The vendor deck shows a finance function that forecasts itself in real time. Your reality is narrower. More than half of finance professionals say they use AI, but most of that is drafting and admin, not the core of the job. The sequence that actually returns money runs opposite to the demo, and getting it backwards is the fastest way to spend a year and a budget on a forecasting agent that sits on bad data.
Start where the volume is, not where the keynote is.
The adoption number behind the hype
Look past the headline. About 59 percent of finance functions now use AI somewhere, up from 37 percent in 2023, but adoption has plateaued and most use is still pilots rather than core finance workflows. The most-cited tool, in CFO Connect’s 2026 survey, is ChatGPT, at 35 percent. And the stack is already sprawling, assembled one point solution at a time rather than designed. Wide adoption, shallow penetration, growing tool count. Even though Deloitte finds 54 percent of CFOs at large companies call integrating AI agents a top priority for 2026, most of that use stays shallow. That is the real starting line, not the autonomous-finance slide.
The sequence that pays, and where it breaks
There is a consistent order to the deployments that reach ROI, and it is unglamorous on purpose.
Phase one, months one to three: accounts payable and reconciliation. High transaction volume, clear success metrics, minimal judgment, and a payoff you can measure inside a quarter. Goldman would call these the boring workflows, the junior grind AI is already absorbing, which is exactly why they pay fast.
Phase two, months three to six: the month-end close and variance commentary, which builds on the ERP integration you already did and cuts real days off the monthly close, with a Stanford and MIT study putting the reduction near 7.5 days. The mechanics are unglamorous: agents that match transactions, resolve discrepancies, and prepare audit logs, with some leaders expecting the concept of month-end itself to shrink as anomalies surface in real time rather than at period close.
Phase three, months six to twelve: FP&A and forecasting. The part the demo led with goes last, because it is the most judgment-heavy and the hardest to measure. This is where the framework has teeth. Autonomous agents in forecasting mean faster decisions and faster mistakes if the inputs are wrong, and a forecast built on data you have not yet cleaned in phases one and two is confident and wrong. The reason to automate forecasting last is not caution for its own sake. It is that the earlier phases build the clean data and the governance the forecasting layer depends on.
Cost the whole thing
Price more than the license. The real spend sits in ERP integration, the data cleanup that phase one forces, change management for a team that has done the close the same way for a decade, and the human review that catches what the agent gets wrong. Among the barriers finance leaders name are cost, integration complexity, and change management, alongside data security and a skills gap; model quality is rarely a leading concern. And vendor forecast-accuracy claims in the 20 to 40 percent range are worth treating as marketing until you have measured them against your own actuals.
What to do Monday
Pick the single highest-volume, lowest-judgment workflow you own, almost always AP or reconciliation, and start there, because it pays inside a quarter and builds the data foundation everything else needs. Put failure criteria on it before you start, the way a risk-aware finance team does any pilot. Measure the things that move: time to close, error and exception rates, hours returned, not a vendor’s forecast-accuracy slide. Only after phases one and two are clean should you let an agent near the forecast. The deck sold you the future in reverse. Build it in order, and the boring part funds the exciting one. Skip the boring part, and the exciting one runs on numbers you cannot trust.
The Counter Brief — one email, every Monday.
The week's AI-for-revenue moves in a 5-minute read: which tools are worth the budget and which to skip, plus what to do this week. Source-checked, no vendor decks.
Edited by Aditya Marin Gasga
Free. One click to unsubscribe.
Frequently asked questions
Where should a finance team start with AI?
With high-volume, low-judgment workflows, typically accounts payable and reconciliation, where success metrics are clear and the payoff lands inside a quarter. These deliver fast ROI and build the clean data foundation later phases require.
Why automate forecasting last?
Forecasting and FP&A are the most judgment-heavy and hardest to measure, and autonomous agents make faster mistakes when inputs are wrong. Automating it last means it runs on the cleaned data and governance built during AP, reconciliation, and close automation.
How widely is AI actually used in finance?
About 59 percent of finance functions report using AI somewhere, but adoption has plateaued and most use is still pilots rather than core workflows, per Gartner. The AI stack tends to grow one point solution at a time rather than by design.
What should I measure to judge finance AI?
Operational outcomes: time to close, error and exception rates, and hours returned, rather than vendor-quoted forecast-accuracy percentages. Validate any accuracy claims against your own historical actuals before trusting them.