As AI coding assistants become more powerful, a small but growing minority of people – including accounting and finance leaders – have turned to "vibe coding" to meet their software needs (
During a one-on-one interview, Jolly said there is a world of difference between what amateur vibe coders make and what a professional enterprise software company like Xero makes. People are certainly welcome to try and match their platform's functionality with AI coding tools, but she is confident they won't get far. This is because, she said, Xero is not just a single tool but what she referred to as a "financial operating system" that is more than its individual functions.
"You need the ability to not just have a system of record. You need to understand all the compliance rules in 13,000 jurisdictions, you need to translate those rules differently in different locations, and translating that is not easy," she said. She added that while certain rote mechanical compliance functions may be easy to code, many need complex judgment and extensive user testing that amateurs cannot bring to bear. "Otherwise, it's been, what, three years, four years? If it was that easy, you would have seen a lot more AI native accounting firms and AI native software companies."
She added that it is not just about data and compliance but also connectivity. Xero has built much of its strength on its third party connections, integrations and partnerships, something that requires more than just good coding.
"Think about connecting to the banks. Connecting to banks in the U.S. and getting accurate data is extremely hard. We've been at it for 10 years. It's still the number one thing that throws things off. Think about connecting to tax authorities. We're connected so we can do e-filing. Think about connecting to payment rails. We have bought a payment rail and we're connected to payment rails. So think about connecting to Shopify or Amazon. So building that web of connectivity in a way that the data doesn't leak, so you can understand the data, it's way harder than people think," she said.
Finally, while an individual vibe coder might get feedback from their colleagues and perhaps their customers on their apps, Xero has feedback from millions of customers that allow it to better fine tune its product for the future.
"We have tons and tons of evals, human evals. We have 5 million clients. That's not easy to replicate," she said.
So, according to Jolly, while people can – and even should – use AI to make custom applications to do things they never could before, this does not mean every single firm is going to build their own accounting systems. Yes, she has heard anecdotes of people using vibe coding to replace 30-40% of the capacity of enterprise applications, and of people saving money by doing so. She has also heard just as many anecdotes of people coding an agent to save $100 on a subscription app but then spending $500 to $1000 on tokens to fuel that agent (which does not even work as well.)
XeroForce
She pointed to tools like XeroForce, the company's recently-announced plain language AI agent builder, as a safe middle ground. People can code their own custom workflows for AI within the platform without having to go to the trouble of vibe coding an entirely new application themselves with all its attendant risks.
"An AI agent is software you write once and it performs a task over and over again. And since XeroForce is powered by JAX [Xero's proprietary AI system] it actually has accounting, payroll, payments, [and] compliance knowledge. You don't have to be a software engineer, you can just define your workflow in plain language and the Xero AI will write the agent for you. It is your rules, your judgment, baked in one and applied everywhere," she said during her presentation.
So for example, someone could make an agent that, on the second Tuesday of each month, fetches a client's P&L from Xero and drafts an email to them summarizing how their month is doing, illustrated with an attached spreadsheet.
XeroForce agents, she said, work across all tools in a practice such as spreadsheets, word processing, presentations, email, shared drives, digital e-signature platforms and more. During the interview she said these connections are partially from native integrations already on the platforms and partially from the agents creating their own connections. The agents themselves are unique programs versus sub-agents of JAX, which was described as the orchestration layer coordinating many different AIs, which include those created by XeroForce. Out of the box XeroForce – which is currently in early access and will be generally available later this year – will also come with pre-built agents for things like the month-end close, which can then be customized for unique workflows.
Tokens
AI consumes what are called "tokens," which can be thought of as an abstract unit of activity. While major AI platforms had at first been subsidizing this cost, increasingly companies are adopting
So, a simple bill with only one line item might be processed by an open source model they host on their own infrastructure, which incurs no cost at all. On the other hand, a complex bill with multiple line items that needs to be disaggregated between fees and sales might be processed with a much more advanced frontier model. In this way, the platform minimizes the costs by selecting the right tool for the right job.
"It's almost like a marketplace playing out," she said.
