet 4.5. Both are fine. If your inputs are messy (mixed date formats, mixed sign conventions), Claude tolerates the mess better.

Good output. A clean 4-row table plus a paragraph that names the specific tight week and its specific driver. “Week of Aug 22 is tightest at $X, driven by semi-monthly payroll landing before the Anthem receivable clears.”

Common failure mode. Not telling the model your minimum operating cash. Without a floor, “the forecast looks fine” means nothing. Always give a threshold.

Prompt 3: Anomaly Detection (vs. 8-Week Trailing Average)

What it does. Scans the current week’s line items against an 8-week rolling average and surfaces what’s off. This is where the LLM earns its keep, because a human doing this by hand is eyeballing 60 rows and hoping.

Copy this exactly:

You are a finance analyst doing anomaly detection on a weekly P&L.

Data (CSV): [LINE ITEM, CURRENT WEEK, 8-WEEK AVG, 8-WEEK STDEV]

Rules:
1. Flag every line where current week is more than 2 standard deviations from the 8-week mean.
2. Also flag any line where current week is more than 25% off the 8-week mean in absolute terms, even if within 2 stdev (because for small-dollar lines, stdev is noisy).
3. For each flag, state: line item, current value, expected range (mean +/- 1 stdev), and a one-phrase hypothesis.
4. Do not flag lines with 8-week mean below $[SMALL DOLLAR THRESHOLD]. They are noise.
5. If nothing is flagged, say so explicitly. Do not manufacture a flag.

Output: bulleted list of flags only. Nothing else.

Which LLM. Either. If you have a code interpreter (ChatGPT’s data analysis feature or Claude’s analysis tool), turn it on. The model will actually compute stdev instead of estimating it, and answers get more reliable.

Good output. Three to six flags. Fifteen flags means either your threshold is too tight or your data has weekly seasonality the model doesn’t know about. Feed that context in.

Common failure mode. Feeding raw weekly totals without the stdev column. The model then estimates stdev, which is where hallucinated numbers creep in. Compute it in a spreadsheet formula and include it.

Prompt 4: Executive Summary for the CEO (5-Bullet Email-Ready)

What it does. Compresses the outputs of prompts 1 through 3 into a five-bullet email your CEO will actually read.

Copy this exactly:

Draft a 5-bullet Friday close email from CFO to CEO. Tone: direct, no throat-clearing, no "great news" framing.

Inputs to synthesize:
- Variance narrative: [PASTE OUTPUT FROM PROMPT 1]
- Cash forecast: [PASTE OUTPUT FROM PROMPT 2]
- Anomaly flags: [PASTE OUTPUT FROM PROMPT 3]

Bullet structure (in order):
1. Revenue vs. plan (one sentence, dollar and percent).
2. Gross margin movement (one sentence, driver named).
3. Cash position and the one tight week ahead (one sentence).
4. The single biggest anomaly worth the CEO knowing (one sentence).
5. What I am doing about #4 by end of next week (one sentence).

Subject line: "Weekly close, [WEEK ENDING DATE]"
Sign-off: none. I add it manually.
Total length: under 120 words.

Which LLM. GPT-5. It writes tighter executive prose out of the box.

Good output. Five short bullets, each one sentence, no adverbs, subject line that names the week. If bullet 5 reads “we will monitor the situation,” reject and rerun with “bullet 5 must name a specific action, owner, and deadline.”

Common failure mode. Not feeding the prior three outputs back in. Skip that step and the model invents generic finance language.

Prompt 5: Open Questions for Monday

What it does. Generates the short list of questions to bring to Monday’s operating meeting. Saves me from the “I meant to ask about that” moment on Tuesday.

Copy this exactly:

Based on the variance narrative, cash forecast, and anomaly flags below, generate 5 to 8 open questions I should raise in Monday's operating meeting.

Inputs:
[PASTE OUTPUTS FROM PROMPTS 1, 2, AND 3]

Rules for questions:
1. Each question must be answerable by an operator (not by another finance question).
2. Each question must reference a specific line item, region, product, or customer from inputs.
3. Do not ask "why is X down?" Ask "is X down because of [specific hypothesis]?"
4. Order by urgency: cash-affecting first, revenue-affecting second, margin-affecting third.
5. Skip anything you cannot ground in the inputs. Do not invent.

Output: numbered list of questions. Nothing else.

Which LLM. Claude Sonnet. Its instinct to ground answers in inputs, rather than pattern-match to “generic CFO questions,” is exactly what you want.

Good output. Questions like “Is the 12% dip in Region 3 revenue driven by the Acme contract slipping from Q3 to Q4, or by lower transaction volume from the existing base?” Testable, specific, pre-hypothesized.

Common failure mode. Accepting the first output. Reject any question that starts with “why” and rerun. “Why” questions run lazy meetings; “Is it X or Y” questions run meetings in 20 minutes.

The Privacy Problem

Real CFOs cannot upload real financials to a public consumer LLM. Even if you personally trust the vendor, your board, auditors, and general counsel will not. Here is how I handle it.

Redaction with normalized scaling. Multiply every dollar figure by a constant (say, 0.37 or 3.14) before pasting. Variance percentages, ratios, and directional patterns are preserved; absolute numbers are meaningless. Unscale after the model returns its analysis. Covers 80% of the weekly review workflow. The one-line formula:

# Python
df["Actual_scaled"] = df["Actual"] * 0.37
df["Budget_scaled"] = df["Budget"] * 0.37
# then paste df[["Line Item","Actual_scaled","Budget_scaled",...]] to the LLM

Synthetic data. For prompt templates you want to iterate on, generate a synthetic P&L with realistic distributions and refine your prompts against it. Once the prompts work on synthetic data, run them on scaled real data.

Enterprise LLMs. ChatGPT Enterprise, Claude for Work (Team or Enterprise), and Microsoft 365 Copilot with commercial data protection all commit that inputs are not used to train the vendor’s models by default, and come with DPAs your legal team can review. This is the path most mid-market CFOs I know are on in 2026. Read the vendor’s data terms yourself before you assume; specifics matter and they update.

What I would not do. Paste unredacted P&L, cap table, or personnel comp data into a free consumer chatbot. Not because the vendor is malicious, but because you have no audit trail. The day your auditor asks how confidential data was handled, “I typed it into ChatGPT” is a bad sentence.

Building the Friday Close Ritual

A tool that lives on your “someday” list produces zero enterprise value. Here is how I make the sequence actually happen every week.

Calendar block, 2:00 to 2:30 PM Friday. Recurring. Titled “5-prompt close,” not “financial review.” The specificity matters; you will not skip a block that names its own workflow.

Pre-block prep (10 minutes, 1:50 PM). Pull weekly P&L, cash balance, and last-8-week trailing data into one worksheet. In Netsuite, QBO, or Sage Intacct, this is a saved report you refresh weekly. If it’s a manual export today, spend an hour this quarter automating the pull; it pays back within a month.

The 30-minute session. Run prompts 1, 2, and 3 in parallel across two browser tabs. Prompt 4 depends on the first three, so it comes fourth. Prompt 5 last. Real time: 18 to 25 minutes on clean data.

Wednesday retrospective. Once a month, spend 20 minutes rereading last Friday’s outputs. Which prompts produced actionable output? Which produced finance-speak? Edit the templates. This compounding step is where the workflow gets sharper each quarter.

Frequently Asked Questions

Which LLM should I use for financial work in 2026?

Claude Sonnet 4.5 for variance narratives and open questions, GPT-5 for executive summaries and forecasting. Pick the one your company has an enterprise contract with and standardize. Switching between three vendors weekly is the fastest way to abandon the workflow.

Can I upload my company’s real P&L to ChatGPT?

To the free consumer version, no. To ChatGPT Enterprise or the Team tier with commercial data protection, yes, subject to your company’s data-handling policies and any client contractual restrictions. Read OpenAI’s enterprise privacy commitments and have legal sign off before making it standard practice.

How much does an AI-native weekly close cost?

Claude for Work Team is roughly $30 per user per month. ChatGPT Business is $25 per user per month. Microsoft 365 Copilot is $30 per user per month on top of your existing M365 license. Roughly $360 per year per finance person. Against the four hours of Friday time this replaces, the ROI is not close.

What if my LLM hallucinates a variance number?

It will, occasionally. Two defenses. The prompts above ask the model to reference specific rows you fed it, not compute new numbers, which shrinks the surface area. And always sanity-check the top-3 variances against the source spreadsheet before you send anything to the CEO. Thirty seconds of eyeballing catches most mistakes.

Can I automate this so it runs Fridays automatically?

Yes. I have a version wired through a scheduled Python job that pulls from QBO, scales the numbers, hits the Claude API, and emails me the outputs at 1:55 PM Friday. But don’t start there. Run the manual version for at least 6 weeks first so you know what “good output” looks like for your business. Automating a workflow you don’t understand produces automated nonsense.

What’s Next in This Series

More prompt playbooks coming: month-end close, board-deck drafting, budget-to-actual walk narratives, and the FP&A hire’s first-90-days prompt set. If there’s a specific workflow you want covered, the site has a contact form; use it.

Cross-Reference

Running a restaurant? The same 5-prompt structure adapted for hospitality unit economics (prime cost, weekly food and labor variance, tip-out reconciliation) lives at restaurantbottomline.com.

Sources

• Anthropic, Claude Sonnet model overview
• OpenAI, Enterprise privacy commitments
• Microsoft, Data, privacy, and security for Microsoft 365 Copilot

Written by The Pragmatic CFO. 15+ years running P&Ls and AI-native finance experiments across restaurants and portfolio companies.