Skip to content

Measurement & Value Realization Framework

Why This Matters Now

Enterprise AI has won the adoption argument and lost the proof argument. Usage is now the norm: 89% of organizations report regular AI use in at least one business function, up from 78% in 2024 (McKinsey, 2026; McKinsey, 2024), and 75% of knowledge workers already use AI at work (Microsoft & LinkedIn, 2024). The tools are deployed, the pilots are funded, and the dashboards are lit up.

What is missing is evidence that any of it moved the financials. Only 37% of organizations attribute any level of EBIT impact to AI (McKinsey, 2026). That leaves roughly six in ten unable to point to a single point of profit. There is also reason to discount what the other four in ten report. A Federal Reserve and Duke survey compared the AI productivity gains firms reported against the gains implied by those firms' own reported revenue and employment. The reported figure was about three times the implied one (Atlanta Fed, 2026). A survey of some 6,000 executives found nine in ten reporting no productivity or employment impact at their own firm (Yotzov et al., 2026). Both of those are working papers. Worker records point the same way on a narrower question. In Denmark, chatbots moved earnings and recorded hours by less than 2% across 11 exposed occupations (Humlum & Vestergaard, 2026). That study measures workers rather than firm profit, and is also a working paper. "We measure it" and "we can show it in the financials" are different claims. A report from MIT's Project NANDA put the figure harsher still. It reported 95% of enterprise generative-AI pilots delivering no measurable P&L impact (MIT Project NANDA, 2025). That report was not peer reviewed, and we hold it through a secondary copy rather than the original. Its own methodology section calls its figures directional, drawn from interviews rather than company reporting, with success defined by the researchers (MIT Project NANDA, 2025). By some estimates more than 80% of AI projects fail, about twice the failure rate of non-AI IT projects, and the root causes are organizational rather than technical (RAND, 2024). In 2024, Gartner forecast that at least 30% of generative-AI projects would be abandoned after proof of concept by the end of 2025 (Gartner, 2024). A January 2026 Gartner article reported at least 50%, without publishing its method (Gartner, 2026).

Government survey data measures something narrower, and lower. Census Bureau researchers found 14.5% of current AI users did not expect to be using it six months later (Census CES, 2024). The figure comes from a working paper rather than an official Census estimate. It describes a firm dropping AI altogether, not a project shelved after a pilot, so it is not the complement of Gartner's figure. A firm can abandon most of its proofs of concept and still be an AI user. Be careful with any published failure rate. Researchers once applied the software industry's most-cited failure definitions to 1,211 real projects, and the figures came apart. The publisher's reply was that its reports should be treated as opinion (Eveleens & Verhoef, 2010). A single headline rate is also the wrong shape of claim. Across 5,392 IT projects, cost overruns followed a power law, most small and a few enormous (Flyvbjerg et al., 2022). Where AI abandonment has been studied directly, it happened mostly after deployment (Chappidi & Singh, 2026). That study samples documented cases and has no project denominator.

Programs are everywhere. Proof is nowhere.

This is the value gap, and it is fundamentally a measurement failure before it is a technology failure. Notably, 59% of leaders already worry about quantifying the productivity gains of AI (Microsoft & LinkedIn, 2024) — they sense the gap but lack the instrumentation to close it. This framework is the conceptual spine of how to close it: instrument before you deploy, separate leading signals from lagging outcomes, and wire measurement into a loop that reallocates capital toward what works.

Instrumentation Before Deployment

The single most expensive measurement mistake is trying to prove value after the fact. Once a tool is live and a workflow has changed, the pre-AI state is gone. You cannot reconstruct a baseline from memory, and self-reported "it feels faster" is not a number a CFO will underwrite.

Instrumentation before deployment means defining the metrics, the baselines, and the measurement method before a single user touches the tool. For every use case, three things must exist on day zero:

  • The metric. What specifically will change — cycle time, cost per transaction, error rate, revenue per rep, throughput. One use case, one or two primary metrics.
  • The baseline. The current value of that metric, measured over a representative window before launch. This is the part that is irrecoverable later; if you skip it, the entire value case becomes unprovable.
  • The method. How the metric will be captured going forward, ideally instrumented into the system rather than reconstructed from surveys.

The MIT funnel shows where programs leak when this discipline is absent: 60% of organizations evaluated generative-AI tools, only 20% reached a pilot, and just 5% reached production (MIT Project NANDA, 2025). Many of the casualties never had a defensible baseline to argue for the next stage of funding. Instrumentation is not a reporting afterthought — it is the precondition for survival past proof of concept. The discipline of choosing which metric to baseline starts upstream, in the AI strategy framework, where use cases are tied to outcomes in the first place.

Leading vs. Lagging Indicators

Measurement programs fail when they confuse motion with results. The distinction that resolves this is leading versus lagging indicators.

Leading indicators are early, fast-moving signals of activity: adoption rate, weekly active users, queries per user, share of workflows touched. They tell you whether the tool is being used. They move within weeks and are easy to collect.

Lagging indicators are the outcomes leadership actually bought: productivity per employee, cost per unit of work, cycle-time reduction, margin, and ultimately EBIT. They move slowly, lag adoption by months, and require the baseline established before deployment to be interpretable.

The classic trap is mistaking the first for the second. Adoption is near-universal — 89% regular use (McKinsey, 2026) and 75% of knowledge workers on board, 46% of them having started within the last six months (Microsoft & LinkedIn, 2024; Microsoft & LinkedIn, 2024) — while measured value remains rare, with only 37% claiming any EBIT impact (McKinsey, 2026). A program can be a runaway success on every leading indicator and a complete blank on every lagging one. Leading indicators are necessary but never sufficient; they are the diagnostic that tells you whether a lack of value comes from non-use (fixable through AI adoption) or from a use case that simply does not pay.

Closing the gap between the two depends on changing how work is done, not just adding a tool to it. In McKinsey's 2025 survey, workflow redesign was the attribute most correlated with EBIT impact. Yet only about 21% of organizations had fundamentally redesigned at least some workflows (McKinsey, 2025).

This is the one leg of the argument with independent support behind it. Across roughly 300 large US firms, technology paid off more where the organizational investments were also present (Bresnahan et al., 2002). Across 7,121 UK establishments, doubling IT capital went with 1.2% higher productivity at domestic and non-US-multinational sites. At those owned by US multinationals it went with 4.9%. Management practice accounted for most of that difference (Bloom et al., 2012). Expect the payoff to lag. Measured returns to computers ran two to five times higher over seven years than over one (Brynjolfsson & Hitt, 2000). Those are the working paper's figures; the published article reports a smaller sample and a narrower range. The complementary investments are intangible and poorly captured while they are being built (Brynjolfsson et al., 2020). A program measuring its own AI returns early should expect to find little. Leading-indicator success without workflow redesign is the most common path to a high-adoption, zero-impact program. Translating a lagging movement back to the intervention that caused it is the job of the companion attribution methodology.

The Measurement-to-Reprioritization Feedback Loop

Measurement that only reports is overhead. Measurement that changes decisions is the asset. The engine of value realization is a loop:

Measure → reallocate → reinvest → measure again.

Each cycle reads the evidence, moves money and attention away from use cases that are not converting leading signals into lagging outcomes, and concentrates resources on the few that are. The reinvested capital then produces the next round of evidence, which sharpens the next reallocation.

The hard part is not the measuring; it is the reallocating. Most organizations are structurally bad at moving resources. The year-over-year correlation of a business unit's budget is 0.92 — near-total inertia, with this year's allocation almost entirely predicting next year's (McKinsey, 2012). AI does not break this gravity on its own. Without a deliberate loop, the strongest use cases get the same incremental budget as the failures, and the portfolio never improves.

The payoff for breaking inertia is well established: companies in the top third for reallocating capital earned roughly 30% higher total shareholder returns annually than the bottom third over the study period (McKinsey, 2012). Applied to AI, the loop turns a static portfolio of bets into a managed one. Two enablers make it run: KPIs that actually reflect AI's contribution — companies that revise their KPIs with AI are 3x more likely to see greater financial benefit than those that do not (MIT SMR & BCG, 2024) — and the operational plumbing to capture evidence on a cadence, covered in standing up AI measurement.

How Measurement Compounds Returns

The loop is not a one-time correction; it compounds. Each turn improves the average quality of the portfolio, which raises the evidence yield, which funds a better next turn. Over several cycles, the organizations that measure and reallocate pull steadily away from those that do not. This is the AI value flywheel, and the data now shows it separating the field.

The separation is stark. More than 60% of "future-built" (leading) firms rigorously track AI value, versus only 17% of stagnating firms — roughly a 3.5x gap in measurement discipline (BCG, 2025). The same firms show about 1.7x the revenue growth and 1.6x the EBIT margins of the 60% BCG calls stagnating or emerging (BCG, 2025). That is an association, not proof that measurement drives the results. And the leading cohort is small — only 5% of companies are future-built, while roughly 60% have not achieved material value (BCG, 2025). The same shape appears in McKinsey's data, where about 6% are "AI high performers" attributing 5% or more of EBIT to AI (McKinsey, 2026).

A small, measurement-disciplined minority is capturing a disproportionate share of the value while the majority stalls. The flywheel explains why the gap widens rather than closes: leaders compound their advantage every cycle, while laggards — unable to prove what worked — keep funding the average. Measurement is not a back-office function in this picture. It is the mechanism that determines which side of the gap an organization ends up on.

Key Takeaways

  • The value gap is a measurement failure first. Adoption is near-universal, but only 37% of organizations claim any EBIT impact and 95% of generative-AI pilots show no measurable P&L impact (McKinsey, 2026; MIT Project NANDA, 2025). Most root causes are organizational, not technical (RAND, 2024).
  • Instrument before you deploy. Define metric, baseline, and method on day zero. A baseline cannot be reconstructed after launch, and without it the value case is unprovable.
  • Do not mistake leading for lagging. Adoption and usage move fast and prove nothing about outcomes; productivity, cost, and EBIT move slowly and require a pre-deployment baseline to read.
  • Redesign work, don't just add a tool. Workflow redesign is the attribute most correlated with EBIT impact, yet only about 21% of organizations had fundamentally redesigned at least some workflows (McKinsey, 2025).
  • Wire measurement into reallocation. The measure-reallocate-reinvest loop fights the 0.92 budget-inertia gravity; top reallocators historically earned ~30% higher annual shareholder returns (McKinsey, 2012; McKinsey, 2012), and KPI-revising firms are 3x more likely to see greater financial benefit (MIT SMR & BCG, 2024).
  • Measurement and results travel together. Leading firms track value at ~3.5x the rate of laggards and show 1.7x the revenue growth and 1.6x the EBIT margins (BCG, 2025; BCG, 2025). Our reading is a flywheel that widens the gap, but the data show only an association.

To put this framework into practice, see the companion pages: attribution methodology for tying outcomes back to AI, standing up AI measurement for the operational build, and measurement maturity scoring to benchmark where your program stands today.

Sources

  • McKinsey — The state of AI in 2026: On the road to ROI, 2026. Nearly nine in ten respondents report regular use of AI in at least one business function. View source · verified 2026-09-13 · primary
  • McKinsey — The State of AI: How Organizations Are Rewiring to Capture Value, 2024. 78% of organizations now use AI in at least one business function (up from 72% in early 2024 and 55% in 2023). View source · verified 2026-06-22 · primary
  • Microsoft & LinkedIn — 2024 Work Trend Index Annual Report, 2024. 75% of knowledge workers now use AI at work. View source · verified 2026-06-20 · primary
  • McKinsey — The state of AI in 2026: On the road to ROI, 2026. About four in ten respondents (37 percent) report that AI has contributed positively to their organizations' EBIT, essentially unchanged from 2025. View source · verified 2026-09-13 · primary
  • Baslandze, Edwards, Graham, McClure, Sparks, Meyer, Waddell & Weitz (Federal Reserve Bank of Atlanta) — Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives (Atlanta Fed Working Paper 2026-4), 2026. the mean reported labor productivity growth attributable to AI is 1.8% ... implied labor productivity growth is substantially smaller at 0.6% ... produced 603 responses. View source · verified 2026-09-16 · primary
  • Yotzov, Barrero, Bloom, Bunn, Davis, Foster et al. — Firm Data on AI (NBER Working Paper 34836), 2026. 69% of firms actively use AI ... nine-in-ten reporting no impact on employment or productivity. View source · verified 2026-09-16 · primary
  • Humlum & Vestergaard — Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI (NBER Working Paper 33777, revised March 2026), 2026. ruling out effects larger than 2% two years after the launch of ChatGPT. View source · verified 2026-09-16 · primary
  • MIT Project NANDA — The GenAI Divide: State of AI in Business 2025, 2025. Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact. View source · verified 2026-06-20 · ⚠ secondary mirror
  • MIT Project NANDA — The GenAI Divide: State of AI in Business 2025 (methodology and limitations), 2025. 95% of organizations are getting zero return ... directionally accurate based on individual interviews rather than official company reporting ... success definitions may differ across organizations. View source · verified 2026-09-16 · ⚠ secondary mirror
  • RAND — The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, 2024. By some estimates more than 80 percent of AI projects fail, twice the rate of failure for information technology projects that do not involve AI. View source · verified 2026-06-20 · primary
  • Gartner — Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025, 2024. At least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value. View source · verified 2026-06-20 · primary
  • Gartner — Why Half of GenAI Projects Fail: Avoid These 5 Common Mistakes (Arun Chandrasekaran, January 26, 2026), 2026. Gartner finds that by the end of last year, at least 50% of generative AI projects were abandoned after proof of concept due to poor data quality, inadequate risk controls, escalating costs or unclear business value. View source · verified 2026-09-14 · ⚠ secondary mirror
  • Bonney, Breaux, Buffington, Dinlersoz, Foster, Goldschlag, Haltiwanger, Kroff & Savage (US Census Bureau CES) — Tracking Firm Use of AI in Real Time: A Snapshot from the Business Trends and Outlook Survey (CES 24-16; also NBER w32319), 2024. a non-trivial fraction (14.5%) of the current users do not expect to use in the future, and another 17.6% don't know whether they will ... The average biweekly response rate over the period of collection for AI-related core and supplement content is about 16%. View source · verified 2026-09-16 · primary
  • Eveleens & Verhoef (Vrije Universiteit Amsterdam) — The Rise and Fall of the Chaos Report Figures (IEEE Software 27(1)), 2010. result in meaningless figures ... Standish opinion and the reader bears all risk in the use of this opinion. View source · verified 2026-09-16 · primary
  • Flyvbjerg, Budzier, Lee, Keil, Lunn & Bester — The Empirical Reality of IT Project Cost Overruns: Discovering A Power-Law Distribution (Journal of Management Information Systems 39(3)), 2022. IT project cost overruns follow a power-law distribution. View source · verified 2026-09-16 · primary
  • Chappidi & Singh (University of Cambridge) — To Build or Not to Build? Factors that Lead to Non-Development or Abandonment of AI Systems (ACM FAccT 2026), 2026. 85% (n=77) of analyzed incidents were abandoned post-deployment ... Less than 5% of all AIAAIC repository entries involved AI abandonment. View source · verified 2026-09-16 · primary
  • Microsoft & LinkedIn — 2024 Work Trend Index Annual Report, 2024. 59% of leaders worry about quantifying the productivity gains of AI. View source · verified 2026-06-20 · primary
  • MIT Project NANDA — The GenAI Divide: State of AI in Business 2025, 2025. 60% of organizations evaluated enterprise-grade systems, but only 20% reached pilot stage and just 5% reached production. View source · verified 2026-06-20 · ⚠ secondary mirror
  • Microsoft & LinkedIn — 2024 Work Trend Index Annual Report, 2024. 46% of users started using AI at work less than six months ago. View source · verified 2026-06-20 · primary
  • McKinsey — The State of AI: How Organizations Are Rewiring to Capture Value, 2025. the redesign of workflows has the biggest effect on an organization's ability to see EBIT impact from its use of gen AI. ... Twenty-one percent of respondents reporting gen AI use by their organizations say their organizations have fundamentally redesigned at least some workflows. View source · verified 2026-06-20 · primary
  • Bresnahan, Brynjolfsson & Hitt — Information Technology, Workplace Organization, and the Demand for Skilled Labor (Quarterly Journal of Economics 117(1)), 2002. The effects of IT on labor demand are greater when IT is combined with the particular organizational investments we identify ... evidence of complementarities among all three of these innovations in factor demand and productivity regressions. View source · verified 2026-09-16 · primary
  • Bloom, Sadun & Van Reenen — Americans Do IT Better: US Multinationals and the Productivity Miracle (American Economic Review 102(1)), 2012. A doubling of the IT capital stock is associated with 1.2 percent higher productivity for a domestic or non-US multinational, but 4.9 percent higher productivity for an establishment owned by a US multinational ... these management practices account for most of the higher output elasticity of IT of US firms. View source · verified 2026-09-16 · primary
  • Brynjolfsson & Hitt — Computing Productivity: Firm-Level Evidence (MIT working paper, April 2000; published Review of Economics and Statistics 85(4), 2003), 2000. the implied returns to computers are two to five times greater when differences are taken over seven years instead of one year ... 600 large US firms over 1987-1994. View source · verified 2026-09-16 · ⚠ secondary mirror
  • Brynjolfsson, Rock & Syverson — The Productivity J-Curve: How Intangibles Complement General Purpose Technologies (NBER Working Paper 25148, revised January 2020; published AEJ: Macroeconomics 13(1), 2021), 2020. Our adjusted measure TFP is 11.3% higher than official measures at the end of 2004 ... these complementary investments are often intangible and poorly measured in the national accounts. View source · verified 2026-09-16 · primary
  • McKinsey — How to put your money where your strategy is, 2012. For our entire sample, the mean correlation between how much capital a business received in one year and how much it got the next was 0.92; for one-third of companies that correlation was higher than 0.99. View source · verified 2026-06-20 · primary
  • McKinsey — How to put your money where your strategy is, 2012. Companies that reallocated more (the top third) yielded on average 30 percent higher total returns to shareholders annually than those that reallocated the least (the bottom third). View source · verified 2026-06-20 · primary
  • MIT SMR & BCG — The Future of Strategic Measurement: Enhancing KPIs With AI, 2024. Companies that revise their KPIs with AI are three times more likely to see greater financial benefit than those that do not. View source · verified 2026-06-20 · primary
  • BCG — The Widening AI Value Gap, 2025. Leaders consistently track and report results. More than 60% of future-built firms rigorously track AI value, compared with only 17% of stagnating companies. View source · verified 2026-09-13 · primary
  • BCG — The Widening AI Value Gap, 2025. Future-built companies already generate 1.7 times more revenue growth and 1.6 times higher EBIT margins than the 60% of companies in the categories we term stagnating or emerging. ... As a result of this investment, they expect twice the revenue increase and 1.4 times greater cost reductions than laggards in the areas where they apply AI. View source · verified 2026-09-14 · primary
  • BCG — The Widening AI Value Gap, 2025. Only 5% of companies in our 2025 study of more than 1,250 firms worldwide are achieving AI value at scale ... Fully 60% of companies are not achieving material value at all, reporting minimal revenue and cost gains despite substantial investment. Another 35% (13 percentage points more than in 2024) are scaling up their efforts and seeing some returns ... the 60% that have little or no value. View source · verified 2026-09-13 · primary
  • McKinsey — The state of AI in 2026: On the road to ROI, 2026. AI high performers - respondents who attribute an EBIT impact of 5 percent or more to AI use and say their organizations have seen 'significant' value from AI use - account for just 6 percent of survey respondents, unchanged from 2025. View source · verified 2026-09-13 · primary