95% of GenAI Pilots Show No P&L Impact. The Cause Is Culture, Not Code

By Daniel Okafor

Ninety-five percent of generative AI pilots have delivered no measurable profit-and-loss impact, according to MIT's NANDA initiative. That number should anchor every AI budget conversation this year. IBM's 2025 CEO study found more of the same: only 25% of AI initiatives met expected ROI, and only 16% scaled enterprise-wide.

The usual reading blames the technology. The evidence does not. Boston Consulting Group surveyed 1,000 senior executives and found roughly 70% of AI implementation failures stemmed from people and processes, 20% from technology, and 10% from the algorithms themselves. Roughly one failure in ten traces to the model. The rest trace to the org chart.

There is a counter-data point worth noting. SAP and Oxford Economics' Value of AI Report 2026 found 69% of businesses worldwide say they are satisfied with AI ROI. Read the fine print: more than two-thirds of that same group still sees room for improvement. Satisfaction, in this survey, is a low bar.

Where do the returns actually come from?

BCG's high performers invested 10% in algorithms, 20% in technology and data, and 70% in people and processes — spending that mirrors the failure distribution almost exactly. That symmetry is the story. Companies that fund the failure mode get the failure; companies that fund the fix get the return.

McKinsey's 2025 State of AI survey adds a wrinkle. Eighty percent of companies set efficiency as an AI objective. The high performers more often added growth or innovation targets, and were nearly three times as likely to have fundamentally redesigned individual workflows. Redesign ranked among the strongest contributors to business impact of every factor McKinsey tested. Buying a copilot license is not redesign.

Timelines are also recalibrating. TEKsystems' State of Digital Transformation 2026 report found only 27% of organizations expect transformation ROI within six months, down from 42% a year earlier. That drop of 15 percentage points in one year suggests boards have stopped accepting early-victory claims. Good. Honest measurement needs room.

What does safe experimentation look like in practice?

A 2025 MIT Technology Review Insights survey found 83% of business leaders believe experiment-friendly cultures make AI initiatives much more likely to succeed. Yet 22% of those same leaders had hesitated to lead AI projects for fear of failure or criticism. The belief and the behavior do not match. That gap is a management problem, not a tooling problem.

MIT's researchers also documented a shadow AI economy: employees using personal AI tools while official pilots stall. Their prescription was moving from central labs to empowered teams. For CIOs the question is blunt — does that initiative happen inside your governance, or around it?

The nonprofit workforce developer Per Scholas offers a working example. Its ResumeBoost tool cut staff time spent on alumni resumes by more than a third. The user experience missed, so the team folded the tool into a broader AI coaching platform rather than defending it. The pilot did its job. It showed promise and showed where to pivot.

The internal build relied on institutional knowledge. Ed Hernandez, who designed Per Scholas's AI-Native Classroom, has been with the organization roughly fifteen years — starting as a learner, then instructor, then leading curriculum development. Leadership's role, the organization says, was to say yes, set guardrails, and weigh possibility against necessity. Candor matters too: people do not experiment freely with a technology they suspect will replace them.

Can collaboration be measured in revenue?

Accenture's research on "Talent Reinventors" — organizations advancing people and AI together — found they delivered revenue growth 1.8 percentage points higher and profit growth 1.4 points higher than peers in 2025. Collaboration treated as infrastructure, not sentiment, shows up in the P&L.

Per Scholas built its job-sourcing tool, Sherlock, with the staff who match graduates to employers, not for them. Staff previously opened a dozen job boards, reran the same searches, and weeded out staffing-firm reposts. Sherlock now reviews more than 125,000 postings a day and delivers one feed matched to training tracks and each campus's recruiting footprint. Saved roles are shared campus-wide and export in the format of the Salesforce job orders staff already create.

Three weeks after launch, staff at nearly three-quarters of campuses had logged in, weekly users had more than doubled, and users had saved over 100 hours of sourcing work. Seven weeks in, the same staff are reporting where the tool falls short. The organization calls it an early success, not a finished story. That framing is the honest one.

Does any of this require a new platform?

No. The argument, made by Per Scholas's leadership, is that AI is a force amplifier, not a force generator — it magnifies whatever culture already exists. Strong culture, distinctive extension; weak culture, the same problems surfacing faster.

Hiring follows the same logic. Equinix's chief people officer, describing the company's early-career pathway investments, said: "We're not just competing for talent; we're creating it." Employers report that technical skills are baseline; what separates candidates is whether they can define a problem, try something new, and work across a team.

Those traits also decide whether the AI budget becomes a line item or a return. Every ROI claim above carries a caveat: none of these surveys measure the same thing. Read them directionally, not literally.