AI That Outputs Real Excel Files (Not Just Formulas)

AI that outputs real Excel files is a file generation capability where AI tools produce downloadable .xlsx or .xlsm files containing formulas, formatting, and multi-tab structures, rather than generating text-based formula snippets that users must manually copy into spreadsheets. This technical distinction separates AI assistants that explain Excel from those that build Excel.

Relevant Articles

• Care about the financial content? See How to Build a Multifamily Pro Forma.
• Need AI to understand deal logic? Review How AI Understands Excel Model Structure.

Working Example: Testing File Output Capability

To see this in action, let's test a specific scenario:

The test is simple: Can the AI hand you a file you can open in Excel, or does it give you instructions to build one yourself?

What Real File Output Means

Real file output is a binary technical capability. Either the AI system produces a downloadable Excel file that contains executable formulas and preserved formatting, or it does not. There is no middle ground.

When you ask ChatGPT or Claude to "build an Excel model," the response is a code block containing formulas like =B5*C5 and instructions: "Copy this into cell D5." This is formula generation, not file output. You are still the one building the file. The AI is a formula consultant.

True file output means the AI writes directly to the Excel file format. It constructs the XML-based .xlsx structure, embeds formulas into specific cells, applies number formatting, creates multiple sheets, and returns a binary file you can download and open. No clipboard. No manual reconstruction.

In our Cascade example, a formula-generation AI would give you text like this:

In cell B12, enter: =B5*(1+B6)
In cell B13, enter: =B12-B7
Format B12:B13 as Currency



A file-output AI would hand you cascade-proforma.xlsx containing those formulas already in place, formatted, and ready to calculate.

This distinction is not semantic. It determines whether you spend 3 minutes reviewing a model or 30 minutes rebuilding one.

The Technical Difference

The gap between formula generation and file output is not complexity—it is architectural. Large language models generate text tokens. Excel files are binary objects structured according to the Office Open XML specification.

To output a real Excel file, the AI must:

1. Construct the .xlsx container: Excel files are ZIP archives containing XML files that define worksheets, styles, formulas, and relationships. The AI must generate valid XML conforming to the ECMA-376 standard.
2. Write formulas in Excel's internal syntax: Excel formulas inside the XML use absolute and relative references, R1C1 notation when required, and function names that match Excel's internal library. A formula written for display (=SUM(B2:B10)) must be encoded as <f>SUM(B2:B10)</f> within the correct cell element in xl/worksheets/sheet1.xml.
3. Apply cell formatting: Number formats, fonts, borders, and conditional formatting are stored in xl/styles.xml and referenced by index. The AI must assign the correct style index to each cell.
4. Link multiple sheets: Cross-sheet references like =Assumptions!B5 require the AI to define worksheet relationships in xl/workbook.xml and ensure the target sheet exists.

In the Cascade example, building a three-sheet model (Assumptions, Pro Forma, Sensitivity) requires the AI to:

• Create three separate sheet.xml files
• Define sheet relationships in the workbook manifest
• Write formulas in Pro Forma that reference Assumptions using valid cross-sheet syntax
• Apply percentage formatting to growth rates and currency formatting to cash flows
• Embed a data table formula in the Sensitivity sheet that Excel will recognize and calculate

ChatGPT cannot do this. It generates formula suggestions because it outputs text, not binary files. AI file output vs formula suggestions is the difference between a compiler and a syntax highlighter.

Workflow Implications

The workflow gap between formula generation and file output compounds over iterations.

Scenario 1: Formula Generation Workflow

1. Describe the model to AI (5 minutes)
2. Receive formula suggestions (instant)
3. Open Excel and create sheet structure (5 minutes)
4. Copy formulas into cells one by one (15 minutes)
5. Debug reference errors from copy-paste mistakes (10 minutes)
6. Apply formatting manually (5 minutes)
7. Request a change from AI (3 minutes)
8. Receive updated formulas (instant)
9. Manually update cells and reformat (10 minutes)

Total time for first iteration + one revision: 53 minutes

Scenario 2: File Output Workflow

1. Describe the model to AI (5 minutes)
2. Download .xlsx file (instant)
3. Open and review (2 minutes)
4. Request a change (3 minutes)
5. Download updated .xlsx file (instant)
6. Review updated model (2 minutes)

Total time for first iteration + one revision: 12 minutes

The multiplier is 4.4x. For the Cascade deal, testing three different equity structures (20%, 25%, 30% down) would take:

• Formula generation: 53 + 10 + 10 = 73 minutes (updating equity cells and dependent formulas twice)
• File output: 12 + 3 + 3 = 18 minutes (downloading three versions)

File output AI also eliminates the category of errors that occur during manual transcription. When you copy =B5*C5 from a chat window into Excel, you might paste it into the wrong cell, miss a dollar sign, or break a reference. These errors do not exist when the AI writes directly to the file.

The workflow advantage is isolation. You are not an intermediary between the AI's logic and Excel's calculation engine. The AI handles both.

Quality of Output Files

Not all file output is equal. The quality of an AI-generated Excel file depends on three factors: formula correctness, structural integrity, and formatting fidelity.

Formula Correctness

Does the formula calculate the right result? In the Cascade pro forma, the Year 1 NOI formula should be:

=Gross_Rent - Vacancy_Loss - Operating_Expenses

A high-quality file output AI will:

• Use named ranges or structured references instead of raw cell addresses
• Apply absolute references ($B$5) where inputs should not shift when copied
• Avoid circular references unless explicitly modeling iterative calculations

A low-quality file output AI might generate =B12-B13-B14, which calculates correctly on the first pass but breaks when rows are inserted.

Test: Insert a blank row between Gross Rent and Vacancy Loss. Does the formula still work? If the AI used relative references to hardcoded rows, it now points to the wrong cells.

Structural Integrity

Does the file open without errors? Does Excel recognize all formulas, or do you see #NAME? and #REF! errors?

In the Cascade model, the Pro Forma sheet references the Assumptions sheet. The AI must:

• Ensure Assumptions is spelled identically in both the sheet name and the formula
• Use valid cross-sheet syntax: =Assumptions!B5, not =Assumptions.B5 or =[Assumptions]B5
• Define the Assumptions sheet before the Pro Forma sheet writes formulas that reference it

A file that contains =Assumption!B5 (missing the 's') will open, but every cross-sheet reference will error. This is a structural failure. The file is not broken—it is wrong.

Formatting Fidelity

Does the file look like a professional model, or does it look like raw data dumped into cells?

For the Cascade pro forma, formatting fidelity means:

• Currency formatting on all dollar amounts: $28,500,000, not 28500000
• Percentage formatting on growth rates: 3.5%, not 0.035
• Thousand separators on unit counts: 180, not 180.00
• Bolded headers and frozen panes on the top row
• Conditional formatting to highlight negative cash flows

How AI understands Excel model structure determines whether the AI treats formatting as cosmetic (applied randomly) or structural (applied based on data type and role).

A high-quality file output AI infers that "Entry Price" is currency, "Hold Period" is an integer, and "Cap Rate" is a percentage—and formats accordingly. A low-quality AI outputs everything as General format and leaves you to fix it.

Evaluating File Output AI

To evaluate whether an AI tool truly outputs real Excel files, run this three-part test:

Test 1: Can it produce a downloadable .xlsx file?

Ask the AI to build a simple three-column table (Item, Quantity, Price) with a Total row that sums Quantity. Do not describe how to structure it—just state the requirement.

If the response includes a download button or file attachment, proceed to Test 2. If the response is a code block with instructions, the tool does not have file output capability. Stop here.

Test 2: Does the file contain working formulas?

Open the file. Check the Total cell. Does it contain a formula (=SUM(B2:B10)), or is it a hardcoded number?

If it is a formula, change one of the Quantity values. Does the Total update automatically? If yes, proceed to Test 3. If the Total is static, the AI generated a data file, not a model.

Test 3: Can it handle multi-sheet structures with cross-references?

Ask the AI to split the table into two sheets: "Inputs" (Item, Quantity, Price) and "Summary" (Total Quantity, Total Price). The Summary sheet must reference the Inputs sheet.

Open the file. Navigate to Summary. Check the formulas. Do they reference the Inputs sheet using valid syntax (=Inputs!B10)? Change a value in Inputs. Does Summary update?

If yes, the AI has file output capability. If the file errors or the sheets are disconnected, the AI has partial capability.

For the Cascade example, a full evaluation would test:

• Three-sheet structure (Assumptions, Pro Forma, Sensitivity)
• Cross-sheet formulas linking Pro Forma to Assumptions
• A sensitivity table that uses a data table formula
• Conditional formatting that highlights negative cash flows

Most general-purpose LLMs (ChatGPT, Claude, Gemini) fail Test 1. They cannot produce .xlsx files. Specialized tools like Apers, sheet-building APIs, or Excel-native AI plugins pass all three tests, but vary in formula quality and formatting fidelity.

The Emerging Category

AI that outputs real Excel files is not a feature of conversational AI—it is a separate product category. The boundary is technical: Can the system write binary files, or only generate text?

As of early 2025, this category includes:

Category 1: Purpose-Built File Output AI

Tools designed specifically to generate Excel files for a domain (real estate, finance, consulting). These systems:

• Accept natural language descriptions of models
• Output .xlsx files directly
• Handle multi-sheet structures, cross-references, and formatting
• Support iterative revisions via re-upload

Example: Apers for real estate models, Julius for data analysis outputs, AI-powered Excel add-ins that write to the active workbook.

Category 2: API-Driven Spreadsheet Generators

Developer tools that provide programmatic file generation. These systems:

• Require structured input (JSON, API calls)
• Generate files server-side and return download links
• Support advanced Excel features (pivot tables, macros, charts)
• Do not interpret natural language directly—developers build the interface

Example: OpenPyXL (Python), ClosedXML (.NET), SheetJS (JavaScript), EPPlus (.NET).

Category 3: Excel-Native AI (Future State)

AI integrated directly into Excel as a file-writing copilot. This does not exist yet as of 2025, though Microsoft Copilot in Excel can generate formulas and pivot tables within an open workbook. It does not generate standalone files from descriptions.

The Cascade test case (180-unit multifamily pro forma) would be handled by:

• Category 1: Describe the deal, download the file
• Category 2: Developer writes code to call an API with deal parameters, API returns file
• Category 3: Not supported (Copilot modifies existing sheets but does not generate complete models)

The category is emerging because the workflow value is compounding. Early users of file output AI report 60-80% time savings on model setup, with the largest gains in iterative scenarios where assumptions change frequently. Generic AI can't build Excel models because general-purpose LLMs were not designed to output structured binary files—they were designed to generate text that humans interpret.

The next wave will be AI that outputs real Excel files and understands the domain logic embedded in those files. For real estate, that means AI that knows a waterfall distributes cash based on IRR hurdles, a pro forma projects NOI from rent rolls, and a sensitivity table should test exit cap rate and hold period—not because a user specified it, but because the AI understands the domain.

That capability is not "file output plus content knowledge." It is AI for financial modeling in Excel: domain-specific reasoning expressed as downloadable models.

‍