Why the Future of Accounting Software Looks Nothing Like a Spreadsheet With John Glasgow

• John Glasgow spent 15 years in finance feeling the pain of outdated accounting software firsthand—then went through Y Combinator to build the modern ERP he had always wished for.

• Campfire is one of the only ERPs running its own foundational model, pairing it with Claude for natural-language queries—and pushing AI beyond simple reconciliations into subjective accounting work, such as accruals.

• His bet: the front end of enterprise software is about to disappear into a human interface, and the companies that own the richest, cleanest data will be the ones whose AI actually delivers.

Let's dive in. No floaties needed…

Revenge of the Nerds

John Glasgow, CEO & CFO of Campfire

He's not your typical founder. John came from a 15-year career in finance, including leading the $625M sale of Invoice2go to Bill.com. He knew the pain of legacy accounting software from the inside—as a customer, as a partner, and eventually as someone who decided to just build something better.

Campfire went through Y Combinator's Summer 2023 batch, has raised over $100M, and now serves companies like Replit, PostHog, and Decagon. The team has grown from 10 people a year ago to about 100 today. John still holds the CFO title because he runs Campfire's own environment—and if the product can't work for him, it's not ready for anyone else.

When the Company Started, It Was Manual Reconciliations Before Moving to AI. What Broke in That Process?

AI had a lot of hallucinations and errors when we first started putting it on the product a couple of years ago. It's gotten a lot better. We still always have a fallback where the AI can silently fail, and the manual reconciliation can still be performed.

Here's a very tactical example. On legacy systems, a human manually does the work—maybe there's some rule-based automation—and then at the end, the user uploads the PDF of the bank statement to confirm the bank feed was reconciled correctly. In Campfire, you upload the PDF at the beginning. AI takes a pass and says, here's the thousand transactions we could handle, and here's the two we couldn't. Instead of 1,002 that humans do with rules and manual work, they just look at two now, but they can still review the AI's work.

And as AI has gotten better, it's moved beyond just the objective work like reconciliations into more subjective territory. There's something called accruals in accounting, which is essentially a cost estimate—very human, very subjective. We've seen AI do an incredible job there. But the human doesn't trust AI by default, so we generate full supporting documentation of all the math. Kind of like in school—even if you get a calculator, you still have to show all your work on the math test. If the human can't see how AI got to the answer, they'll just redo it themselves, and then you're not adding any value.

How Do You Know When the Model Fails, & What Has the Error Rate Looked Like Over Time?

Two things have truly changed. Of course, the models themselves have gotten better. But we've also made material investments in how we work with the model. Previously, we'd just take a bunch of accounting data, feed it into the model with a prompt. Now, in many scenarios, we have the AI write code—whether it's SQL or Python—and the code actually performs the work on the financial data. AI is great at writing code, but it's not always as great at directly handling accounting data. So our methods of working with AI have materially evolved alongside model development. There have been two vectors of improvement.

On accuracy, a lot of our customers are now seeing 100% match rates. In the early days, it was around 60 to 70%.

We have two approaches. If it's a natural language query—a chat interface in our product—we go to Anthropic. They actually published a study on how we're reinventing accounting. But because we have such a rich financial data set, we saw an opportunity to have a more secure environment where we also run our own foundational model.

Here in SF, we've got an AI engineering team that runs our own model. Very specific accounting workflows go into that model. We've got our own GPUs; there's a lot of inference work behind it. And then for natural language queries—because those are so open-ended, you could literally ask anything in a chat interface—we go to Claude for those.

How Competitive Is It to Hire AI & ML Engineers in San Francisco Today?

It's quite competitive. This is where OpenAI is, where Anthropic is—a lot of the frontier model headquarters. But a lot of folks say they want to go to an early-stage startup where they can make a huge impact on a lean team. So we lean into that. You can be on the founding AI team.

One thing that gets engineers excited in interviews is our data. Because accountants are so meticulous about their work, it's an incredibly clean data set. Clean, labeled data is an engineer's dream. And it's an area the existing models haven't spent much time on—all the code is on the internet, all the legal databases are on the internet, but nobody's accounting data is on the internet. So there are areas where we've found people genuinely get excited. We pitch them on that, and we've done quite well hiring amazing people. The team is three people right now—lean, but we're actively hiring.

AI Is a Benefit for Campfire Today. How Long Does It Remain a Competitive Advantage, & When Does It Become Table Stakes?

AI is not homogeneous anymore. It's not about whether you have AI or not—it's about the specifics. We now have an agent platform. We're the only ERP with our own foundational model. Maybe everybody can slap an MCP on their product and connect it to OpenAI, but are you building your own model?

It's getting down to a finer point. For us, it's more about AI velocity than just having AI. We consistently see customers choose us because they say they got a sandbox of every solution on the market, and our AI performed the best. On paper, they all kind of had the same AI, but from a performance and workflow standpoint, we stood out.

New AI products are notorious for churn because they demo great, but often don't deliver in reality. So we try to get customers as much hands-on time with the product as possible before they sign—because we're aware things are moving quickly and people want to test it on their own data before they commit.

What Might Campfire Look Like in Three to Five Years? What Does the Future of Accounting Look Like?

I think a lot about when the iPhone came out. Everything was a flip phone with buttons, small screens, and essentially no app stores. Then out came something that was just a screen with a single button. There's a cannibalization moment happening right now—the drop-down menus, the 30 different pages of a product, a human having to go perform all of this work. That's starting to fade away.

Whether it's a chat interface or an agent platform, ultimately it's a human, AI, and then ones and zeros in a database behind it. It's not drop-down menus and configurability anymore. The key insight for me is that a lot of knowledge workers are moving from specialization—doing one specific task all day—to generalization, where you're managing a bunch of agents doing specialized tasks. That's where Campfire is going.

What Is Campfire's Right to Win Against Incumbents Who Are Adding AI on Top of Their Existing Platforms?

All the incumbents are adding AI, but throwing AI on top of a very old, outdated database that can't scale well with large volumes of data is a problem. Our incumbents struggle with data volume, so their customers end up summarizing data into the solution. Adding AI on top of summarized data is actually pretty bad AI—it doesn't have good visibility into the underlying information.

Great data is the foundation for great AI. I spend a lot of time ensuring our product is incredibly scalable, with great APIs for both internal and customer consumption. When you have access to a rich data set, AI can do amazing things with it. A lot of our customers are still doing work in spreadsheets around their accounting software. My big push to them is: whatever is still in a spreadsheet, we need to get into software, because then we can unlock AI onto that offline data.

And there's an important nuance around context. There are a lot of corner cases and decisions that get made during the day. Why were exceptions made? If that context is documented in the system, the AI can understand how to handle it next time. Otherwise, it won't be able to handle the exception without a human stepping in.

Tell Me About Internal AI Tooling. How Are You Thinking About AI at the Organizational Level?

Everyone talks about AI on the product side. Fewer talk about it internally. I held an all-hands and said everybody needs to be using AI and attempting to do every single manual task with AI before they do it manually.

Everyone gets a Claude enterprise license. I'm comfortable with people putting our own company data in there—customer data is a totally different topic, but company data, even financial data, I'm okay taking a little risk. We launched a Slack channel called Show & Tell, where people share what they've built. Literally every day, every team is posting new apps, widgets, and automations.

I even built a little app that runs locally on my laptop—never leaves my machine, all the data stays on it. It's not even on the internet. So we're using homegrown apps for just about everything within Campfire. Of course, the core workflows still have their tools—HubSpot, GitHub, our own product for accounting. But any nook and cranny that's a manual task or an integration that doesn't exist between two systems, we just build it ourselves.

Quick Poll

Additional Reads

• Campfire: The AI Native ERP for the fastest growing companies. | Y Combinator

• Campfire accelerates accounting with Claude

• Tiny AI ERP startup Campfire is winning so many startups from NetSuite; Accel led a $35M Series A

Rate This Edition