Yogi Goel pushes back against some of the trends in finance AI. Contrary to popular belief, finance leaders do not need to “rip and replace” existing systems with AI-powered tools, he says. Token maxxing is counterproductive. And despite the sensitivity of finance, there are ways to ensure the work is safe and accurate.
The co-founder, CEO and CFO of Maxima, an accounting software firm based in San Mateo, California, Goel spoke with CFO Leadership to share his insights into the best ways to use and build AI in finance.
Do CFOs really need to rip and replace their systems of record to bring AI on?
No. That’s a common narrative I’m seeing right now, and as an accountant for more than 20 years, I can confidently say that ripping and replacing an existing ERP is every accountant’s personal idea of torture.
There was never anything wrong with systems of record to begin with, and they exist for a reason. Their job is to store financial information in a structured, reliable way. They are deeply embedded in how a company runs. They tie into vendors, customer relationships, approval flows, compliance requirements. All of that logic is already encoded and enforced.
If you try to replace that just to adopt AI, you are effectively doing the equivalent of tearing down your house to install solar instead of just building a solar-proof roof. It’s far easier and far more practical to layer something on top.
That is why we are seeing a new system of work emerge that sits between systems of record and accounting teams. It pulls data from existing systems, like an ERP, then applies company-specific logic, and performs the work before looping it back to accounting teams to check for errors and confirm the work is accurate.
How do you get finance teams using AI without sacrificing security and auditability?
This goes back to the idea of ripping and replacing your ERP. If you’re asking an accountant to uproot everything they’ve known about how work gets done, all in the name of AI, it’s not going to happen. Finance teams are trained to be risk-aware, and they’re not going to adopt something that feels like a loss of control.
Instead, give them AI that supports their day-to-day work and actually lessens their workload. Don’t ask them to change how they operate overnight. Meet them where they are. And most importantly, it can’t be a black box. In finance, that just doesn’t work. It has to be a glass box.
They need to see exactly what the system is doing, what data it pulled, how it transformed it, what logic it applied and how it got to an output. Every step needs to be traceable.
In many cases, this also actually improves auditability. Today, a lot of work lives in spreadsheets, email threads or in someone’s head. With AI, you can standardize that process and create a clear proof of work every time.
When you’re not forcing an entire workflow change, adoption happens much faster. Teams can start with repeatable work, see the benefit immediately and build from there.
How should CFOs think about capital allocation in a GPU-first world?
The biggest mistake I see right now is treating AI spend as something you just need to maximize.
There’s this idea of “token maxxing” where you give every team as much access as possible and measure success by how much they use. That’s just adding a new form of headcount, except it can scale faster and get expensive quickly. The better way to think about it is closer to labor planning.
If you’re running a project, you estimate how many hours you’ll need, you price it accordingly and then you track actuals against that plan. Compute should be treated the same way. You estimate usage, tie it to specific workflows or projects, and then monitor whether it’s delivering the output you expected.
But you also have to be realistic about timing. You’re not going to get clean ROI on day one. This is closer to training your team. You invest upfront, you get people comfortable and then you start identifying repeatable workflows where AI can take meaningful time out of the process.
How do you actually quantify ROI on AI?
It really comes down to two things: Are you getting more output from the same team, and are you having to add people or not?
In most companies, especially in finance, the majority of the cost is labor. So if your team can do more work without you adding headcount or burning them out, that’s real ROI.
Where I think people get this wrong is they try to measure usage. They look at tokens used, licenses assigned, how many tools they rolled out. That’s just activity. For a while, the push was “adopt AI at all cost.” That’s not where we are anymore. People want to see actual output.
And the place it really shows up is in repeatable work. Every team has things that happen every month, every quarter. If something that used to take eight hours now takes one hour, and that keeps happening, it’s pretty obvious what changed.
Over time, it compounds. People move faster, they take on more and you realize you’re not hiring at the same pace even though the business is growing. That’s when you start to see the impact in a real way.
