VC Modeling Using the NEW ChatGPT for Excel

VC and growth equity modeling has always been one of the most labor-intensive parts of the privates investing workflow.

You’re underwriting companies with no public financials, no audited revenue, and no sell-side coverage to lean on. Building an institutional LTV/CAC model means weeks of primary research just to get to a credible assumption set … and that’s before you’ve touched Excel.

AI is starting to meaningfully accelerate this process, so I wanted to put it to a real stress test. We're going to use two new AI tools that are drawing a lot of attention.

Perplexity Computer

ChatGPT for Excel

The target I chose: Anthropic. Arguably one of the hardest private companies to model right now given how fast the product, pricing, and customer mix are all moving at once.

I’m going to walk you through the full workflow I used to go from a blank page to a LTV/CAC model, including what worked well and the pitfalls to watch for along the way.

Here’s how the workflow runs:

1. Use Perplexity Computer use to build a cited research data pack (Prompt 1)
2. RAG that data pack into ChatGPT for Excel + run a structured modeling prompt (Prompt 2)
3. Audit the output and iterate to close the gaps

What would have taken a junior analyst a week of work now realistically takes a few hours of focused prompting. That’s a meaningful shift for anyone running lean on the diligence side (e.g., solo GPs or funds that are drowning in dealflow)

The Prompts

We are using two prompts to run this workflow. The first runs inside Perplexity Computer to build a cited research data room and extract a clean assumptions pack. The second takes that assumptions pack directly into ChatGPT for Excel to build the model.

Prompt 1: Use Perplexity Computer to Build the Data Room + Assumptions Pack

This step does the heavy lifting on research. Perplexity scrapes and synthesizes all available data on the company, then structures it into a cited, investor-grade assumptions file. This is the document you will RAG into ChatGPT for Excel in the next step. Do not skip this. The quality of your assumption sourcing determines everything downstream.

Prompt: https://www.finprompter.com/share/f60a6431-37e1-4ea2-bf96-3eb3be7e8246​

Result: anthropic_ltv_cac_assumptions.docx​

Prompt 2: Use ChatGPT for Excel to Build the Model

Feed the assumptions file from Prompt 1 directly into ChatGPT for Excel as context. The prompt instructs the model to map every assumption first, flag anything missing or ambiguous, and build a fully auditable multi-tab workbook.

Prompt: https://www.finprompter.com/share/416eb762-c7f8-48a9-b89d-3e3d253d36ac​

Result: anthropic_ltv_cac_model_v2.xlsx​

​

The Result

The model came out with nine clean tabs covering everything from source assumptions and segment operating build through cohort retention, a sensitivity suite, and a QA check layer. Segment-level analysis covered enterprise, team/mid-market, self-serve/developer, and API/usage-based customers, with both segment and blended outputs across LTV, CAC, payback, NRR, GRR, and churn.

Headline outputs from the base case:

• Blended PV LTV/CAC: 5.6x across all four segments
• Blended contribution payback: ~7 months, well inside what you’d expect for a high-velocity API business
• Enterprise is the value driver as expected: $779k PV LTV per customer, $150k CAC, 7-month payback, 135% NRR
• API/usage shows the most attractive LTV/CAC at nearly 12x, reflecting low CAC and strong net expansion on a surviving base
• 2026E revenue: $18B modeled, consistent with the ~$19.5B ARR run rate corroborated by official sources

A few caveats worth keeping in mind before running with these numbers.

The first pass had mechanical issues I had to work through via audit prompting. The biggest was around churn linkage in the cohort engine and some blended weighting inconsistencies across sheets. The good news is that iterating with the model to flag and fix these is straightforward if you know what to look for. A few source assumptions also needed reclassification.

After several rounds of iteration, we got to a more finalized version sense checking the underlying numbers and the model mechanics.

The Bottom Line

This workflow genuinely works, and you should budget time for an audit loop before trusting the outputs. The research data pack step is non-negotiable here. The quality of your assumption sourcing determines everything downstream.

Personal: What I’m Thinking About

The last few weeks have been crazy - everyone is shipping. Claude keeps lauching. Perplexity launched Perplexity Computer. Now ChatGPT for Excel.

It’s honestly such a privilege to be building with AI right now. I don’t say that lightly. I’ve been in this industry for some time now, and the pace today is unlike anything I’ve ever seen.

Which brings me to a question I get constantly: will AI replace my job?

Every time, I think the framing is off. AI doesn’t replace jobs. It replaces tasks. Those are very different things.

We’ve seen this before. ATMs didn’t kill bank tellers. Spreadsheets didn’t kill accountants. The internet didn’t kill journalists. The job changed. Who survives in the job changed.

Organizations won’t eliminate roles overnight. But they’ll need fewer people in those roles. The people who can’t outperform what AI produces on its own get squeezed out. The ones who 10x their output become indispensable.

Same jobs. Fewer seats. Higher bar.

If history is any indicator, the people who adapt to the tool always outlast the people who fear it.

Act accordingly.

- Felipe