Accounting platform Digits announced the launch of Digits Accounting Agents, which CEO Jeff Seibert said can — when running on its Autonomous General Ledger product — automate 95% of the entire bookkeeping workflow in a way that, according to tests the company performed, outperforms even human professionals in terms of accuracy and speed.
Speaking during the Scaling New Heights conference in Orlando, Seibert heralded this development as true end-to-end automation with little to no need for human intervention or supervision beyond exception management and final approvals. The agents, he said, run 24/7 in the background, even “while you sleep,” to automatically do things like collect and reconcile bank statements; manage work papers; categorize, match and book transactions; book payroll; and clean up financial data, among other things.
“The problem is there’s just too much manual work, and you have to balance between way too many apps to complete it. So an AI-native workflow is flipped. It’s actually a new mindset. It’s not time to close the books. The books are closed, and your close checklist is already almost complete, ready for your review and approval. So you can dedicate your time to advise and budget compliance. This is now possible thanks to technology, thanks to autonomous agents,” he said during his presentation.
In a follow-up interview, Seibert reiterated that the agents are capable of doing the vast majority of the bookkeeping workflow with virtually no need for human supervision. While humans can examine what the agents are doing at any given time, and view the data-driven insights they surface, it’s not strictly necessary: The bots can run completely independently.
“It does 95% of it. You plug in your bank’s cards, payroll, almost everything is done. The things that the AI isn’t quite sure about, it surfaces for you in an inbox, and so you go in and categorize the remainder. Of course, if you’re doing advanced accruals, or project-based accounting or so on, you’ll still have to do those pieces, but we try to automate the 95% of the tedium that you’re just trying to do every month and save time,” he said.
Digits emphasized the accuracy of their AI outputs. Regular users of large language models may be familiar with a concept called “AI hallucination,” which is a more artful term for “making things up wholesale,” but Seibert said that Digits’ autonomous Accounting Agents “never hallucinate” due to the nature of the product itself.
He said that this is because the solution layers both LLMs and predictive models to do the work. The LLMs are not doing the calculations themselves, and in fact are specifically prevented from doing so. Instead it is the predictive models, trained on a massive amount of transaction data, that do the math. The role of the LLM is, instead, to orchestrate agent activity and communicate their results.
“LLMs are generative. They hallucinate. Predictive models don’t, they can’t. They predict things like: Where should transactions go? We layer the two of them. So we use LLMs to orchestrate our agents. The agents have access to tools which are all predictive. And so when you look at Digits’ technology stack, we run 18 different models in production. Almost all of those are custom-trained prediction models, and then we use LLMs to orchestrate,” he said.
He said during his presentation that the models are so accurate they not only outperform general models like ChatGPT and Claude by a significant degree, they also outperform human professionals. Digits pitted its agents against professionals from 12 outsourcing companies, with experience as CPAs ranging from a few years to over 30, to see who was faster and more accurate. He said that the humans were about 80% accurate versus 98% accuracy for the bots.
Meanwhile, in terms of speed, the humans took an average of 34 seconds per transaction, while the bots clocked in at 40 milliseconds per transaction. According to the Digits study, the transactions contained relatively fewer edge cases and incorporated a higher proportion of repeat transaction types compared to typical operational datasets. This composition contributed to an observable increase in accuracy rates across all evaluated systems, including both the Digits platform and the LLMs under assessment.
While he hesitated to call it a wholesale replacement for a professional accountant, he noted that it is a replacement in the case of the tedious tasks that accountants don’t like to do and that clients don’t necessarily value.
“We do want it to take over the really low-value work that you’re just honestly wasting time on every month, and the clients don’t appreciate [because they just] assume the bookkeeping is accurate. They don’t appreciate all the time that takes to make it happen. And so we are trying to take over the tedium while leaving the meaningful work for you to really focus on,” he said.
During his presentation, Seibert noted that the agents represent a third path that sidesteps the problems of both outsourcing and using LLMs.
“It’s so painful to search for accountants with outsourcing. They need a lot of guidance, they make mistakes, and they don’t know the details or history of the business they’re working on … . GPT is exactly the same. You have to tell it what to do. It hallucinates frequently, and it doesn’t know that business at all. It’s just trained on general accounting knowledge from the Internet. Digits is completely different. Our agents know the accounting workflows. They run them 24/7, they’re based on predictive models, so they can’t hallucinate, and they have secure access to historical context with each business,” he said.
When asked what’s next, he said Digits will be working on automating that remaining 5% of accounting tasks. This area represents use cases that are a little more tricky, such as splitting transactions, though he was confident in the future.
“So we are constantly teaching the agents new skills, and so over the next few years, expect Digits to get better and better and better at getting rid of this tedium for you,” he said.
