How to measure AI ROI - a practical three-level framework
Task-level dashboards look great and predict almost nothing about business impact. Here is the measurement stack that does.
Short answer: Most AI measurement stops at the task level - hours saved, outputs produced - because that data is cheap and always looks good. It does not predict business impact. The fix is a three-level model: task metrics for adoption, workflow metrics like cycle time and error rate, and business-outcome metrics like EBIT attribution. Only the third answers the board's question.
We rebuild AI measurement for operators heading into a budget review. The reporting almost always has to move up two levels before it survives the first board question.
Why do task-level AI metrics fool executives?
Task metrics are cheap to collect and flattering to read. That is exactly why they mislead.
Adoption rate, messages drafted, hours saved: this data is easy because the tools emit it for free, and it climbs because people do use the tools. None of it tells you whether the business is worth more than it was last quarter. The gap between those two things is now measurable. Deloitte's Q4 2025 State of Generative AI survey, covering 3,235 leaders across 24 countries, found that 74% say their most advanced gen AI initiative is meeting or exceeding ROI expectations. In the same survey, only 20% are actually growing revenue from AI. Another 74% say they hope to.
So most leaders report a win their P&L has not seen yet. The board-level view is bleaker. IBM's 2025 CEO study, a survey of roughly 2,000 chief executives across 33 countries, found that only 25% of AI initiatives have delivered the ROI executives expected, and just 16% have scaled enterprise-wide. The confidence sits at the task level. The disappointment sits at the outcome level. They are measuring the first and getting graded on the second.
What is the board actually asking when it asks about AI ROI?
The board asks one question in different words: what did this return, in money or in better decisions? Adoption data cannot answer it, and directors know that.
The reason the question lands so hard is that the base rate is poor. BCG's 2025 "Build for the Future" study, which surveyed more than 1,250 firms across nine industries, found that only 5% of companies are achieving AI value at scale. Sixty percent report minimal or no material gains in revenue or cost. A director who has read one of these reports is not going to accept a chart of messages sent.
The pressure is about to get sharper for anyone running agents. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear business value, drawn from a poll of 3,412 webinar attendees. When budgets tighten, the projects that survive are the ones that can show a line to the P&L, and the ones defended with usage stats are first to go.
What is the three-level AI ROI framework?
Measure AI at three levels, and treat a number as real only once it closes the loop back to the P&L. Each level answers a different question, and most dashboards never leave the first.
Level 1: task metrics. Adoption rate, messages drafted, hours saved, outputs produced. These answer one question: are people using the tool? Worth tracking, because a tool nobody uses returns nothing. On its own, though, it measures activity and stops well short of value.
Level 2: workflow metrics. Cycle time, error rate, throughput, cost per unit of work, rework rate. These answer a harder question: did the work itself get better? A contract that used to take six days now takes two. A support queue clears with a third fewer escalations. The workflow layer is where the task-level activity either becomes a real operational change or disappears.
Level 3: business-outcome metrics. Revenue per AI-enabled employee, EBIT attribution, and what we call the decision-quality delta - whether the calls the business makes are measurably better with AI in the loop. These are the framework's own terms; you will not find them as standard line items in a vendor report, and they are the ones a board recognizes. A number reaches Level 3 only when you can trace it back to cost or revenue.
The bridge between the levels is workflow redesign, and the evidence for that is specific. McKinsey's State of AI, March 2025 tested 25 organizational attributes against whether a company sees EBIT impact from gen AI, and found workflow redesign has the single largest effect of any of them. The same study found only 39% of organizations report EBIT impact at the enterprise level, and among about 2,000 respondents, only around 5.5% attribute more than 5% of EBIT to AI. Initiatives pile up Level 1 numbers easily and rarely produce a Level 3 one, and the gap between them is Level 2, where most programs quietly stall.
What metrics belong at each level?
The table below is the whole framework in one view. Report all three columns, and make the arrows explicit: a Level 1 number should point at a Level 2 number, which should point at a Level 3 number.
| Level | The question it answers | Example metrics | Who cares |
|---|---|---|---|
| 1. Task | Are people using it? | Adoption rate, hours saved, outputs produced, messages drafted | The team, the tool vendor |
| 2. Workflow | Did the work get better? | Cycle time, error rate, rework rate, throughput, cost per unit of work | The function head |
| 3. Business outcome | What did it return? | Revenue per AI-enabled employee, EBIT attribution, decision-quality delta | The board, the CFO |
The rule is simple, and rarely followed: no Level 1 metric goes in a board deck without the Level 3 metric it is supposed to move. If you cannot draw the line, you have found a program that is producing activity and no return, and the budget review is the wrong time to find that out.
Heading into a budget cycle? Get the next piece by email - short, evidence-led writing for the exec who has to defend AI spend.
How is this different from McKinsey's measurement stack?
This three-level model is deliberately simpler than the prior art. McKinsey has published its own multi-layer measurement stack that runs from technical and adoption metrics up through engagement to bottom-line financial results, and argues that no deployment should scale without a traceable line from usage to P&L.
The underlying principle is the same: usage has to connect to money before it counts. The difference is what an operator can hold in their head during a live board discussion, where a stack with several intermediate layers gets hard to keep straight under questioning. Three levels stay simple enough to defend on the spot. If the fuller version fits your governance, use it. If you need something you can run tomorrow, three levels is the working minimum.
How long before AI ROI actually shows up?
Plan for two to four years to a satisfactory return on a typical AI use case, and expect the early numbers to disappoint against the pitch. Measuring at three levels is what keeps a slow-burn investment fundable through the years before Level 3 turns.
The agentic numbers make the timing concrete. Deloitte's 2025 "AI ROI" study, based on 1,854 executives across Europe and the Middle East, found that among firms already using agentic AI, only 10% report significant ROI from it. Only 6% saw payback within a year, against a typical expectation of 7 to 12 months. Most report that a satisfactory return on a typical AI use case takes two to four years.
Read that against the earlier confidence figures and the three-level model earns its keep. A program that shows nothing at Level 3 in year one is normal. Without the workflow layer in between, you have no way to confirm the program is actually on track. Level 2 is your early warning. Cycle time and error rate move well before EBIT does.
What do you put in front of your board?
One page, three columns, and an honest read of where each initiative sits. The framework is the argument; these are the moves.
- Lead with the Level 3 number, even if it is small or still zero. A named EBIT figure with a two-year horizon beats a dashboard of adoption stats. A modest, traceable number holds up better under questioning than an impressive one you cannot defend.
- Show the Level 2 evidence that it is coming. Cycle time down, error rate down, cost per unit down. That downward trend makes a zero at Level 3 credible instead of alarming.
- Name the workflow redesign behind each gain. McKinsey's data says this is the attribute that separates the companies seeing EBIT impact from the ones that are not. If a gain has no redesign behind it, it will not last.
- State the timeline out loud. Two to four years, per Deloitte. Set that expectation before someone else sets a twelve-month one for you.
- Kill the initiatives that only ever produce Level 1 numbers. They are the ones Gartner expects to be canceled anyway. Cut them yourself, on your terms, and move the budget.
Do that, and the budget review becomes a status update on a return you have already mapped.
Frequently asked questions
What is AI ROI and how is it actually calculated? AI ROI is the business return on an AI investment, measured against its full cost. Calculate it at three levels: task usage, workflow improvement (cycle time, error rate), and business outcome (revenue per AI-enabled employee, EBIT attribution). Only the third is true ROI; the first two are leading indicators that the return is on its way.
Why do most companies fail to show ROI on their AI investment? They measure adoption instead of outcome. Usage data is cheap and always positive, so it fills the dashboard while the P&L stays flat. McKinsey's 2025 data points to the cause: companies buy tools without redesigning the workflow, and workflow redesign is the attribute most tied to EBIT impact.
What percentage of companies are seeing real ROI from AI? Few. BCG found only 5% achieving AI value at scale, with 60% seeing minimal gains. IBM's 2025 CEO study found only 25% of initiatives delivered expected ROI. McKinsey found only about 5.5% of firms attribute more than 5% of EBIT to AI.
How long does it realistically take to see ROI from an AI investment? Two to four years for a satisfactory return on a typical use case, per Deloitte. Only 6% of firms saw payback within a year, against a common 7-to-12-month expectation. Budget and govern for the longer horizon, and watch workflow metrics as the early signal.
Is agentic AI actually delivering ROI yet? Rarely, so far. Among firms already running agents, Deloitte found only 10% report significant ROI. Gartner expects more than 40% of agentic projects canceled by 2027 on cost and unclear value. The three-level framework is how you tell an early-stage agent program from a failing one.
Get the free human-writing skill
The one AI output every exec still reads word for word is the writing. We packaged our method for making AI drafts sound human - the sentence-level fix table, the rhythm rule, and a 60-second check - as a free Claude skill. Add your email and we will send the Notion doc.
Want the writing without the download? Just the newsletter - we write about running an AI-native organization, for the person who funds it.
Related reading
- Why AI pilots stall before they scale
- Why adding AI to your existing workflow makes it worse
- The AI-native operating model, function by function
Sources
- Deloitte, "State of Generative AI in the Enterprise," Q4 / Wave 4 (3,235 leaders, 24 countries, August-September 2025). deloitte.com
- IBM Institute for Business Value, "CEO Study" (roughly 2,000 CEOs, 33 countries, May 6, 2025). newsroom.ibm.com
- BCG, "The Widening AI Value Gap: Build for the Future 2025" (1,250+ firms, September 2025). bcg.com
- Gartner, "Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (poll of 3,412 attendees, June 25, 2025). gartner.com
- McKinsey, "The State of AI: How Organizations Are Rewiring to Capture Value" (March 2025). mckinsey.com
- McKinsey, "From Promise to Impact: How Companies Can Measure and Realize the Full Value of AI" (2025). mckinsey.com
- Deloitte Global, "AI ROI: The Paradox of Rising Investment and Elusive Returns" (1,854 executives, Europe and the Middle East, 2025). deloitte.com
By Christopher Kliebenstein. We build and run AI-native workflows for operators who want results, not demos.