Vibe-coding an accounting tool with AI

Some AI proponents claim that soon, pretty much everyone will be able to vibe-code their own tools and agents. But many companies have run into roadblocks attempting to implement or scale AI: messy data, rising costs, and reluctant staff among them. Skeptics question whether AI has much value, or even whether its ROI can be measured at all.

But what some accountants are achieving with AI might have SaaS companies looking anxiously over their shoulders. Nathaniel Dycus, VP and global controller at Estonian CRM software company Pipedrive, built an accounting close tool in a matter of months using Gemini—without any programming experience. Dycus’s vibe-coded app has replaced the finance team’s commercial close management tool, his CFO, Regi Vengalil, told CFO Brew.

A messy manual close. Pipedrive, a 16-year-old company targeting small and medium-sized businesses, reached “unicorn” status in 2020 with a $1.5 billion valuation. It has more than 800 employees.

That global reach caused headaches for the accounting team, which had members spread across time zones in the US, Estonia, the UK, Ireland, Portugal, and other countries, Dycus said. Sometimes a team member would mark close-related tasks as completed or reviewed on different days than the rest of the team. “We had a struggle with collaboration on month-end close and how to look at it as one consolidated process,” he said.

Plus, the close process involved many manual steps that Dycus thought could be automated, such as copy-pasting data from NetSuite into Google Sheets, marking when tasks were completed or reviewed, and updating the trial balance when changes needed to be made.

The team did look into commercially sold close software as a potential solution, but found that most of it was too robust—they didn’t need “the entire start-to-finish close process,” Dycus said. He turned to Gemini to see if AI could help solve the problem.

At the time, Dycus said, his AI skills were limited to prompting and asking chatbots to recommend Excel formulas. His work experience, which included 11 years with PwC and four years as SVP of accounting at HireVue, was all accounting-related. “He had never coded before,” Vengalil said. “I don’t know the number of times I asked him, ‘Are you sure there wasn’t a minor in college?’ No, he just started with a blank canvas.”

Built piece by piece. The close tool started out as a simple sidebar added to Google Sheets, which allowed the team to mark items as completed or reviewed without having to scroll horizontally past many columns. Over time, Dycus added more features and made the sidebar larger. It evolved into an app on top of Google Sheets and then a web app. “These are more or less iterative prompts that did one thing, and then slightly more, and slightly more, and then you end up seeing a much different world available to you,” Dycus said.

Now, the app shows team members a timeline of each item with its status, and allows them to set due dates and to see what’s completed and what’s past due, and it emails them when someone assigns them a task. They can “see how things are getting prepared, how things are getting completed as we move to the cycle relative to their due dates,” Dycus said, “and that helps us to not only monitor performance in close, but helps us to rewind after close and see where our bottlenecks were.” 

Dycus also added a visualization component to the app to make it easier for the team to see the status of the close process. He used Gemini to build a close matrix and added a charting feature that creates visualizations of data, such as balances at different points in time or compared with previous months. Instead of looking at “500 rows, we get a sense of how close is going respective to our deadlines,” Dycus said.

Real ROI. Vengalil is delighted with the close tool, which he says the team now uses instead of the close management tool that came with NetSuite.

“Does it match the functionality 100%? No,” Vengalil said. “It is confined to our needs and our use cases. It’s not apples to apples, but that’s a good example of a place where we might be spending a couple of thousand a year in terms of inference [the cost of running a trained AI model] and automation costs relative to a cost base of something like $30,000 to $40,000.”

The close tool has other advantages as well: It’s freed up time for the finance team, which has allowed them to do “more rigorous reviews,” Vengalil said. It’s also expanded “the range of transactions that can be validated. It is allowing us to look at processes that we haven’t been able to spend time on.”

A team effort. Dycus built the close tool over several months, with plenty of feedback from his team along the way, he said. The original Google Sheets sidebar only took him a couple of hours to put together, while the first version of the app that the team was able to use for a close took “probably three or so weeks,” he said. Over the next few months, he refined it and added more features. The team also “demoed the tool and tried to make the tool break,” and they all learned about the “pain involved in the build phase” as opposed to the relative seamlessness of buying something already developed.

What helped, Dycus said, was knowing he had a safe place to experiment in. The app was built on Google Sheets, and the team wasn’t tinkering with anything that could harm the business. “We weren’t pushing journal entries into NetSuite, we weren’t doing anything that wasn’t resolvable. Anything we did was visible on the screen to us,” he said. “Worst-case scenario, we were back to the same Google Sheet we had.”

Zoom out. In the future, Vengalil believes, finance staff will need to become more technologically adept. “The expectation is, and it’s scary to a lot of people, that we all have an engineering mindset to how we’re approaching things,” he said. Functions such as FP&A and treasury, he said, may need to learn more about data engineering and become knowledgeable about “the architecture of how your data interacts with the data warehouse.”

Companies can encourage experimentation with AI. “For me the biggest problem is the ‘cold start problem’” or how to turn a blank screen into an output, Vengalil said. Pipedrive holds trainings, demos of staff-created AI tools, and “build days” when employees start working on business problems in Gemini. “Many of these are efforts to show people that it’s not as hard as they think it might be,” he said.

As for Dycus and his team, they’re starting to work on more advanced AI tools. They’ve connected to a NetSuite secure testing environment to trial them, he noted.