Why AI fails in finance ft. Didi Gurfinkel, CEO of Datarails
I asked Didi Gurfinkel a simple question. Three years from now, who's still on your finance team? He didn’t soften it. He splits finance into two kinds of people.
I just had a conversation with a CFOTech CEO of one of the fastest-growing companies in the US that made me rethink which finance jobs will survive the next three years and which ones are already gone.
So I spoke with Didi Gurfinkel, CEO of Datarails.
Didi sits in a rare seat.
He listens to hundreds of finance teams every month, and he sees exactly where they’re all heading. So when he told me half the work in his finance department won’t need a person soon, I was surprised.
I asked Didi, Three years from now, who’s still on your finance team?
He didn’t soften it.
I look at my finance department, and I can’t see why we need someone just processing a payment to a vendor. Or chasing a receivable. I literally can’t see the reason anymore.
Didi runs Datarails.
He works with hundreds of real CFOs and controllers telling him what they’re trying to do with AI and exactly where it breaks. Almost nobody on earth has that view. And he’s telling you… The people who move numbers from one box to another are done.
He splits finance into two kinds of people.
The ones who think, and the ones who simply follow processes.
The process-followers move the numbers. Execute the invoice, chase the bill, and key the entry. The thinkers decide what the number means.
The process-followers are going. All of them.
If that shocked you, good.
The conversation got heated.
Here’s everything Didi covered:
The last finance jobs
Why Didi told his customers to not use his own software
Why AI fails in finance
The role of Anthropic and OpenAI
The future of AI in finance Didi is betting on
If you’re a CFO, CEO, VP of Finance, or Controller/FP&A, this is your AI blueprint; you must read it to move forward.
Let’s dive in.
The last finance jobs
Didi splits every finance person into two kinds.
The ones who follow, and the ones who think.
The process-followers move a number from one place to another
The work is necessary today, but it’s entirely mechanical. There’s no judgment in it. You’re a pair of hands moving data across a gap that software is about to close.
Just to be clear, the process itself isn’t bad. It’s that the human simply following the process is the redundant part. The thinker isn’t a maverick; they’re someone who knows when the process is wrong.
Didi looks at that work and can’t find a reason it survives.
I don’t see why we need someone who takes care of payments to vendors. I literally can’t see why we need it.
The thinkers look at a number and ask why
By contrast, thinkers see COGS jump while revenue falls, and they know something’s wrong before anyone tells them. They decide what the number means and what to do about it.
That work doesn’t shrink when AI arrives. It explodes because, for the first time, one thinker can cover the ground that used to need five process-followers just to assemble the data.
Most finance people think their job security comes from knowing the tools. Knowing the ERP, being fast in Excel, and owning the close process. That’s backwards. The better you are at the mechanical work, the more replaceable you become, because that work is the first thing to go. The people who feel safest are the most exposed.
Why Didi told his customers to not use his own software
A few weeks ago Didi ran a customer advisory board in New York. Dozens of his own paying customers in one room, the kind of event where you expect customers to ask for more of what they already love.
He walked from table to table asking how Datarails could be better.
Every single request was about the screen.
Make the dashboard print better.
Add this metric to that view.
Move this column.
Customers who pay him real money, asking him to polish the window they look through. And he told them to think differently.
Take Claude and build any dashboard you want in minutes. It stays connected to the same data, the same permissions, the same source of truth. Why use our interface at all? No reason.
How unexpected. A software CEO, standing in front of the people who pay for his software, telling them the part they can see is the part that no longer matters.
He understands something that the customers in that room didn’t.
The dashboard was never the product. It was just the window.
The actual product, the thing worth paying for, was always the data foundation underneath: the consolidated, trusted, defined source of truth that the dashboard happened to display. For 30 years you needed the vendor’s window because you had no other way to look at the data.
Now you have Claude. You can ask the foundation anything, in plain English, and get a dashboard, a memo, a board deck, whatever you want, in next to no time. The window is anywhere and everywhere. Only the foundation is the hard part.
So Didi did something almost no software company has had the nerve to do. He restructured his entire company around making his own interface obsolete. He built new teams whose whole job is moving customers off his screens and onto Claude, with Datarails feeding the data underneath.
He calls it the finance operating system, or FinanceOS
He’s cannibalizing the visible part of his product on purpose, because he’s certain the hidden data layer is where all the value was hiding the whole time.
Look at every finance tool you pay for and ask which part is the window and which part is the foundation. You’re probably paying a fortune for window-dressing.
Why AI fails in finance
AI and finance are built on opposite values.
For a moment, walk through what finance actually demands:
Consistency: the same number every time.
Trust: you can rely on it.
Repeatability: run it multiple times and get the same result.
Auditability: trace every figure to its source.
Those four principles are the entire foundation of the profession. A number that changes when you ask twice isn’t a number; it’s a guess.
Now look at what generative AI is.
It’s probabilistic.
It doesn’t calculate the answer. It predicts the most likely answer and hands it to you, dressed up with the confidence of a veteran CFO. Ask it the same question twice and you can get two different answers, because generating a fresh response every time is literally how it works.
As Didi put it.
Everything finance admires, consistency, trust, repeatability, and auditability, is exactly what AI is not.
Didi has watched a lot of technology cycles, and he says he’s never seen anything like it.
There’s always a gap between how people talk about new technology and the reality. But here it’s two completely separate worlds.
The labs: AI research companies such as Anthropic, OpenAI, Google, and DeepMind are convinced that finance is the function AI changes most. The finance teams trying it are the most stuck. Those two facts are landing on the same desk, and it’s making CFOs feel frazzled.
If a chatbot drafts a contract and drops a clause, a lawyer reads it and catches the mistake. Finance has no such safety net.
If an AI runs through a hundred million transaction rows and hallucinates a one percent error, you will never find it. It’s buried. And one percent of a large company’s numbers is millions of dollars; walking into your board deck wearing a suit, looking exactly like every correct number around it.
So your pilot didn’t fail because AI is weak.
It failed because you aimed the most powerful tool in a generation at the one function that can’t survive a confident wrong answer, and you gave it nothing underneath to keep it honest.
It handed you something that looked brilliant, and you couldn’t stake your name on a single line of it. That’s not a failure of the tool.
That’s the predictable result of skipping the foundation.
The role of Anthropic and OpenAI
Anthropic shared which jobs AI will replace first.
Every CFO must see this chart. Blue is what AI can do to your function. Red is what it’s doing right now. The gap in finance is massive.
The biggest promise for AI and the worst progress in execution are seen within the CFO’s Office.
For a couple of years, the assumption was that you win AI by owning the computation. You run the models, you burn the tokens, you charge a premium on top. Datarails’ own early pricing worked like that: buy a block of tokens and pay more when you need more. That whole concept is collapsing.
As I told Didi on the call, all the tokens are moving to Claude. You don’t want to own the token business anymore. Let Anthropic and OpenAI burn the compute and fight over the models.
It’s a brutal, expensive, winner-take-most race, and it’s not your race.
So if the model isn’t where the value is for finance, where does it land?
In the layer that makes the model trustworthy for your business. The labs win the raw intelligence. Whoever builds the foundation that intelligence stands on wins finance. That’s the whole bet, and it’s why Didi stopped trying to out-feature everyone else and went all-in underneath the model instead of on top of it.
For you as a CFO, the takeaway is direct.
Don’t get into a bidding war over which model is smartest this month. They’ll all be brilliant and they’ll all be roughly equal. Spend your energy and your budget on the layer underneath, because that’s the part the labs will never build for you, and it’s the part that’s actually yours to own.
The future of AI in finance Didi is betting on
Now Didi zooms all the way out, and this is the part no CFO has heard yet.
He isn’t building a product. He’s building a category that doesn’t exist yet.
Call it the Snowflake for finance. A financial operating system. It’s a new category.
And he said something I haven’t stopped repeating.
Building a new category is like running a marathon, and if you look behind you and nobody is there, it means you’re running the wrong way. For months he was alone with this message. Now he’s hearing his competitors start to use his exact words:
Foundation.
Consistency.
The layer underneath.
To most founders that’s a threat. To him, it’s the first proof he picked the right race because finally there are runners behind him.
Today, only companies with genuinely complex reporting ever bought an FP&A system. Everyone else, the vast majority, limped along on spreadsheets and called it good enough. Didi thinks that era is over. Every finance team on earth will need this foundation, because you simply cannot run AI on top of nothing.
Everyone will have a financial operating system. You can’t run Claude on one MCP for Salesforce, one for your ERP, and one for your budget spreadsheet, and expect consistency. There is no question about it. That turns a few-hundred-million-dollar niche into a market ten times the size.
Didi has a rule at Datarails for measuring customer health.
The single biggest jump in value happens at one specific moment: when a customer stops feeding the system only finance data and starts feeding it operational data too. Inventory. Shopify orders. Sales pipeline. Headcount.
Why does that one move matter so much?
Because the second your foundation holds operational data next to financial data, FP&A stops being a finance function and becomes a company function. Now the FP&A person isn’t explaining last month’s variance to the CFO.
FP&A teams are catching a shift in inventory or a crack in the sales pipeline and walking it straight to the COO before anyone asks. Their reach jumps from the finance department to the whole business, all the way up to the CEO. In fact, Didi baked flexibility into Datarails’ architecture: FinanceOS uses a “relational database,” giving AI tools, such as Claude or ChatGPT, an unlimited environment with endless dimensions, columns, and rows.
Then I asked Didi the question that matters most to you.
What happens to the companies that try to get there without building this foundation?
Most finance teams will try to build without a foundation, but they won’t see the full cost of this decision until it’s too late. He laid out two paths. Look closely, because you’re about to pick one.
The first is the big-company path
You take your finance data and pour it into the corporate data lake, the Snowflake or Databricks or BigQuery your IT team already runs, right next to all the other company data. It works, sort of.
But here’s the trap nobody mentions.
Finance data is alive in a way other data isn’t.
The definitions shift, a new budget line appears, the mapping changes, and the model gets revised constantly. So you end up hiring a dedicated finance-IT person whose entire job is keeping the finance definitions aligned with the data lake.
In fact, hundreds of Datarails customers choose Datarails and Snowflake at the same time. Didi points out that FinanceOS isn’t competing with Snowflake or Databricks.
The FinanceOS category exists because finance teams have to update their model, adding a department, restructuring a P&L, changing an allocation, or locking a forecast version… without filing a JIRA ticket and waiting for the next sprint. In addition, FinanceOS is built around a long-running finance reality: teams work in Excel. FinanceOS fills a significant gap of things Excel is genuinely bad at, such as acting as a database and store of data.
FinanceOS keeps Excel connected to a governed system, while Snowflake forces finance to export data and rebuild models every month.
Didi says a data warehouse alone is the wrong place for finance models, because finance can’t own a model that lives in Snowflake.
You typically wait on IT to change a definition that used to be yours to control. Their structure often forces finance to export data and rebuild models. For a function whose entire reason to exist is owning the numbers, you just quietly handed away your independence. Most CFOs walk straight into this and only feel the loss years later.
The second is the duct-tape path
You wire Claude to a few MCP connectors yourself and make do.
You’ll pull a trial balance out of the ERP, sure. You’ll automate a little. It won’t be zero. But you’ll get a sliver of the potential, running the same manual consolidation you run today with a chatbot stapled to the side.
And here’s the specific thing that breaks: pulling a trial balance is easy, but a proper cash flow statement or a real invested-capital view… those take deep knowledge of how to build them correctly.
That knowledge has to be embedded once, into the foundation, so it runs the same way every month. Improvise it each time across a tangle of connectors, and you get your old chaos back, just with better autocomplete.
The future he’s building takes neither path.
The foundation is built once. The definitions are set once and held. And the AI on top finally behaves the way finance needs it to.
The Bottom Line
For two years the vendors sold you the tip of the iceberg.
The pithy insights, the pretty dashboards, the auto-generated board narrative. They skipped the foundation because building it is slow and unglamorous work that doesn’t close a sale in a thirty-minute demo.
But finance doesn’t run on impressive.
It runs on trust. And trust comes from exactly three things. One source of truth. A layer that holds your definitions. And a human who owns the result. Skip any of them and no model on earth will save you, because you’ll get an answer that’s right this week and wrong the next, and you’ll stake your name on it without knowing which week you’re in.
Everyone is about to have the same AI.
The same Claude, the same models, the same raw horsepower.
So the model cannot be your edge. The finance teams that win over the next three years won’t have a smarter AI than you. They’ll have done the boring, brutal work underneath it that you were tempted to skip.
The process-followers are going.
The thinkers are about to inherit the most valuable seat in the building, the one that sees the whole company, is right next to the CEO, and finally allows them to spend their time on judgment instead of data entry.
Make absolutely sure that’s the seat you’re aiming for.
The foundation takes time, and the clock started without you.
The hard part was never the numbers.
It was always the foundation.
And that’s all for today.
See you on Thursday!
Whenever you’re ready, there are 2 ways I can help you:
If you’re building an AI-powered CFO tech startup, I’d love to hear more and explore if it’s a fit for our investment portfolio.
I’m Wouter Born. A CFOTech investor, advisor, and founder of finstory.ai
Find me on LinkedIn