KPMG’s 2026 Global AI in Finance survey found active use of AI in the finance function has more than doubled since 2024, rising to 75%. But adoption and outcomes are two different things. In my own working conversations with middle-market and PE-backed CFOs over the last 18 months, only about 12% could name a specific, measurable business outcome AI had produced in their finance function. The rest offered variations on “improved productivity” or “faster analysis.”

That gap – near-universal adoption, one-in-eight naming a real outcome – is the actual state of AI in finance right now. Not the vendor decks. Not the LinkedIn posts. The gap.

I have been in and out of AI implementations at nine middle-market and PE-backed companies in the last 18 months. I have a working theory on what the 12% do differently. This post is the theory.

What the 88% do

The typical CFO deployment pattern I see:

1. Buy ChatGPT Enterprise or Claude for Business, roll out to the finance team.
2. Send an email announcing the deployment.
3. Do a lunch-and-learn on prompt engineering.
4. Watch the team use it as a slightly better search engine for two months.
5. Conclude that AI is “overhyped” for finance.

The tool sits in a browser tab. It generates draft emails, summarizes meeting notes, sometimes reformats a table. It never touches the general ledger. It never runs a forecast. It never produces anything that gets consumed by a decision-maker outside finance.

This is the failure mode. Not the tool. The deployment shape.

What the 12% do

The pattern I see in the 12% is different in three specific ways.

1. They ship a workflow, not a tool

The 12% do not “roll out AI.” They ship a specific workflow with a defined input, a defined output, and a defined human in the loop. Example: “Every Friday, the FP&A analyst runs the 5-prompt weekly financial review, and the CFO reviews the output at Monday’s operating meeting.” That is a workflow.

Compare: “Everyone has access to Claude.” That is a tool.

The workflow has a chain of custody. Someone owns the input quality. Someone owns the output review. Someone consumes the result. If you cannot name those three people for a specific AI use case in your org, you do not have a workflow. You have a tool.

Read The 5-Prompt Weekly Financial Review for the shape of one.

2. They rebuild the source data first

The 12% do the boring work first. I noted in How to Structure Your Chart of Accounts for AI-Powered FP&A that recent CFO surveys from McKinsey and KPMG have consistently identified poor data structure and data trust as the leading blocker to scaling AI in finance. The 12% see this and fix the chart, the sub-ledger tagging, the cost center hierarchy, before they buy the tool.

The 88% see this and buy Cube or Datarails hoping the tool will paper over the data mess. It will not. Datarails is a good product. It is not a data cleanup service.

The order matters. Data first. Workflow second. Tool third. Every case study I have seen where AI actually moved a metric follows that order. Every case study I have seen where AI became shelfware inverted it.

3. They pick one metric and prove ROI on it

The 12% pick a single measurable outcome and build the AI use case to move that outcome. Examples I have seen work:

• Days sales outstanding, down from 52 to 39 in six months via an AI-run weekly collections triage.
• Board deck prep time, down from 22 person-hours to 4 via a templated prompt chain (see The Board Deck AI Prompt That Cuts Prep From 8 Hours to 45 Minutes).
• Monthly close, down 1.5 days via automated flux commentary and journal entry review.
• Forecast accuracy, root mean square error down 22% on the top-3 P&L lines via better driver mapping.

Notice what these have in common: a metric that existed before AI, a delta you can measure, and a specific person who owns both the metric and the workflow.

The 88% pattern: “We rolled out AI and productivity went up.” How measured? “People say they save time.” That is not an ROI story. That is a survey.

What the 12% look like in practice

The CFO of a $340M PE-backed distribution business I work with runs the following stack:

• Claude for Business, project-based, one project per workflow.
• The chart of accounts got rebuilt in Q4 2025 before any AI work started.
• Six named workflows in production: weekly review, monthly flux, board pack, board Q&A prep, budget season narrative, and 13-week cash update.
• Each workflow has a named owner and a documented prompt (versioned in a shared doc).
• The CFO reviews AI-generated outputs but does not run the prompts. That is the analyst’s job.

Total finance headcount: 11. Same as before AI. Total time reclaimed per week: about 35 hours. Reallocated to strategic finance work, not headcount reduction.

The stack is not fancy. There is no agentic autonomous system. There is no proprietary model. There is a chart of accounts that works, six prompts that work, and a controller who cares.

The counter-argument

“Some functions genuinely benefit from generic AI without workflow discipline. Drafting, editing, brainstorming.”

True. If your team is using AI as a writing tool and your writing has gotten better and faster, that is a win. Do not knock it. Just do not call it AI transformation of the finance function. It is a spell-check upgrade.

The second objection: “Autonomous agents will change this. The workflows will build themselves.” Maybe. Not yet. Gartner’s 2026 Hype Cycle for Agentic AI places agentic AI squarely at the Peak of Inflated Expectations, and Gartner’s 2026 CIO survey found only 17% of organizations have actually deployed AI agents to date. Anyone telling you they have deployed autonomous agents in finance in 2026 is doing a demo, not production.

The third objection: “The 12% is a self-selection bias. Those companies were going to succeed anyway.” Fair. But the pattern is consistent enough across sizes, geographies, and industries that I do not think it is only selection. The workflow-first, data-first, metric-first pattern shows up in first-time CFOs and 30-year veterans, in $80M companies and $2B companies. It is a shape you can copy.

The read

If you are in the 88%, the fix is not “try harder.” The fix is “start over with a workflow.” Pick one metric. Pick one workflow that would move that metric. Fix the data feeding that workflow. Then buy the tool.

The specific playbook for what workflows to build first is in The 2026 AI CFO Benchmark, which cuts the 19,435 EDGAR 10-K analysis into what real companies actually deployed vs. what they filed.

The AI-Native CFO Prompt Pack Pro has the 30 workflows I have seen actually work, with the prompts and the data-prep checklists for each.

Note: Company details in this piece have been anonymized. Any figures drawn from Spencer’s advisory work with middle-market and PE-backed finance teams are directional.