How to rebuild finance for AI - the CFO operating model
Most finance teams have already bought the AI. Far fewer can show what it paid back. The agents are in the close and the AP queue, and the board still asks where the return went.
How CFOs sequence the move to an AI-native finance function so the money reaches the P&L.
Short answer: Rebuilding finance for AI means sequencing three layers in order: a clean data foundation, then automation of routine work, then a layer of intelligence that models and decides. Skip a layer and the program stalls. Among finance organizations that have fully deployed AI, only 21% report clear, measurable ROI (Deloitte, 2026).
The order of the build predicts the return more reliably than the tooling does, and every credible 2026 framework converges on the same sequence.
Why do most AI investments in finance fail to show ROI?
Most finance AI programs miss on ROI because they automate on top of messy data instead of fixing the foundation first.
The numbers are stark. Deloitte's Finance Trends 2026 surveyed 1,326 finance leaders at companies above $1 billion in revenue across 23 countries in spring 2025. Among the 63% whose organizations have fully deployed AI, just 21% report clear, measurable ROI. Only 14% have fully integrated AI agents into finance. The tools are everywhere. The return is rare.
Bain's April 2026 survey of more than 100 CFOs points at why. Overall, 31% rate their AI outcomes as strongly positive. Among the CFOs who pushed AI all the way into full production, that jumps to 41%. Among those still stuck in pilots, it drops to 25%. How far the work got past the pilot stage and into production moves the outcome most.
What is an AI-native finance function, actually?
An AI-native finance function runs its routine processes on agents and moves its people onto the judgment. Bolting AI tools onto the existing org chart speeds up individual tasks and leaves the operating model exactly where it was.
The clearest built example is recent. In May 2026, PwC and OpenAI built a native finance function with agents across the close, accruals, forecasting, procurement, payments, tax, and treasury. The finance professionals stopped executing those processes. Their job became supervising and improving the agents that run them.
BCG describes the same shift in the CFO seat. As routine work falls to agents, the CFO moves from producing the numbers to architecting value. The finance functions leading this aim at a close that runs in real time and scenario models the team can rerun on demand.
Where should a CFO start rebuilding finance for AI?
Start with the data foundation. Every credible 2026 framework puts it ahead of automation and well ahead of agents, and the organizations that skip it are the ones that stall.
BCG's AI-first finance work lays out three layers built in strict order. A data foundation comes first: financial data that is clean, connected, and governed. Automation comes second, once the data can be trusted. An intelligence layer comes third, where agents model scenarios and surface anomalies. Build a higher layer on a weak one and it collapses.
KPMG frames the first move the same way: make AI the orchestration layer across finance workflows. It automates the routine work and routes each exception to the person who has the context to handle it.
Here is the sequence, and what each layer buys you.
| Layer | What you build | Why it comes here |
|---|---|---|
| 1. Data foundation | Connected, well-governed financial data | Automate on messy data and the agents scale your errors |
| 2. Automation | Reconciliations, AP, close mechanics, variance narratives | Trusted data makes routine work safe to hand off |
| 3. Intelligence | Forecasting, scenario modeling, anomaly detection | Agents decide inside the parameters; people own the exceptions |
The temptation is to buy an agent and point it at the forecast. That is layer three on top of a foundation nobody built. It is also a fair description of the programs sitting in the 21% that show no ROI.
Following the sequence? Get the next piece in your inbox. We write about making finance, operations, and sales AI-native, without the hype and without the tutorials.
What happens to the finance team?
The finance headcount pyramid inverts. For decades finance was a pyramid: a wide base of analysts preparing numbers for a thin layer of people who decided anything with them. As agents take the preparation, the shape turns into a diamond, widest in the middle where people interpret, advise, and govern the agents.
That matches what the built functions report. In the PwC and OpenAI model, the people moved up a level, from running the process to running the agents that run the process. The entry analyst role and the FP&A seat change the most. We mapped how those roles get restaffed across the whole company in a separate piece on the AI-native team.
The mistake here is treating the freed capacity as a headcount cut and banking it. Redeploy it into forecasting and business partnering, and finance turns into an advisory function. Cut it, and you own a cheaper version of the finance team you already had.
How much of finance can you automate today?
Less than the vendor decks imply, and finance is behind the other functions.
CFO Connect's State of AI in Finance 2026 finds 56% of finance leaders now use AI, double the 2023 rate. Adoption is climbing fast, but deployment trails well behind it. Finance ranks last among business functions on that measure: 45% remain in limited pilots, and only 17% use AI in core workflows.
The rule-bound work automates first. Reconciliations and the mechanics of the close are the safe early moves, and they map cleanly onto the routine workflows worth handing off. We wrote up which workflows to start with. The judgment work stays with people for now. Anyone selling a fully autonomous finance function today is running ahead of what the deployment data supports.
What the CFO actually decides
An AI-native finance function is a set of decisions the CFO owns. No vendor installs it for you. Four calls carry the budget.
- Is the data foundation ready? If financial data is fragmented across systems and lightly governed, that is the first project. Automating before it is fixed produces fast, confident errors.
- What order do you build in? Data, then automation, then intelligence. The sequence is the difference between the CFOs reporting strong outcomes and the ones reporting a pilot that never scaled.
- Where does the freed capacity go? Into advisory work, or into a spreadsheet showing the saving. Only one of those changes what finance is for.
- Who owns the agents and their failures? Every agent needs a named human accountable for the parameters it runs inside and the call when it gets one wrong. That accountability does not transfer to the vendor.
Answer those four and you have an operating model. Skip them and you have a tool subscription and a close that runs the same way it did last year.
Frequently asked questions
Will AI replace the finance team, or specific finance jobs? It replaces tasks and reshapes jobs. Agents take on the preparation: matching transactions, running payables, and drafting the first variance read. The people who did that work move up into advising the business and overseeing the agents that now do it. The base of the pyramid shrinks and the middle widens.
What is an AI-native finance function versus AI tools bolted on? An AI-native function runs its routine processes on agents and puts people on judgment and oversight. A bolted-on setup adds AI to the existing org chart, so individual tasks get faster while the division of labor stays fixed. The first changes the P&L. The second usually shows up in the 21% reporting no clear ROI.
Where should a CFO start? With the data foundation. Financial data that is trustworthy and wired together across systems comes before any automation and well before any agent. BCG and KPMG both put it first, and skipping it is the most common reason finance AI programs stall short of measurable return.
How much of finance can realistically be automated today? The rule-bound, high-volume work: processing payables, reconciling accounts, and closing the books each month. Only 17% of finance leaders currently use AI in core workflows, and 45% are still in limited pilots, so most functions are earlier than the marketing suggests. Judgment work stays with people for now.
Why do most AI investments in finance fail to show ROI? They automate on weak data and stay stuck in pilots. Among organizations that fully deployed AI, only 21% report clear ROI, and CFOs who scaled into production are far more likely to see strong outcomes than those still piloting. The gap is sequencing and scale. The model is rarely the problem.
We write about what it takes to run an AI-native finance function, with the operating decisions the exec actually has to make and no hype in between. Join the newsletter to get the next piece.
Sources
- Deloitte, "Finance Trends 2026 / CFO Guide to Tech Trends" (1,326 finance leaders, $1B+ revenue, 23 countries, spring 2025). deloitte.com
- Bain & Company, "CFOs Funded the AI Revolution. Now They're Joining It." (100+ CFOs, April 2026). bain.com
- PwC and OpenAI, "PwC and OpenAI Build a First-of-Its-Kind OpenAI Native Finance Function" (May 2026). pwc.com
- BCG, "The AI-First Finance Function" (2026). bcg.com
- BCG, "From Hindsight to Foresight: The CEO Mandate for an AI-First Chief Financial Officer" (2026). bcg.com
- KPMG, "The New Finance Operating Model: Four Moves to Make It Happen" (2026). kpmg.com
- CFO Connect, "State of AI in Finance 2026" (2026). cfoconnect.eu
By Christopher Kliebenstein. We build and run AI-native workflows for operators who want results, not demos.