Nearly every large company now pays for AI: chat licences, coding assistants, copilots inside the office suite, pilots with a vendor, an internal platform team. Ask what that spend returned and the answers get vague. Usually it is a count of seats, a chart of prompts, or an “hours saved” figure that nobody can trace to a task. None of those is a return. This note sets out what the research shows, where the common metrics mislead, and a method for measuring AI ROI that holds up when finance asks how you got the number.
What the evidence says about AI returns
Surveys of executives agree on one thing: returns lag adoption. Their exact numbers are not comparable, because each one defines success differently, so read them as a pattern rather than a trend line.
of respondents say AI has contributed positively to their organisation’s EBIT, essentially unchanged from 2025.
McKinsey, State of AI 2026 1of AI initiatives delivered the ROI expected of them in recent years, CEOs report. Only 16% scaled enterprise-wide.
IBM Institute for Business Value, 2025 4of 1,250+ companies are achieving AI value at scale. 60% report little to no value so far.
BCG, The Widening AI Value Gap, 2025 3organisations track KPIs for their gen AI solutions, the practice most linked to bottom-line impact.
McKinsey, State of AI, March 2025 2There are brighter readings. In a Wharton and GBK survey of about 800 US leaders, 74% reported positive ROI from gen AI and 72% said they formally measure it.5 The widely shared MIT NANDA claim that 95% of organisations get zero return sits at the other extreme. It is based on a small sample and a narrow definition of success, and its method has been contested.6 The honest summary is that nobody knows the average return on corporate AI, because most companies are not measuring it in a way that could tell them.
That is also the most useful finding. McKinsey tested 25 practices and found that redesigning workflows had the biggest effect on whether gen AI shows up in EBIT. Tracking well-defined KPIs had the biggest effect of the adoption practices.2 The companies that see returns are the ones that change how work is done and then measure it.
Usage is not value
The easiest AI numbers to get are about usage: licences assigned, weekly active users, prompts sent, suggestions accepted. They matter, because a tool nobody opens cannot pay for itself. But they measure inputs. A thousand prompts a week tells you people are trying something. It does not tell you whether a report is better, a customer waited less, or an error was caught.
Some dashboards blur the line. Microsoft’s Copilot Dashboard reports “assisted hours” by counting Copilot actions and multiplying each by a fixed, research-based time value. Microsoft’s own documentation says these should be viewed as a general estimate.7 That is a modelled number, not a measured saving in your workflow.
Seat counts also hide how uneven use is. In a six-month field experiment across 6,000 workers at 56 firms, people who used Microsoft 365 Copilot spent about three fewer hours a week on email. Averaged across everyone who was given access, the effect was about 1.4 hours.8 Paying for a seat and changing a job are different events.
Scene 01 / Felt vs measuredMETR, 2025
That gap between perception and measurement is not a one-off. METR itself notes that survey estimates of AI productivity have tended to run higher than field experiments.10 If your ROI case rests on a staff survey asking how much time AI saves, you are measuring sentiment.
Where AI pays, and where it quietly costs
Rigorous studies do find real gains, but they are specific to the task and the person. In a Fortune 500 customer support centre, access to an AI assistant raised issues resolved per hour by 15% on average. The least experienced agents gained most, and the most experienced saw small speed gains and small declines in quality.12 Across three field experiments with 4,867 software developers, coding assistants raised completed tasks by about 26%, again with larger gains for less experienced developers.13
Scene 02 / The jagged frontierHBS & BCG, 2023
Then there are the costs that never reach the licence invoice. In a Stanford and BetterUp survey of 1,150 US employees, 40% said they had received “workslop” in the past month: AI-generated work that looks finished but needs fixing. Each instance took close to two hours to deal with.14 Workday’s research with 3,200 AI users found that nearly 40% of the time AI saved was lost to rework, and only 14% of employees consistently got net-positive outcomes.15 Both are surveys, and Workday sells software. Still, the direction is consistent: gross time saved is not net time saved, and the review cost often lands on someone else.
Five layers, from spend to value
The fix is to stop asking one number to do every job. Measure AI in layers, each feeding the next. Spend you cannot trace to a workflow is not yet an investment. It is an option you are paying to hold.
Scene 03 / Where spend leaksIllustrative
- Spend
Every cost, not just licences: usage-based API bills, build and integration time, training, review time and maintenance. Track a unit cost. The FinOps Foundation’s advice is to move from cost per token toward cost per successful outcome, because “the token is the billing unit, not the value unit.”16
Owner: finance and IT - Usage
Who uses what, how often, and for which tasks. Look at the distribution, not the average: a few heavy users and many idle seats is a different problem from light use everywhere.
Owner: IT and team leads - Workflow
Named, repeated tasks that now run through a changed process, each with an owner, a documented routine and a fallback for when the output is wrong. This is the layer McKinsey’s research ties most closely to EBIT impact.2
Owner: the workflow owner - Outcome
The same task measured before and after: time including review, rework rate, error rate, cycle time, customer wait. Scored by a person against a written quality bar, on a sample, not taken from a self-report survey.
Owner: the workflow owner, with a reviewer - Value
Outcomes turned into money or capacity, with assumptions written down. Capacity only becomes cash if the freed time goes somewhere: more volume, faster revenue, avoided hires or avoided cost.
Owner: finance, signed off by the business
Put a number on it
For one workflow at a time, start from time saved per run and subtract what the research says gets lost along the way: review, fixes and outputs that fail the quality bar. Then subtract the full cost of the AI, not just the licence line.
Net AI return, one workflowPer year
minutes saved per run, after review÷60 ×runs that pass the quality bar ×loaded hourly cost −licences, usage, build, training and upkeep
This values freed capacity at cost. It is an estimate, not a cash saving: capacity only pays if the time is redeployed. A run that fails the quality bar earns nothing, and if the new way is slower after review, the saving is negative.
- Net return, year one
- —
- Value of capacity
- —
- ROI
- —
- Hours freed
- —
- Payback
- —
Every figure comes from the sliders. Nothing is stored or sent. Use your own measured baseline: the defaults are an example, not a benchmark.
A worked example, using the defaults above: a task that took 40 minutes now takes 20 with AI, plus 5 minutes of review. It runs 300 times a month, and 85% of runs pass the quality bar. At a loaded cost of ₹1,000 an hour, that frees about 765 hours a year, worth roughly ₹7.7 lakh. Take away ₹1.8 lakh of annual AI spend and ₹1.5 lakh of one-off build and training, and the first year nets about ₹4.3 lakh. Now drop the runs to 80 a month, and the same workflow loses money. Volume matters as much as speed.
Two more rules keep the number honest. Do not multiply one team’s result across the company unless other teams do the same task at the same frequency. And if outputs fail the agreed quality bar, count zero benefit until the workflow is fixed.
When to expect a return
General-purpose technologies tend to show a productivity J-curve. Early on, companies spend on tools and on the less visible work of changing processes and skills, and measured productivity can dip before it rises.18 AI is following the same pattern. That is why a return measured at the end of a pilot often looks worse than the same workflow a quarter later.
Scene 04 / The J-curveIllustrative
That cuts both ways. Gartner predicted that at least 30% of gen AI projects would be abandoned after proof of concept by the end of 2025, citing poor data, risk controls, cost and unclear business value.17 Stopping a pilot is healthy when it is a decision made against criteria. It is waste when nobody wrote the criteria down.
| Layer | Before | Day 30 | Day 90 and after |
|---|---|---|---|
| Spend | Budget and unit cost target | Actual cost per run | Cost per successful outcome |
| Usage | Who will use it, for what | Weekly use by the named team | Use without prompting, spread across the team |
| Workflow | Owner, routine and fallback agreed | Runs through the new process | Handed to a colleague and still runs |
| Outcome | Baseline sample: time, quality, rework | Same task, with review time counted | Repeat sample without the builder in the room |
| Value | Kill and scale criteria written | Early estimate, assumptions shown | Signed-off figure, or a decision to stop |
Cost and risk guardrails
Two things change an AI return after launch: the price of running it and the risk it carries. On price, model costs are falling fast. Stanford’s AI Index found that the cost of GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024.21 Usage can grow faster than prices fall, though, especially once agents run tasks unattended. Put a budget and an owner on usage-based spend, and review cost per successful outcome monthly.
On risk, a saving that leaks confidential data or ships a wrong answer to a customer is not a saving. The NIST AI Risk Management Framework organises this into four functions: govern, map, measure and manage.19 Its generative AI profile lists risks specific to gen AI, such as confabulation and information integrity.20 Neither is a certification. Use them to decide what gets tested before a workflow counts toward value: which data may enter the tool, what the reviewer checks, and who owns an incident.
Questions to ask before the next AI renewal
- Which named workflows does this spend support, and who owns each one?
- What was the baseline for each, measured before the tool arrived?
- What is our cost per successful outcome, not cost per seat or per token?
- How much review and rework time are we counting against the time saved?
- Which seats have not been used for a real task in the last 30 days?
- What result would make us stop, and when do we check it?
None of this needs a new platform. It needs one workflow, a baseline, an honest count of the costs and a date to check again. Start there, and the company-wide number will have something real underneath it.
Sources
- McKinsey, The state of AI in 2026: On the road to ROI, August 2026. Self-reported executive survey.
- McKinsey, The state of AI: How organizations are rewiring to capture value, March 2025. Correlational, self-reported.
- BCG, The Widening AI Value Gap: Build for the Future 2025, September 2025.
- IBM Institute for Business Value, 2025 CEO Study, May 2025. 2,000 CEOs, with Oxford Economics.
- Wharton Human-AI Research and GBK Collective, Accountable Acceleration, October 2025. About 800 US leaders, self-reported.
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Interviews and conference surveys; success defined as measurable P&L impact about six months after a pilot. Widely cited and widely criticised.
- Microsoft Learn, Microsoft Copilot Dashboard, documentation on assisted hours.
- Dillon, Jaffe, Immorlica and Stanton, Shifting Work Patterns with Generative AI, NBER Working Paper 33795, 2025. Authors include Microsoft researchers.
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025. 16 developers, 246 tasks.
- METR, uplift study update, February 2026; and AI usage survey, May 2026.
- Dell’Acqua et al., Navigating the Jagged Technological Frontier, HBS Working Paper 24-013, 2023. The published journal version reports quality gains as over 30%.
- Brynjolfsson, Li and Raymond, Generative AI at Work, Quarterly Journal of Economics 140(2), 2025.
- Cui, Demirer, Jaffe, Musolff, Peng and Salz, field experiments on AI coding assistants, Management Science, 2026. The estimate is imprecise.
- Niederhoffer et al., AI-Generated “Workslop” Is Destroying Productivity, Harvard Business Review, September 2025. Self-reported survey.
- Workday, Beyond Productivity: Measuring the Real Value of AI, January 2026. Vendor survey of 3,200 AI users.
- FinOps Foundation, Token economics, and unit economics capability.
- Gartner, press release, July 2024. A prediction, not a measurement.
- Brynjolfsson, Rock and Syverson, The Productivity J-Curve, American Economic Journal: Macroeconomics, 2021.
- NIST, AI Risk Management Framework 1.0 (NIST AI 100-1), January 2023.
- NIST, Generative AI Profile (NIST AI 600-1), July 2024.
- Stanford HAI, AI Index Report 2025, April 2025.
Scenes 03 and 04 and the calculator defaults are illustrative, not client results. Figures from surveys are self-reported. Where a study has a newer or published version with different numbers, the note says so.