AI Transformation Reality Gap: Why 90% Use AI but Few See ROI

Updated November 2026. Figures reflect McKinsey’s August 2026 survey and NBER’s summer 2026 wave.

Nearly nine in ten organizations in McKinsey’s State of AI 2026 survey now use AI. Among respondents who use it in their own jobs, 80% say it made them more productive. Ask about the company’s earnings, though, and only 37% say AI has added anything to EBIT. A year earlier the figure was about the same.

That distance between what employees feel and what finance can find is the AI transformation reality gap. Below, we look at how wide it is, where the value goes missing, and what the small group of companies getting real returns does with AI that others don’t.

What Is the AI Transformation Reality Gap?

The AI transformation reality gap is the distance between how widely an organization uses AI and how much measurable business change follows from it. Adoption is easy to count: licenses, users, prompts per week. Transformation is harder, because it only shows up when cost, revenue, cycle time, or the shape of the work itself changes.

In practice, the gap has three layers. The first is financial: plenty of usage, little profit to show for it. The second is perception, since leaders tend to believe more has changed than the people doing the work. The third is expectation. Executives forecast far bigger gains than anything they’ve measured so far.

MIT’s Project NANDA gave the financial layer a name: the “GenAI Divide.” Its 2025 report found clear structural disruption in just two sectors, technology and media. Tools like ChatGPT and Copilot spread quickly everywhere else, but mostly they made individuals faster without moving company P&L.

How Wide Is the Gap? Four Studies Compared

Four major studies point in the same direction. Each one measures something different, so their numbers shouldn’t be added up or ranked against each other.

Every figure in the table comes from what respondents told researchers. None of it comes from audited financials.

McKinsey 2026: More Scale, Same Profit

McKinsey’s latest wave has an odd shape. The share of organizations scaling AI across the enterprise climbed from 38% to 44%, and the share reporting any EBIT impact stayed exactly where it was.

“Any EBIT impact” is also a low bar. McKinsey keeps a much stricter category for “high performers”: companies that attribute at least 5% of EBIT to AI and describe the impact as significant. That group has stayed at about 6% of respondents.

NBER: Modest Gains, Larger Hopes

Executives in NBER’s summer 2026 wave credit AI with 0.2% annual productivity growth over the last three years. Over the next three, they expect 0.6% a year.

An earlier NBER wave surveyed nearly 6,000 executives in the US, UK, Germany, and Australia. It asked the question another way, and 89% reported no impact on labor productivity at all. Because one wave reports an average effect and the other reports the share seeing nothing, the two shouldn’t be read as a trend line.

How Much Should You Trust MIT’s 95%?

The 95% figure is useful as a direction of travel. As a precise rate, it is shakier than the headlines suggest.

MIT published it as preliminary research, with anonymized organizations and outcomes reported by interviewees. Success meant deployment beyond pilot with measurable KPIs, and ROI was measured six months after the pilot. The authors themselves say six months may be too short to judge complex enterprise systems.

There’s also some slippage inside the report. The 95% describes organizations seeing no P&L return, while a separate figure says only 5% of custom enterprise AI tools reached production. The sector count drifts between eight and nine in different sections. Quote the direction with confidence, and the decimals with care.

Why Individual Productivity Rarely Reaches the P&L

Saved time turns into financial return only when something around it changes. That might be the process, staffing, pricing, or how much the team produces. Speed up one task inside an otherwise identical workflow and the P&L doesn’t notice.

Take a support agent who now drafts replies in half the time. If ticket volume per agent stays flat and overtime doesn’t drop, the company books no gain. The same is true if those saved minutes go into fixing weak drafts. That last problem is common enough to have a name, and AI workslop that hurts productivity can quietly eat whatever time the tool saved.

A simple four-stage lens helps show where projects stall. It’s our editorial framing, not a validated maturity model.

Most of the surveys above are good at capturing stage one. Very few get anywhere near stage four.

Where Enterprise AI Value Leaks Out

No single cause explains the gap, and weak models are a smaller part of it than vendors tend to admit.

MIT’s main explanation is what it calls a learning gap. Most GenAI systems don’t hold onto feedback, adapt to context, or get better with use, so they stay stuck at roughly the same level of usefulness.

How companies deploy matters at least as much. In McKinsey’s 2026 survey, nearly three-quarters of high performers had redesigned workflows around AI, compared with about a quarter of other respondents. Most organizations still drop AI into processes built before it existed.

Money tends to flow to the wrong places, too. MIT found GenAI budgets leaning heavily toward sales and marketing. Back-office automation often paid back better, but its savings were harder to show a board.

Sourcing seems to make a difference. In MIT’s interview sample, tools built with outside partners reached deployment about 67% of the time, while internal builds managed roughly 33%. The authors caution that this is a correlation. Companies that buy may differ from companies that build in other ways.

Then there’s the running cost. About one in five McKinsey 2026 respondents say AI operating costs, token spend included, are holding back their use.

Finally, a lot of AI use happens where the company can’t see or measure it. Workers at more than 90% of the companies in MIT’s survey used personal AI tools for work, yet only 40% of those companies had paid for an official LLM subscription. That hidden usage brings its own shadow AI security risks for company data.

Is Some of the Gap Just Measurement Delay?

Partly, yes. MIT’s six-month window is one sign of it. NBER Working Paper 34984 is another. It surveyed about 750 executives, most of them CFOs, and found that the productivity gains they perceived ran ahead of what revenue data showed. The authors’ explanation is that revenue arrives later.

AI usually pays off first in time and only later in money. Finance sees the gain once hiring slows, an outsourcing contract ends, output rises, or prices change. None of that closes the gap on its own. Still, a pilot judged at six months may simply be judged too soon. A pilot judged on usage numbers is being judged on the wrong measure entirely.

What AI High Performers Do Differently

The companies getting returns redesign the work around AI. McKinsey’s high performers also tend to aim at growth as well as efficiency.

MIT’s report includes some self-reported results from organizations that made it across the divide:

• Cutting outsourced customer-service and document work saved some of them $2 million to $10 million a year.
• Spending on external creative and content agencies fell by around 30%.
• Most of the savings came from lower external spend rather than internal layoffs.
• The fastest mid-market firms went from pilot to full rollout in about 90 days. Large enterprises typically needed nine months or more.
• Line managers drove the rollouts, rather than a central AI lab working alone.

A few companies go further and reorganize teams from scratch, with small AI-native pods handling work that used to span whole departments.

Invoice processing shows what redesign looks like in practice. In a basic setup, AI pulls data from each invoice, staff still check every result, and approvals follow the old chain. In a redesigned version, reviewers see only the exceptions AI flags, and approvals route by risk and value. The team then tracks processing cost and payment cycle time rather than tool usage. (This is an illustration, not a documented case.)

Why Leaders and Employees See Different Transformations

Where someone sits changes what they see. In the Wharton–GBK 2026 study, 73% of leaders said AI had significantly changed workflows. Only 58% of entry- and mid-level employees agreed. On training, the split was wider: 78% of leaders said employees were adequately trained, against 56% of employees.

This was the first year the study included junior staff. Its comparisons with earlier years therefore cover managers only.

NBER’s data shows the same pattern in expectations. About 3,000 US employees expected AI to lift productivity by 0.9%, while US executives expected 2.3%. When leadership’s view runs ahead of the front line, training budgets and rollout timelines get built on the rosier picture.

Will the Gap Close on Its Own?

Maybe, but the signals conflict. Executives expect bigger gains ahead, and the CFO research suggests some value is late rather than missing. On the other hand, McKinsey’s EBIT figure hasn’t moved in a year.

Agentic AI raises a further warning. In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, citing rising costs, unclear value, and weak risk controls. It’s a forecast, but it lines up with everything the surveys show about projects that start without a clear business case.

Five Questions That Show Whether an AI Project Is Working

Usage dashboards won’t tell you much. A better test is whether the project can answer the five questions below. If it can’t, it’s still a pilot, however busy the logins look. The checklist is ours, so treat it as an editorial tool rather than a validated one.

The cost question is usually where business cases fall apart, mostly because human review rarely makes it into the savings estimate. A task-level automation audit that prices supervision as its own cost line, and compares scenarios at the same output, gives a far more honest number before any budget is committed.

Closing the Gap Starts With the Workflow

Some of the reality gap is timing. Most of it comes from structure: AI layered onto unchanged workflows, usage tracked instead of outcomes, and nobody accountable for the result. The companies pulling ahead can point to a workflow that changed and a number that moved because of it. That is the standard worth holding any AI project to.

FAQs

Q.Is it true that 95% of AI projects fail?

Not quite. MIT found that 95% of organizations in its 2025 sample saw no measurable P&L return six months after a pilot. That describes one studied group, not a failure rate for AI in general.

Q. How long does it take for AI ROI to show up?

No study gives a reliable timeline. NBER’s CFO research suggests revenue lags behind perceived productivity gains, so judging ROI at six months may be premature.

Q. Is agentic AI closing the gap?

Not yet at scale. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027.

Related: AI Transformation Is a Governance Problem (Not Tech) — 2026 Truth