If your AI scorecard leads with hours saved, you are reporting the metric that Wharton's annual survey of US enterprises links to the weakest results. In Wharton and GBK Collective's fourth annual study, released October 13, 2026, 35% of departments that track at least one growth measure report strongly positive AI returns, against 13% of departments that track efficiency alone. Hours saved is still the most common AI metric (40%), and net-new revenue is the least used (27%).
That gap is a correlation drawn from self-reported answers, and the report's own definition of "growth" is looser than the word suggests. It is still the most useful thing to take into a 2027 budget meeting, because it tells you which number your CFO will find hardest to dismiss, and which number your own leadership is most likely to be wrong about.
What Did Wharton and GBK Actually Measure?
The study is a 15-minute online survey of 1,416 people at US companies with 1,000 or more employees and over $50 million in revenue, fielded July 9 to July 29, 2026, according to the methodology page of the report. Of those, 1,007 are managers and above; 409 are entry- and mid-level employees, surveyed for the first time this year so the researchers could compare what leaders report with what staff experience.
The headline numbers, per the Business Wire release carried by FinancialContent: 80% of leaders report positive AI returns, and 28% call them strongly positive. "The vast majority of companies are using AI and most are seeing some return," said Stefano Puntoni, a Wharton professor who co-directs Wharton Human-AI Research.
Compare that with last year. The 2025 edition, published October 28, 2025, found three in four leaders seeing positive returns and 72% formally measuring ROI, and four in five expected their investments to pay off within two to three years. So the share reporting any return barely moved. The 2026 report leans on a harder bar ("strongly positive"), and only 28% of leaders clear it.
One rule in the method matters for how you read every ROI figure below: departments that do not formally measure their return are counted as not having shown one. Measurement discipline is baked into the outcome variable.
Why Growth Metrics Show 35% and Hours Saved Shows 13%
Departments that judge AI on at least one growth outcome are far more likely to report strong returns, and the share rises with each growth metric added. In the report's breakdown, strongly positive ROI runs 13% for efficiency-only departments (n=189), 32% with one growth metric (n=376), 37% with two (n=311) and 43% with three (n=65, which the authors flag as a small base to read directionally).
The scorecards look like this, among the 941 managers who measure ROI at least occasionally (each picked up to three):
| KPI | Type | Share using it |
|---|---|---|
| Hours saved per employee per week | Efficiency | 40% |
| Reduction in error rates | Efficiency | 37% |
| Increase in total output volume | Growth | 35% |
| Reduction in operational costs | Efficiency | 34% |
| Reduction in task completion time | Efficiency | 34% |
| Customer satisfaction scores | Growth | 34% |
| Increase in sales conversions | Growth | 31% |
| Net-new revenue from AI-enabled products | Growth | 27% |
Three caveats before you reorganise anything around this table.
First, "growth" is generously defined. The four growth measures are net-new revenue, sales conversion, customer satisfaction and output volume, which the survey illustrates with "lines of code, marketing campaigns, reports generated." A department counting AI-written reports is in the growth bucket. A finance team will read that as throughput, and it will discount it accordingly.
Second, most of the efficiency-only group still reports a return: 13% strongly positive and another 65% moderately positive, so roughly four in five. What separates the two groups is how strong the return looks.
Third, the direction of cause is unknown. A department that already sells something AI touches can measure conversion; a back-office team processing invoices mostly cannot. Teams that pick three KPIs are probably more mature measurers overall. The report itself concedes the same problem elsewhere: managers very confident they can track total AI spend report strong ROI at 56% against 10%, but "those same respondents are also more enthusiastic about AI in general, so some of the difference may reflect conviction rather than capability."
The intent gap is the part with no caveat attached. 83% of managers expect AI to fundamentally transform or majorly affect their industry within two to five years, yet 52% say AI's primary job over the next 12 to 24 months is productivity (44%) or cost reduction (7%). Only 8% name new revenue or products, and 14% customer experience. Practitioners have been making the same complaint all year: Sam Laidlaw at Atomic Object called hours saved "the vanity metric of AI adoption" in May, and UC Today's Kristian McCann described a "Time Saved Fallacy" in April, where freed time disappears into more meetings and email unless someone decides where it goes. The same leak showed up in the data behind AI saving 11 hours a week while workers spend 6.4 babysitting it.
Does a Chief AI Officer Raise ROI, or Just Signal Commitment?
A dedicated Chief AI Officer goes with much stronger reported returns; a CAIO title bolted onto an existing executive does not. Among managers whose organisation created a newly designated CAIO position (n=252), 58% report strongly positive ROI. Where the responsibility was added to an existing role such as the CIO (n=421), it is 24%, which the report says is no better than organisations still deciding. With no CAIO and no plans for one, it falls to 8%, on a base of 63 that the authors call small and directional.
The trend line explains why this matters now. 67% of managers report a CAIO or equivalent, up from 61% in 2025 and 47% in 2024. But 42% say the responsibility was added to an existing job, against 25% who describe a new position, and the report notes that nearly all of this year's growth is reassignment.
IBM's Institute for Business Value reached a similar direction with a different method. Its July 2025 study with the Dubai Future Foundation, based on more than 600 CAIOs in 22 countries, found organisations with a CAIO see 10% higher ROI on AI spend, and only 26% of organisations had one. Neither study can separate the role from the budget and executive attention that usually come with creating it. A company that funds a new C-level seat has already decided AI is a priority; the 58% may be measuring that decision.
One Platform Beat the Central Hub, Which Contradicts IBM
Organisations that standardise on one approved AI platform report the strongest returns, and those still experimenting report the weakest. Among managers who influence AI decisions, 50% of those standardised on a single platform (n=184) report strongly positive ROI, against 30% of those allowing multiple approved tools under a formal policy (n=472), 19% of those running a central hub or orchestration layer across models (n=193), and 9% of those still experimenting without a defined approach (n=85), per the report. Only 19% of organisations have standardised; 49% run the multi-tool policy.
The hub result is where the evidence disagrees. IBM's CAIO study reported 36% higher ROI where CAIOs run a centralised or hub-and-spoke operating model. Wharton's central-hub respondents sit near the bottom. The two studies define the hub differently (an operating model in IBM's case, a technical layer managing access to multiple models in Wharton's), and they measure different things: an ROI uplift against a share reporting strongly positive returns. If you are about to fund an AI gateway on the strength of either number, you are reading more into it than the data carries.
The Wharton authors' own takeaway is narrower than "pick one vendor": the main goal is any settled approach. Read the hub figure with care, though: by the report's own significance markers, the hub's 19% is not statistically distinguishable from the experimenters' 9%. Only single-platform and multi-tool policy clear that bar.
Your Training Numbers Are Probably 22 Points Too High
Leaders think their people are better trained than the people do, and training is the largest gap in the study. 78% of managers say employees receive adequate AI training; 56% of entry- and mid-level employees agree. The report finds the same split on every readiness question:
- Leadership has set a clear long-term AI vision: 79% of managers, 64% of staff.
- Confident using AI in daily work: 79% against 62%.
- Managers empowered to lead AI change: 84% against 65%.
- Adequate investment in responsible AI policy: 77% against 59%.
- Workflows changed significantly because of AI: 73% against 58%.
The authors call this seniority gap the most consistent finding in the data. It appears in every industry and company size, and in every department except Purchasing/Procurement. Overall, 39% of managers describe an organisation that is aligned on AI (four of five factors, training among them); 26% of junior employees do.
This connects back to ROI. Departments reporting strong returns are more likely to adequately train employees (91% against 73%), include AI in performance reviews (87% against 65%) and monitor AI use (92% against 74%). They did not cut more staff than others did. Overall, 37% of managers report AI-related staff reductions and 33% slowed or frozen hiring, and where cuts happened, only 20% describe an aligned organisation, against 53% where they did not.
If your training-readiness figure comes from a manager survey or an LMS completion report, the study suggests it overstates what the people doing the work believe by roughly 22 points. The gap matters more as junior roles change shape: 24% of managers expect fewer entry-level hires over the next two to five years, and Goldman Sachs already has new hires supervising AI agents before they learn the job.
What to Change Before the 2027 Budget Is Locked
Budgets are going up, so the scorecard you attach to them now will be the one you are judged on next October. 87% of managers expect AI budgets to rise in the next 12 months and 47% expect more than 10%, while 64% agree that rising token, licence and API costs are a major barrier to using AI more (74% of VP-and-above). Only 40% are very confident they can track total AI spend once costs bundled into other software and cloud bills are counted.
This Week:
- Pull the KPI list your AI programme reports to the executive team and tag each line efficiency or growth using Wharton's eight definitions. If there is no growth line, that is the first gap.
- Ask finance whether "hours saved" has ever been converted into a cost line, a headcount plan or deferred hiring. If the answer is no, stop presenting it as return and relabel it as capacity.
This Month:
- For each production AI use case, name one growth metric the owning team can actually observe: conversion rate, CSAT, throughput on a revenue-bearing queue, revenue from an AI-enabled feature. Set a baseline from the last full quarter before the AI went live. Do not count "reports generated" as growth in your own deck, even though the survey does.
- Run a short anonymous readiness survey of individual contributors with the same five questions Wharton used (training, vision, confidence, manager empowerment, responsible AI policy). Put the result next to your management numbers.
- Write down who owns AI. If it is a CIO with "and AI" added to the title, list what that person stopped doing to make room. If the answer is nothing, the 24% cohort describes you.
Before Budget Sign-Off:
- Decide on a platform posture and publish it: one approved platform, or a formal multi-tool policy with named tools. Do not take "still experimenting" into 2027; it was the 9% group.
- Build a total-AI-spend view that includes AI features bundled into existing SaaS and cloud contracts, so the ROI denominator is the real one.
The Bottom Line
The pattern in this year's data is the one economists expected. Erik Brynjolfsson, Daniel Rock and Chad Syverson's productivity J-curve paper argued that general purpose technologies need complementary investment in processes and skills that goes unmeasured early, so measured productivity lags and then catches up. Wharton finds the same shape inside AI deployments: departments using agents to their fullest extent reach transformational success at 31% and advise-only departments at 25%, while those approving every action (16%) or sampling afterwards (18%) sit in a trough. Only 7% of organisations succeed on all three of the study's markers.
The counterweight is still the MIT NANDA study from mid-2025, which reported that 95% of organisations saw no business return from $30 billion to $40 billion of generative AI spending, based on 52 interviews, 153 survey responses and a review of more than 300 public initiatives. Wharton's 80% positive and MIT's 95% zero are measuring different things: a manager's view of their own department, against P&L impact a researcher could find.
Put a growth metric with a pre-AI baseline on every AI line item before the 2027 budget is signed.
Continue Reading
- Deutsche Telekom's €2.5B AI Savings Target Excludes Token Costs
- M&T Published 3 Copilot Counts. None Was Weekly Actives.
- Walmart and Amazon Both Said 40%. Neither Ran a Holdout.
- Airbnb Shipped 80% More Features. One Number Is Auditable.
- Workday Finds Basic AI Skill Demand Fell 25% After January Peak
- Sysco Tied Equity Pay to Its AI Cost-Out Push. Net of What?