Short answer: Measure AI ROI one workflow at a time. Record a baseline before launch, measure the same numbers on a comparable mix of cases after go-live, and count only value that changes cash, risk or capacity you actually use. ROI is counted value minus total cost, divided by total cost, over a fixed period.
Measure AI ROI one workflow at a time. Record a baseline for that workflow before launch, measure the same numbers on a comparable mix of cases after go-live, and count only value that changes cash, risk or capacity you actually use. Freed hours are not savings until someone decides what happens to them. The formula is simple: counted value minus total cost, divided by total cost, over a fixed period, with review time and your own staff’s time on the cost side.
Below: why the big AI ROI studies disagree, which metrics fit which workflow, how to take a baseline in two weeks, a worked example, what to tell the board at 30, 90 and 180 days, and a measurement plan you can copy.
Why are most published AI ROI numbers unreliable?
Most headline figures come from surveys that ask executives how AI is going. That tells you what leaders believe, which is useful. It is not a measurement. Here is what the most cited studies actually asked:
Study · sample | What it measured · headline |
|---|---|
McKinsey State of AI 2026 · 1,719 respondents, 97 nations | Whether respondents attribute EBIT impact to AI · 37% report some |
Wharton and GBK Collective 2025 · about 800 US enterprise leaders | ROI as judged from internal conversations · 74% positive |
IBM CEO Study 2025 · 2,000 CEOs, 33 countries | Share of AI initiatives that delivered expected ROI · 25% |
Google Cloud ROI of AI 2025 · 3,466 leaders, vendor-commissioned | ROI within the first year · 74% say yes |
Deloitte, Europe and Middle East 2025 · 1,854 executives | Time to satisfactory ROI on a typical use case · mostly 2 to 4 years; 6% under a year |
MIT NANDA 2025 · 52 interviews, 153 survey responses | Tools with marked, sustained impact, as remarked by users or executives · 95% of organizations get zero return |
Sources: McKinsey, Wharton, IBM, Google Cloud, Deloitte, MIT NANDA.
These are the most recent figures of each kind we could find as of October 2026, and each study states what it asked. The trouble starts when a headline strips the question away. Seniority alone moves the answer: in the Wharton study, 81% of VPs and above believed ROI was positive, against 69% of mid-level managers.
MIT NANDA’s own authors call their figures directionally accurate, drawn from interviews rather than official company reporting. And individual and company views split. McKinsey found that 80% of respondents say AI has improved their own productivity, while only 37% report any EBIT contribution. In the report, senior partner Dan Tinkoff writes: “A striking finding this year is the gap between individual gains and enterprise impact.”
What happens when someone measures instead of asks?
Controlled studies are rarer and narrower, but they show why your own measurement matters:
Measured gains are real, and uneven. In a study of 5,179 customer support agents, Brynjolfsson, Li and Raymond found that a generative AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled agents, with minimal effect for the most experienced.
People misjudge their own speed. In METR’s 2025 trial, 16 experienced developers took 19% longer with AI tools, yet afterwards believed AI had sped them up by 20%. METR now treats that result as historical: in February 2026 it said developers are likely faster with today’s tools, while calling its own new data weak evidence. Its May 2026 survey note adds that “public estimates of productivity impacts from surveys have tended to be greater than those from field experiments or quasi-experiments.”
Saved time does not show up on its own. Linking adoption surveys to Danish payroll records, Humlum and Vestergaard found that workers report productivity benefits, yet effects on earnings and recorded hours two years after ChatGPT’s launch were precisely zero, ruling out anything above 2%.
The lesson for a CEO: do not borrow anyone’s ROI number. Measure your own, on one workflow, against your own baseline.
We hold our own numbers to the same standard. On a regulated ops invoice workflow we run in production, most routine prep is automated — boundary and period still under validation; freed time is not cash until you decide where it goes.
What counts as AI value?
Count value in three buckets, and only these:
Cash. Costs that stop (overtime, temp staff, a vendor, a hire you no longer make), fees and penalties avoided, discounts captured, revenue you can trace to the workflow.
Risk. Errors, late filings or duplicate payments that no longer happen, measured against the baseline rate and priced at what they cost you before.
Capacity actually used. Freed hours that went to named work with its own result: a backlog cleared, a new service, faster replies that you can show kept customers.
Everything else, such as logins, prompts sent or “hours saved” in a survey, is an activity metric. Watch it, but do not put it in the ROI line. BCG’s 2026 Applied AI Index makes the same point from the other end: among the most AI-mature companies, 95% measure AI through the P&L or KPIs, and those tracking directly in the P&L report three times the AI value of companies that do not formally measure it.
Which metrics work for which workflow?
Pick one primary metric that the workflow exists to move, and one or two guardrails that must not get worse.
Workflow type | Primary metric · guardrails |
|---|---|
Document processing (invoices, claims, intake) | Cycle time and cost per case · exception rate · errors found in audit |
Customer service | Time to first answer · resolutions per agent hour · reopen rate · satisfaction |
Sales support (research, proposals) | Time from lead to first reply · revenue per rep · win rate |
Internal or legal knowledge questions | Time to answer · accuracy on a sampled audit · escalations |
Quality and compliance reports | Cycle time · reviewer corrections · late submissions |
Content production | Throughput per person · revision rounds · approval rate |
Leading and lagging indicators
Lagging indicators, such as cost per case, revenue per rep and error rate, are the ones that pay. They take a quarter or more to settle. Leading indicators tell you in the first weeks whether you are on track:
share of eligible cases going through the new workflow;
share of AI outputs accepted with no edits or light edits;
review minutes per case;
escalations and exceptions per 100 cases;
incidents caught by the approver before anything left the building.
If leading indicators stall, the lagging ones will not arrive. Deloitte’s 2026 State of AI in the Enterprise survey of 3,235 leaders shows the usual gap: 66% report productivity and efficiency gains from AI, but only 20% report increased revenue so far.
How do you take a baseline in two weeks?
Two weeks is enough if you lean on system data and sample the rest. Avoid holiday weeks and month-end peaks unless the workflow lives there.
Days 1 and 2: define the case. Write down the event that starts a case and the event that ends it. Choose the primary metric and the guardrails from the table above. Find where timestamps already exist: the helpdesk, the ERP, the document system.
Days 3 to 10: collect. Pull cycle times and volumes for the last three months from the systems, so you can see seasonality. For hands-on time, have each person time a random sample of cases (30 to 50 each is plenty) rather than log everything. People work differently when they know they are timed, so prefer system timestamps wherever they exist.
Days 11 and 12: audit quality. Have someone who did not do the work check 50 finished cases against a checklist. That gives you an error or rework rate you can repeat later with the same checklist.
Days 13 and 14: sign off. Put the numbers on one page. Have the workflow owner and someone from finance sign it, together with the decision on what counts as value and what will happen to freed hours.
Keep the audit checklist and a set of past cases with known correct outcomes. You will reuse both after go-live. If you have not picked the workflow yet, start with how to choose your first AI workflow.
How do you calculate AI ROI?
Use one formula and state the period:
ROI = (counted value − total cost) ÷ total cost, over 12 months.
Payback in months = build cost ÷ (monthly counted value − monthly run cost).
Total cost includes the build, the monthly run cost, the owner’s time and the review time that people spend checking outputs. Our AI implementation cost guide breaks those lines down.
Worked example: supplier invoices
This example is hypothetical. It shows how the numbers fit together. It is not drawn from any client.
A 150-person consumer health company processes 2,000 supplier invoices a month. The baseline, taken over two weeks: 9 minutes of hands-on time per invoice (300 hours a month), a 6-day median cycle time, 12% of invoices corrected or sent back, and about €1,500 a month in late fees and missed early-payment discounts.
After go-live, the workflow reads each invoice, matches it against the order and the goods receipt, and drafts the posting. A person approves every payment. By day 90, 75% of invoices match cleanly and take 1.5 minutes of review. The other 25% are exceptions and take 10 minutes each. The owner spends 10 hours a month on spot checks.
Hands-on time after: 37.5 + 83.3 + 10 = about 131 hours. Freed: about 169 hours a month, not the 225 you would claim if you ignored review, exceptions and oversight.
Cycle time falls to 2 days, which removes most fees and captures the discounts: €1,500 a month in cash.
Build: €20,000 including internal time. Run: €900 a month.
The result depends on what happens to the 169 hours:
Scenario | Counted value · payback · 12-month ROI |
|---|---|
A. Nothing changes: same team, same work | €1,500/month cash · 33 months · −42% |
B. A planned hire is not made (€4,500/month loaded) | €6,000/month · 3.9 months · +134% |
Spreadsheet that values every freed hour at €45 | €9,113/month · 2.4 months · +255% |
The third row is what many calculators show. It assumes every freed hour turns into money, which is only true if a cost actually goes away or the hours go to work with a measurable result. BCG’s AI at Work 2026 survey of 11,749 workers found 66% get limited or no guidance on what to do with the time they save. Decide before launch whether you are in scenario A or B, and write it into the plan.
What should you report to the board after 30, 90 and 180 days?
Day 30: is it working? Leading indicators only: cases through the workflow, acceptance rate, review minutes per case, exceptions and incidents. Run cost against plan. No ROI claim yet. Say what you are fixing.
Day 90: is it better than before? The primary metric and guardrails against the baseline, on a comparable mix of cases. The first cash items. Where the freed hours went, by name. Any change in error rate, good or bad.
Day 180: is it worth it? Counted value against total cost, signed off by finance, and payback to date. Then a decision: scale it, fix it or stop it, and which workflow comes next.
Keep expectations sober. Deloitte’s survey of 1,854 executives in Europe and the Middle East found most achieved satisfactory ROI on a typical use case in two to four years, and only 6% within a year. One narrow, well-measured workflow can move faster, but the board should hear that from your numbers, not from a vendor’s.
AI ROI measurement plan template
Copy this into a document and fill in the right-hand column. The examples come from the hypothetical invoice case.
Field | Your entry (example) |
|---|---|
Workflow, start and end | Invoice received → posted and scheduled for payment |
Owner | Head of accounts payable |
Value type | Cash · capacity used |
Primary metric | Cycle time and cost per invoice |
Guardrails | Errors found in a 50-invoice audit · exceptions per 100 |
Baseline period and source | Two weeks plus three months of ERP history |
Baseline values | 9 min · 6 days · 12% corrected · €1,500/month fees |
Day-90 target | 2 days · under 5% corrected |
Leading indicators (day 30) | Share matched · review minutes · incidents |
What happens to freed hours | Planned hire not made; or named backlog work |
Costs counted | Build · run · review time · owner time |
Attribution method | Pilot team vs comparison team, or staggered start |
Other changes to log | New ERP release · volume spikes · staff changes |
Sign-off | Owner and finance at day 90 and 180 |
Stop rule | Error rate above baseline for two months |
What are the most common traps?
Counting usage as value. In the Wharton study, the most common way companies measured generative AI ROI (47% of those measuring) was employee engagement or productivity, with clicks and hours used as examples. Usage tells you people opened the tool, not that a cost went away.
Ignoring review time. Workday’s 2026 research with 3,200 respondents reports that nearly 40% of AI time savings are lost to rework: correcting errors, rewriting content and verifying outputs. That figure is self-reported too, so do not borrow it. Time the approval step yourself and put it on the cost side.
Attribution. Volume, mix, staffing and other projects change at the same time. Keep a log of other changes. Where you can, start with one team and compare it with a team that has not switched yet, or stagger the start dates.
Trusting estimates of time saved. Ask people and they will tell you what they feel. METR’s trial is a reminder that feelings and stopwatches can point in opposite directions.
Counting freed hours as savings. The Danish payroll data show how easily saved minutes disappear into other work. Name where the hours go or do not count them.
Claiming the return too early. At day 30, report leading indicators. Financial claims wait for day 90 at the earliest.
IBM’s 2025 CEO study shows the gap between confidence and results: 68% of CEOs said their organization has clear metrics to measure innovation ROI, yet they reported that only 25% of AI initiatives had delivered the expected ROI. A plan with a baseline, a value rule and a sign-off closes that gap for one workflow at a time.
Where to start
Pick one workflow, take the two-week baseline and fill in the template before anyone builds. If you are still choosing, read how to choose your first AI workflow. To price the cost side, use the worksheet in what it costs to implement AI. We agree the baseline before we build and measure real use after launch. Implementation projects start at €10,000 (from $12,000 in the US). Bring your own numbers. Book a fit check — leave with a Map in 48h.
Frequently asked questions
What is the formula for AI ROI?
ROI is counted value minus total cost, divided by total cost, over 12 months. Payback in months is the build cost divided by monthly counted value minus monthly run cost. Total cost includes the build, the run cost, the owner’s time and the time people spend reviewing outputs.
Why are published AI ROI figures unreliable?
Most come from surveys that ask executives how AI is going, which tells you what leaders believe. Seniority alone moves the answer: in the Wharton study, 81% of VPs and above believed ROI was positive, against 69% of mid-level managers. Do not borrow anyone’s number. Measure your own.
How long does an AI baseline take?
Two weeks is enough if you lean on system data and sample the rest. Define the case, pull cycle times and volumes from your systems, and have people time a sample of cases. Then audit 50 finished cases for quality, and have the owner and finance sign the numbers off.
Are hours saved by AI the same as money saved?
No. Freed hours count only when a cost goes away or the hours go to named work with a measurable result. Humlum and Vestergaard linked adoption surveys to Danish payroll records. They found that effects on earnings and recorded hours two years after ChatGPT’s launch were precisely zero.
What should you report to the board after 30, 90 and 180 days?
At day 30, leading indicators only, such as acceptance rate and review minutes per case, and no ROI claim. At day 90, the primary metric and guardrails against the baseline, plus the first cash items. At day 180, counted value against total cost, signed off by finance, and a decision to scale, fix or stop.
Related reading
Sources
McKinsey & Company, The state of AI in 2026: On the road to ROI, August 25, 2026. Online survey of 1,719 participants in 97 nations, May 4 to June 8, 2026.
Wharton Human-AI Research and GBK Collective, Accountable Acceleration: Gen AI Fast-Tracks Into the Enterprise, October 2025. Online survey of about 800 US enterprise leaders, June 26 to July 11, 2025.
IBM, IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles, press release, May 6, 2025. Survey of 2,000 CEOs in 33 countries, February to April 2025.
Google Cloud, Google Cloud Study Reveals 52% of Executives Say Their Organizations Have Deployed AI Agents, press release, September 4, 2025. Survey of 3,466 leaders in 24 countries, commissioned by Google Cloud and conducted with National Research Group.
Deloitte, AI ROI: The paradox of rising investment and elusive returns, October 22, 2025. Survey of 1,854 executives in 14 countries in Europe and the Middle East, August 15 to September 5, 2025.
MIT NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. Preliminary findings.
Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, Generative AI at Work, NBER Working Paper 31161, April 2023, revised November 2023.
METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 10, 2025.
METR, We are Changing our Developer Productivity Experiment Design, February 24, 2026.
METR, Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity, May 11, 2026.
Anders Humlum and Emilie Vestergaard, Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI, NBER Working Paper 33777, May 2025, revised March 2026.
BCG, The Formula for Agentic AI Value (Applied AI Index 2026), September 30, 2026. Survey of more than 1,300 CxOs and senior leaders.
Deloitte, The State of AI in the Enterprise, 2026. Survey of 3,235 leaders in 24 countries, August to September 2025.
BCG, AI Is Reshaping Jobs Faster Than Companies Are Reshaping Work, press release, June 3, 2026. Survey of 11,749 workers in 14 markets.
Workday, New Workday Research: Companies Are Leaving AI Gains on the Table, press release, January 14, 2026. Survey of 3,200 respondents, November 2025.