🚀 How to Have Claude Audit Your Startup's Financials | The AI CFO

My favorite thing about AI is that it has created a golden age of experimentation. Execution is becoming commoditized, and value is increasingly shifting towards asking the right questions. In this week's essay, I explore what's possible when you ask Claude to build a 'finance second brain' from the data your business is already producing.

Elsewhere, what Rillet's fundraise says about the future of the finance stack, and perspectives on the future of work in a post-AI world. Let's get into it!

How to have Claude audit your startup's financials

One afternoon a few weeks back, I asked myself a question that sent me down a rabbit hole:

"How much could Claude learn about our business from running a structured analysis of our financial data?"

I started a Claude Code session and set out to run a 'Claudit' of airCFO's financials. A couple hours later, I shared a 3,330 word output with my cofounder Justin (a CPA with 20+ years of experience in startup accounting).

His feedback told me I was onto something:

What's more, the outputs Claude created are more than just a document to be filed away. They serve as invaluable context that I can provide to Claude in future sessions, a 'finance second brain'.

In this essay, I break down the step-by-step approach I took to create this context profile, and explore what this could unlock for AI-powered finance professionals.

The Context Gap

LLMs have gotten significantly better at completing finance tasks over the past year, judging from the benchmarks. And yet Anthropic's analysis shows that finance is one of the areas with the biggest gap between potential & demonstrated usefulness.

If you drop your financial statements into a fresh Claude session and ask it to produce a breakdown, you'll get a report that is technically correct but mostly unhelpful. The model has all the intelligence it needs; what's missing is the context.

So how do you get Claude up-to-speed so it can operate as a competent financial analyst? The same way you'd onboard a new hire: giving it access to your systems, providing a baseline of knowledge, letting it explore on its own and compile a set of questions that only a human can answer.

1) Your Systems: The Data Itself

Every finance team is sitting on a goldmine of context that they haven't extracted: their general ledger system. A GL is an ultrarich source of data on a company's mechanical context: it shows how a business generates revenue, the costs required to service that revenue, who the key customers/vendors are, and how the shape of the business has shifted over time.

The main blocker is that most GL tools do not provide connections that give Claude full access to the data inside. I've written about my frustrations with QBO's Claude Connector, so I created my own custom QBO MCP that lets Claude access our full dataset.

After giving Claude full (read-only) access to our QBO environment, I ran a quick back-and-forth to generate a list of questions that an analyst would ask when reviewing a company's financials for the first time. Then, after validating the research plan, I let it loose. A few minutes later, it came back with a shockingly thorough breakdown of our financials.

This initial pass gave Claude a solid understanding of airCFO's business (it self-graded as a B+), but there were a couple big gaps that still needed closing.

2) Your Model: The Knowledge Around It

The general ledger shows where your business has been, but it says nothing about where the business is heading or which performance indicators matter to leadership.

To fill this gap, I shared airCFO's financial model and a few board write-ups. Claude ingested this information and incorporated it into the writeup. Now, Claude could see the full set of key metrics that we use to steer the business, and it understood the drivers that will determine airCFO's performance moving forward.

At this point, the profile looked fairly comprehensive to me (Claude's self-grade: A-), but I knew there was additional context that didn't live in any system or spreadsheet.

So I decided to flip the script and let Claude ask some questions.

3) Your Head: The Part Only You Can Supply

To round out our finance second brain, I told Claude:

"By now you've created a comprehensive writeup of airCFO's business, but there is some information that lives only in the heads of our Finance team. Imagine you have a one-hour working session with company leadership, and your goal is to fill in any remaining gaps in your understanding of how our business works so you are able to analyze new results in future cycles. Give me a list of the most important questions you still need answers to."

This prompt resulted in a list of 24 questions (not sure how Claude came up with that number) - a few of them were off-base, but the majority were valid & easily answerable from my seat as CFO.

I gave my answers, Claude folded them into its second brain, and it achieved an A-level understanding of how our business works.

Your Finance Second Brain

This exercise was eye-opening for me on several levels:

1. It helped me think through which types of knowledge are most important for onboarding an LLM to a finance team
2. It made me realize that Claude can pull this information itself if I give it the tools & prompts to do so
3. Additionally, Claude's analysis surfaced some interesting questions about our finances that I took back to our finance team (and we're an accounting firm!)

My next step here is setting up the infrastructure that turns this one-off process into an always-on system, so Claude can maintain memory of the current state of our finances. I'm 100% convinced that finance teams who do the work to give their LLMs this background knowledge will give themselves a powerful tool that helps their company make better strategic decisions.

Have you experimented with building your own 'finance second brain'? If so, I'd love to hear what worked / what didn't. Also, if you'd be interested in learning more about this system, let me know & we will set up a webinar to break it down further.

Rillet raises $100M at a $1B valuation

Rillet continues to win against legacy players in the ERP space, showing that finance teams are increasingly comfortable adopting AI-native tools. ICONIQ's writeup on the Series C describes Rillet as a 'harness for agentic finance,' showing where Rillet/Campfire/DualEntry are heading.

Stripe's leaked investor letter predicts 'intelligence is the new capital'

Stripe announced its OpenRouter acquisition in a letter to investors and laid out where they think the economy is heading in a post-AGI world (which they think we're already in). Stripe sees intelligence as the next resource that all companies will need to manage in the future, and they want to capture a small slice just like they did with payments. Pairs well with my previous AI Chimera posts.

Every's Thesis Statements

AI consulting firm / media company Every just launched an essay series called Thesis Statements, asking founders to make predictions about the post-AI future of work. Highly recommend going through for a range of perspectives on what work will look like in 2027 & beyond.

That's all for this edition of The AI CFO - we'll be back in your inbox soon with more musings on the future of AI-powered finance & operations. Please reply to this email directly with any feedback/suggestions/just to say hi!

Cheers,

Alex Wittenberg, CEO @ airCFO