Build a Restaurant P&L in 20 Minutes With Claude or ChatGPT

Written by The Pragmatic CFO. 15+ years running restaurant P&Ls (and now teaching AI to do it).

Last updated 2026-08-04.

The 250-restaurant hook

I spent three weeks building the first restaurant P&L for a client in 2016. Sales report, bank statement, credit card statement, vendor invoices, every line categorized by hand. Between 2016 and 2024 I built and refined P&Ls for more than 250 restaurants the same way. Yesterday I did the same work for a similar 3-unit concept in 22 minutes. The difference isn’t talent. It’s the tools. In the LLM era there’s no excuse for not having a working P&L.

That is the entire premise of this piece. If you operate a restaurant in 2026 and you don’t have a monthly P&L that you actually read, the problem isn’t time or money or complexity. The problem is that nobody told you the playbook has changed. So here it is.

Why most restaurants still don’t have working P&Ls

Every excuse I have heard for 15 years is dead. Let me kill them one by one.

• “We’re too small.” If you have a bank account and a POS, you have enough data. A one-store taqueria has the same four inputs a 20-store chain has.
• “I don’t have an accountant.” You do not need one to produce a working monthly P&L. You need one to file taxes and to defend it under audit. Different jobs.
• “QuickBooks is a mess.” Then stop opening QuickBooks. Start with your bank statement CSV. The LLM does not care that your chart of accounts has 400 line items and half of them are duplicates.
• “No time.” Twenty-two minutes. I timed it. You have twenty-two minutes.

What you need before you start

Gather these four files first. If you have all four, the whole workflow takes about 20 minutes. If you have three, it takes 30. If you have two, stop and get the other two, because the P&L you get out the other end will lie to you.

• Bank statement CSV for the last 3 months.
• POS export with weekly or monthly sales by category.
• Credit card statement CSV for the last 3 months.
• A restaurant P&L template. You can grab mine on the Restaurant Finance Toolkit page.

You will also need access to a capable LLM. Claude Sonnet 5, released June 30, 2026, handles this workflow well on the $20/month Pro plan, which allows file uploads up to 30MB and 5 files per conversation. ChatGPT Plus at $20/month runs GPT-5.6 Sol as the flagship and GPT-5.5 Instant as the default, and either will do this job. Free tiers can technically get you there, but the file size and message limits will slow you down enough that the whole point of “20 minutes” goes out the window.

The 4-step LLM P&L workflow

Here is the actual sequence. One conversation, four prompts, roughly five minutes each. Do not skip a step and do not reorder them. Each prompt builds on the output of the last one.

Step 1: Categorize transactions

Upload the bank statement CSV and the credit card CSV. Ask the LLM to add a Category column and, where it can’t tell, a Notes column that says what it would need to know. This is the piece that used to take me three days per restaurant. It now takes about five minutes and produces something a human can review in another five.

You are helping me build a restaurant P&L. I've attached a CSV of bank transactions for a 3-unit fast-casual concept. For each row, add a column called "Category" and assign one of: Food Cost, Beverage Cost, Paper/Packaging, Hourly Labor, Salaried Labor, Payroll Taxes, Benefits, Rent, Utilities, Repairs & Maintenance, Marketing, POS/Tech, Insurance, Bank/CC Fees, Owner Draw, Other. If you can't tell, use "Uncategorized" and add a "Notes" column explaining what you'd need to know. Return the full table as CSV.

What you get back: a full CSV with two new columns. Most vendors get categorized correctly on the first pass. Sysco, US Foods, and PFG all land in Food Cost. ADP and Gusto land in Hourly or Salaried Labor. What you review are the Uncategorized rows and the odd manual transfers.

How to review: sort by Category, scan the Uncategorized bucket, and answer the Notes questions. Then re-upload the corrected file and ask it to reapply the categories. Spot-check 20 random rows in categories you did not touch. If more than one is wrong, you probably have a vendor the LLM didn’t recognize, and you need to add a plain-English rule (“Anything from Ecolab is Repairs & Maintenance, not Food Cost”).

Step 2: Aggregate into P&L categories

Now take the clean categorized table and turn it into a real P&L structure. The shape here matters. You want the order that lets you read a restaurant P&L the way an operator reads one, which means Sales at the top, COGS next, Gross Profit, Labor, Prime Cost as a subtotal, then Occupancy and Other Operating below.

Using the categorized transactions from the last step, aggregate by category and produce a monthly P&L in this exact order: Sales (from POS, I'll paste separately), COGS (Food + Beverage + Paper), Gross Profit, Labor (Hourly + Salaried + Taxes + Benefits), Prime Cost, Occupancy (Rent + Utilities + R&M + Insurance), Other Operating (Marketing + POS/Tech + Bank Fees + Other), Total Operating Expenses, Net Operating Income. Include both dollars and % of sales for each line. Show three months side by side.

What you get back: a three-column P&L in the exact structure a lender or a buyer would recognize. The math is done. The % of sales column is done. You should already be able to see whether your prime cost is where it needs to be for your concept.

How to review: check that Sales matches your POS export exactly. If it doesn’t, the LLM guessed, and you need to paste the POS numbers explicitly. Then verify that the subtotals foot: COGS + Labor = Prime Cost, and Gross Profit minus Total OpEx = NOI.

Step 3: Generate the Excel structure

This is the piece that separates the operators who use LLMs from the ones who watch demos on LinkedIn. Both Claude and ChatGPT can now produce actual Excel files with working formulas, not just tables. Claude added native file creation for Pro/Max/Team plans (enable it under Settings, Experimental, “Code execution and file creation”). ChatGPT for Excel and Google Sheets went generally available on May 5, 2026, powered by GPT-5.5, and works as a side-panel that reads formulas in place.

Rebuild that P&L as an Excel file. Column A = line items in the order above. Columns B-D = three months. Column E = trailing 3-month average. Column F = % of sales. Use SUM() for subtotals (COGS, Labor, Prime Cost, Occupancy, Other Operating, Total OpEx). Prime Cost = COGS + Labor. Net Operating Income = Gross Profit - Total OpEx. Format currency as $#,##0 and percentages as 0.0%. Bold the subtotal rows.

What you get back: a downloadable .xlsx with live formulas. Open it in Excel or Google Sheets and confirm the SUM() ranges are pulling the right rows. This is the file you send your accountant, not your first draft.

How to review: click into every subtotal cell and read the formula. If SUM(B3:B7) should be SUM(B3:B8), the LLM missed a row. Fix it once, save the template, and reuse it every month.

Step 4: Anomaly detection

This is the step everyone skips and it is the one that pays for the whole workflow. You are asking the LLM to be the second set of eyes you used to pay a bookkeeper for.

Look at the categorized transactions and the aggregated P&L. Flag anything that looks off: duplicate vendor payments in the same week, transactions that seem too large for their category, missing categories (e.g., no payroll tax entries when there's hourly labor), COGS % or labor % outside industry norms for a fast-casual concept (COGS 28-32%, labor 28-34%, prime cost target under 60%). Return a numbered list with "what to check" for each flag.

What you get back: a numbered list of 5 to 20 flags. Some will be false positives. Two or three will be real, and one of them will pay for the software subscription for the year. The most common finding I see is a duplicate ADP run because payroll got processed twice, or a vendor invoice that got hit against both the debit card and the checking account.

How to review: walk each flag to ground. Do not clear one until you have the receipt or the reason. If you clear a flag without checking, you are lying to your own P&L.

The prompts in one place, ready to copy

All four, no scrolling.

Prompt 1, Categorize transactions:

You are helping me build a restaurant P&L. I've attached a CSV of bank transactions for a 3-unit fast-casual concept. For each row, add a column called "Category" and assign one of: Food Cost, Beverage Cost, Paper/Packaging, Hourly Labor, Salaried Labor, Payroll Taxes, Benefits, Rent, Utilities, Repairs & Maintenance, Marketing, POS/Tech, Insurance, Bank/CC Fees, Owner Draw, Other. If you can't tell, use "Uncategorized" and add a "Notes" column explaining what you'd need to know. Return the full table as CSV.

Prompt 2, Aggregate to P&L:

Using the categorized transactions from the last step, aggregate by category and produce a monthly P&L in this exact order: Sales (from POS, I'll paste separately), COGS (Food + Beverage + Paper), Gross Profit, Labor (Hourly + Salaried + Taxes + Benefits), Prime Cost, Occupancy (Rent + Utilities + R&M + Insurance), Other Operating (Marketing + POS/Tech + Bank Fees + Other), Total Operating Expenses, Net Operating Income. Include both dollars and % of sales for each line. Show three months side by side.

Prompt 3, Generate Excel structure:

Rebuild that P&L as an Excel file. Column A = line items in the order above. Columns B-D = three months. Column E = trailing 3-month average. Column F = % of sales. Use SUM() for subtotals (COGS, Labor, Prime Cost, Occupancy, Other Operating, Total OpEx). Prime Cost = COGS + Labor. Net Operating Income = Gross Profit - Total OpEx. Format currency as $#,##0 and percentages as 0.0%. Bold the subtotal rows.

Prompt 4, Anomaly detection:

Look at the categorized transactions and the aggregated P&L. Flag anything that looks off: duplicate vendor payments in the same week, transactions that seem too large for their category, missing categories (e.g., no payroll tax entries when there's hourly labor), COGS % or labor % outside industry norms for a fast-casual concept (COGS 28-32%, labor 28-34%, prime cost target under 60%). Return a numbered list with "what to check" for each flag.

Common mistakes to avoid

• Uploading unredacted bank statements. Strip or redact account numbers, routing numbers, SSNs, and personal transactions before you upload. Both Anthropic and OpenAI have enterprise controls, but on consumer Pro plans, you should assume the file lives somewhere. Treat it like you would an email to your accountant.
• Skipping sample verification. Spot-check 20 random categorizations after Step 1. LLMs are right most of the time and wrong in ways that will bite you if you don’t look.
• Trusting formulas without checking SUM() ranges. The most common Excel error I see out of an LLM is a subtotal that misses the last row. Read every formula once before you trust the file.
• Not knowing your concept’s benchmark ranges. An LLM will give you generic ranges (COGS 28-32%, labor 28-34%) that apply to fast casual and mislead pizza, fine dining, and bars. Pin your own targets in the prompt.
• Treating this as a diligence deliverable. It isn’t. It is a working operator’s P&L. Your accountant still owns your tax filings.

When the LLM can’t replace a human CFO

I get this question every week. The honest answer: an LLM will replace 70% of what a bookkeeper does. It will replace roughly 20% of what a fractional CFO does. And it will not replace a real restaurant CFO for the following:

• Judgment calls on GAAP vs cash accounting, and when to switch between them.
• Complex owner add-backs. Personal expenses, family payroll, prep-kitchen allocations across units. The LLM will make plausible guesses. A CFO will make defensible ones.
• Multi-entity consolidation with intercompany eliminations. If you have three LLCs, a management company, and a real estate holdco, an LLM will produce a P&L per entity. It will not consolidate them correctly.
• PE-diligence-grade reporting. Quality of Earnings work has a specific shape, and buyers know it when they see it.
• Anything that requires context beyond the P&L. Leases, insurance policies, debt covenants, franchise fees, minimum guarantees. Those live outside the transactions and require a human who has read the actual documents.

If you are wondering whether your business has crossed the line where an LLM P&L stops being enough, read when a restaurant actually needs a CFO. The short answer: when the decisions cost more than the CFO does.

Where the industry is heading

Restaurant365’s 2026 AI Strategy Survey found that 61% of operators say AI has already reduced food costs and 62% report lower labor costs. That is not marketing copy. That is operators who took the tools out of the box and used them. Their Transform 2026 conference, November 1-3 in Scottsdale, is where the operators who have moved fastest will trade notes. Oracle NetSuite’s Restaurant Operations announcement on March 31, 2026, told you the same thing from the enterprise side. The direction is set.

What that means for you: the operators who build their own P&L in 20 minutes are not doing something clever. They are doing the new baseline. The ones still waiting on a monthly bookkeeper package that arrives on the 18th are the ones getting priced out of refinancing conversations.

What’s coming next in the AI series

This is post one of a five-part series on running the finance function with LLMs. Here is the rest of the map:

• Coming next: AI-Powered Cash Flow Forecasting. The 13-week rolling forecast, built from the same source files, plus the two prompts that catch a cash crunch six weeks out.
• Later: Menu Engineering With Claude. Feeding item-level POS data and recipe costs to get a real profitability quadrant, not a pretty matrix.
• Later: Automating Weekly Prime Cost Reports. The Monday-morning report every operator should have on their phone, generated end-to-end.
• Later: AI Vendor Contract Review. Feeding a broadline food contract to an LLM and getting the three clauses to negotiate before you sign.

Subscribe on the toolkit page to get the whole series as it drops. No filler, no cross-posts, just the workflow.

FAQ

Which LLM works best for restaurant P&L work?

Claude Sonnet 5 and ChatGPT with GPT-5.5 or GPT-5.6 Sol both handle the four-step workflow in this post. I use Claude for the categorization and anomaly steps because it handles long CSVs cleanly, and either tool for Excel generation. If you have to pick one, pick the one you will actually open every Monday.

Is it safe to upload my bank statement to Claude or ChatGPT?

Safe enough for an owner-operator using a Pro plan, if you redact account numbers and any personal transactions first. If your business has an enterprise agreement or you handle sensitive customer data, use the enterprise or team tier where inputs are excluded from training and retention is controlled. Never paste employee SSNs.

How do I know the LLM categorized my transactions correctly?

Spot-check 20 random rows across categories you did not review. If more than one is wrong, add a rule to the prompt (“Ecolab is Repairs & Maintenance”) and rerun. Then run the anomaly-detection prompt in Step 4, which will catch most of what a spot-check misses.

Can I automate this monthly?

Yes. Save your Step 1 through Step 4 prompts as a project or a saved conversation. Each month, drop the three fresh CSVs in and rerun the four prompts in order. The whole thing takes about 20 minutes once you have the template.

Do I still need an accountant or a restaurant CFO?

You still need an accountant for tax filings, audits, and anything that gets signed under penalty of perjury. You need a CFO when your decisions get expensive enough that being wrong costs more than being right pays. An LLM covers the middle: the working monthly P&L that used to require a bookkeeper.

Cited sources

• Anthropic, Claude Sonnet 5 announcement.
• Anthropic, Create and edit files with Claude.
• Anthropic, Use Claude for Excel.
• OpenAI, ChatGPT for Excel and Google Sheets.
• Restaurant365, Transform 2026 announcement and 2026 AI Strategy Survey.