Not subscribed? Sign up to get it in your inbox every week.

Claude can solve unsolved math problems, but it still can't reliably understand your financial model.

That sounds backwards. A three-statement model with a few dimensions and four years of monthly columns is not harder than frontier math. Yet anyone who has pointed an AI assistant at a big Excel file has watched it lose the plot.

Siqi Chen thinks he knows why, and his answer has nothing to do with how smart the model is.

Siqi is the CEO of CFO.ai, the company most of you knew until recently as Runway. He's a three-time founder (a games company acquired in his twenties, a startup acquired by Postmates, then CEO of Sandbox VR right before COVID hit) and was one of the earliest investors in Amplitude. Six years ago he started the company he wanted to build, then had to wait for the models to catch up.

This conversation has been edited and condensed for clarity, and some answers have been grouped under questions to keep it readable. The answers are Siqi's own words.

Chris: Who are you and what are you building?

Siqi: I'm the CEO of a company called CFO.ai. We used to be known as Runway, but we have a brand new product that is very much the product I was hoping to build six years ago when we founded the company. The technology wasn't quite there yet at the time, but now we have it.

The funny thing, and a lot of people don't know this: the entity that a16z invested in back in April 2020 was called CFO.ai, Inc. So this is actually our original name. It was nice to finally get a chance to use it to build the thing that does what it says in the name, which is an AI CFO.

Chris: How did Runway start?

Siqi: The genesis of this company was when COVID did hit and every VC was triaging their portfolio for who's going to get saved. We were making these Excel and Google Sheets scenario plans for how long COVID was going to last and how we were going to survive.

Long story short, we had to lay off 95% of the company where I was CEO. We went from 400 employees to 15. I laid myself off, and about 30 minutes after, I was talking to our now ex-CFO, and I said: all that stuff you're doing, the scenario planning, surely there was something better we could have used? Surely we can do better than this. And they said: no, you should try to build something.

Chris: Why didn't you build the AI CFO back then?

Siqi: Initially we wanted to build a product for smaller companies and founders. What we realized is that the technology to serve that market just didn't exist back in 2020.

If you think about the mentality of a founder, finance is just this black box. So you hire a fractional CFO, or if you're scaled up you hire a real CFO, and there's an intermediary, and you're hopeful that someone will just explain it to you. Software couldn't do that, at least in 2020.

So we realized: okay, we're going to have to wait this out. We're going to zig and wait for the technology to develop. What we built with Runway was, I would say, a nicer Anaplan. AngelList ran on it, AG1, Lambda.

Around 18 months ago, it really felt like the models were almost there. This was right when Claude Code first came out. And really, all of these capabilities had an inflection point around last Christmas.

Chris: A lot of finance people are already using Claude for Excel. Why isn't that enough?

Siqi: It's a great product. It makes Excel a lot easier. It'll write your formulas. But what every person we talk to who uses Claude for Excel finds is that when your model gets large enough, it is incapable of really understanding the context.

The thing we learned the hard way is that this is not an issue with the model. It's a super unintuitive point. Models can solve unsolved math problems. They can write very sophisticated code. How is a given financial model more complex than any of those things? You would expect it to just know how to do it. But when you actually use it, it doesn't quite do the job.

The reason is it's not an issue with reasoning capabilities. It's an issue with the format of the spreadsheet, which has been designed for humans.

Here's a really simple example. How would you model something like 10% month-over-month growth? You'd enter a million dollars in cell A1. In the cell next to it, you'd write A1 * 1.1. You copy that across, drag it over, and that's how you model it. What happens is that now every cell has a different formula.

So think about even a moderately complex model. Sheets has a 10 million cell limit. With just a few dimensions and four years of columns, you're going to hit that pretty quickly. And the problem with LLMs is there's a notion called a context window, and most LLMs have a 1 million token context window. So even if every cell was one token, it immediately floods the context window.

It's not an issue with math. It's not an issue with code. It's an issue with spreadsheets.

Chris: So what did you build instead?

Siqi: Look at what makes this different from a Harvey or a Legora. You have documents, they're in English, you can have a model read them, and you can do really useful work for lawyers. You can't do it for FP&A, for this reason.

There's no shortage of companies building agents for spreadsheets. For reporting, for getting a query done, it's fine. But when you're trying to understand and modify the context of a model, it doesn't do as well.

So instead of building an agent for spreadsheets, we had to build a spreadsheet for agents. And then on top of that is Ari, our CFO agent.

Chris: What does it actually feel like to use?

Siqi: I was on a call with our external general counsel and I said, "Hey, you should just try it." And he's like, "I have a finance team, and there's processes and procedures, I've got four spreadsheets, I don't know how to get started." I'm like, "Just try it."

Right afterwards, once Ari starts working, he's losing his mind. Ari's building his model in real time, just based on web research. I asked, does it make more sense now? And he said: once you experience it, you can't unexperience it. My mind is racing.

Chris: How does that compare to how finance software usually gets rolled out?

Siqi: The issue with every company in our space is the sales cycle is 30 to 60 days. You talk to a salesperson, you talk to a solutions architect, you have an RFP. Then when you onboard, it's: okay, work with our solutions team to write the SQL and shape the data in, and then from the data we're going to build a model, and then after we build the model, maybe we'll get some nice reports. Typically in our space it's a three to six month implementation. If you're a large enterprise customer, it can literally be a two-year implementation.

Everything I described, Ari will do automatically with one click. It will search your company. You connect your data. It will explore your schema, write the SQL, aggregate your data, and create a multi-dimensional model bespoke to your type of business.

Chris: It sounds like what Claude Code did for non-coders. Code was this scary thing until they tried it.

Siqi: Exactly. It's Claude Code for people who live in modeling and spreadsheets. But it also codes, which is the crazy thing. It'll make your board decks. It'll maintain your data rooms. You can give it a schedule and say, look at my variances. All of our charts are code. You can literally make Flappy Bird where the sides of the pipes are your margin every month.

We have a fractional CFO for a VC who, over a weekend, was modeling one of their portfolio companies. We got on a call, and he now wants to put all 50 portfolio companies he works with on it. We never saw this kind of pull with the last version of the product.

Chris: What's the use case that constantly blows people's minds?

Siqi: The number one thing is investor reporting. People who are raising a round want to maintain a data room and create a board deck.

But when you're building something, there's a level of value just from increasing the understanding you have of your business. This is the thing that enables that, in a way you're not going to get if you have to go through an intermediary or try to digest a model you didn't make that's indecipherable.

Chris: Data science used to work like that. You'd ask a question, someone would go off for a month and come back with an answer. If you can just ask, you end up with ten more questions.

Siqi: Totally. I was one of the first investors in a company called Amplitude. They do product analytics.

The really underrated thing about a product that makes analytics faster is not about saving time. Of course it does that. The more subtle thing is that when analytics are more widely available, accessible, and faster, it changes the nature of the questions that are worth asking.

If you have a question about your business and you know it's going to take two or three days and 500 bucks to get an answer, the questions you would ask are very, very different than if you knew you could get it right now for two cents. And that's fundamentally what we've done here.

Chris: Any closing thoughts for readers?

Siqi: If you're in finance, I think you kind of understand it. If you aren't in finance, the default thought is: this is boring, this is hard, this is complicated. All I can say is it's free. Go try it. It'll take you 10 minutes, and you can find out for yourself.