At Digits, accounting isn’t just getting a makeover—it’s being reinvented from the ground up. The company’s modern accounting software promises streamlined financial operations for the more than 30 million small businesses across the United States, bringing the same level of real-time insights and automation to finance that other sectors have enjoyed for years.

Samir Naik, Head of Data and AI at Plaid, spoke with Digits’ CEO and Co-Founder, Jeff Seibert, about the company’s mission, the role Plaid plays in powering it, and why the future of AI-powered accounting is anything but boring.

Samir Naik: What inspired you to start Digits?

Jeff Seibert: At my previous start-ups, and then at Twitter, I saw firsthand how advanced product and engineering tools were—all real-time, intuitive, and easy to use. But when it came to finance, I’d wait weeks for black-and-white PDFs from our accountant. The entire industry was stuck. And the culprit was outdated software. Accountants are incredibly hardworking—but their tools hadn't evolved in decades.

In 2018, we started building the modern general ledger from scratch, which we soft launched last year by standing up our own accounting firm to test it with hundreds of businesses. We finally launched publicly last month on stage at HumanX in Las Vegas. It’s been a journey!

Naik: Congrats. That’s really exciting. Can you share how Plaid fits into your business model?

Seibert: We’ve been a Plaid customer for over six years now—it was the very first third-party provider we worked with, and we’re so appreciative. Plaid is an essential part of our flow, pulling transactions and balances, and providing us access to all the banking data we need to make Digits run for our customers.

Naik: Now speaking of smart technology, explain how Digits uses AI.

Seibert: We run 18 models in production, most of which are not LLMs. That’s because accounting is not generative. You don’t want an LLM hallucinating your books. So we’re very, very intentional about where we use LLMs and we actively prevent them from doing any math. Instead, we focus on task-specific models—like one that decides if a transaction is with a business or a person, which affects downstream categorization.

That said, LLMs are powerful for predicting things that aren’t yet in the dataset. But once we have it in the dataset and it's reviewed by an accountant, we push all that work down into our predictive models because they're extremely reliable.

Naik: What kind of impact has all this had for accountants?

Seibert: It’s been incredible. Digits auto-books 93% of transactions. From the perspective of an accounting firm, that's a major structural change that alleviates the need for the majority of bookkeeping labor. When we started, I was nervous about pitching automation to accountants. But today, the industry is welcoming it. CPA enrollments are down 33% over the past eight years, and 75% of CPAs are nearing retirement. Firms can’t hire enough junior talent, so they need tech to automate data entry to allow their teams to focus on client advisory work.

Naik: A lot of folks are talking about AI agents. What are your thoughts?

Seibert: We started using agents over a year ago with a specific use case: categorizing novel transactions in bookkeeping. If a small business makes a purchase from a new vendor, say a local restaurant no one's seen before, the fallback AI agent kicks in. It does what a human bookkeeper would do—it searches the web, gathers context, and builds a profile before handing it off to a language model to make a categorization decision.

One of the big wins is the absence of human bias. We had a case where a bookkeeper incorrectly tagged a transaction for “Stratton” as a ski resort expense because they were familiar with such an establishment in Vermont. Our agent, however, identified the correct vendor—a San Francisco-based business consultant named Stratton—just by doing research without prior assumptions.

Naik: Were you already doing this before LLMs came into the picture?

Seibert: No, we used to rely on a large outsourced team for that kind of work. Since rolling out agents, we’ve wound that down—it's all automated now.

Naik: Shifting gears a bit—security and privacy are top-of-mind. How do you approach these concerns?

Seibert: We have tons of sensitive, proprietary data. That’s why we avoid sending it to third-party models. Where we do have to use GPT, we redact business names, dollar amounts—everything—then sub it back in after. We're SOC 2 certified and leverage very few vendors in order to really control the data stack end to end.

Naik: Any final thoughts or advice for companies moving into the AI age?

Seibert: It reminds me of Twitter’s transition from web to mobile—disruptive, inevitable, and full of opportunity. This shift to AI is similar. It’s not just about plugging in new tools—it’s about reshaping your product, your team, and your thinking in order to take full advantage of what’s possible.