95% of Enterprise AI Pilots Show No P&L Impact, MIT Finds

By Tom Whitfield

MIT researchers analyzed 300 enterprise AI deployments and found that roughly 95 percent produced zero measurable impact on the profit and loss statement. McKinsey's most recent global State of AI survey points the same way: only about 6 percent of organizations could trace more than 5 percent of operating profit to AI.

The failure is not at the task level. In a study of 5,172 customer support agents, published in the Quarterly Journal of Economics, those given an AI assistant resolved about 15 percent more issues per hour. The hours saved are real; they simply get absorbed by other work that looks productive and does not drive revenue. For HR and operations leaders, that gap turns every AI rollout into a workflow and redeployment decision, not a procurement decision.

What separates the 5 percent?

The companies that generate real returns tend to make four choices differently. Each carries a caveat, because several of the examples cut both ways.

Start with back-end work. Klarna announced in February 2024 that its AI assistant handled two-thirds of customer-service chats, matched the work of 700 agents, cut resolution time from 11 minutes to under two, and was expected to drive a $40 million profit improvement that year. A year later, CEO Sebastian Siimiatkowski said the company had over-prioritized cost and needed to restore a human customer-service option. The money was saved; the trust was not. Lumen Technologies took the safer route. Generative AI cut a seller's pre-call customer research from about four hours to 15 minutes, an improvement Lumen valued at $50 million in annual revenue.

Target the highest-friction tasks. Samsung's semiconductor division gave engineers open ChatGPT access in March 2023. Within roughly 20 days, employees had pasted proprietary source code and a confidential meeting transcript into the public tool across three incidents; the company then banned generative AI on company devices. The return was negative and unrecoverable. Bayer ran the opposite play: leadership collected more than 700 application areas across 100,000 employees, then built around the most frequent tasks. One result, an AI training system for pharmaceutical sales representatives, reached 1,000 reps in six months, saved 6,000 to 7,000 hours of training time a year, and improved role-play coaching efficiency by more than 50 percent.

Redesign the workflow, not just the task. McKinsey found that 55 percent of effective AI users had redesigned their workflows, against roughly 20 percent of everyone else. The law firm Levidow, Levidow and Oberman skipped the redesign: ChatGPT invented six cases, lawyers filed them in federal court, and a judge sanctioned the firm. JPMorgan Chase did the opposite. Its contract intelligence system extracts about 150 attributes from 12,000 commercial credit agreements a year, work that had consumed 360,000 hours annually. Lawyers became final-mile reviewers rather than first-draft readers.

Where do the reclaimed hours go?

This is the step nearly everyone skips. A JAMA study of ambient AI scribes across five academic health systems over two years found documentation time fell about 16 minutes a day, yet after-hours EHR time barely moved and revenue rose a nominal $167 per month per clinician. Nobody had decided what the freed minutes were for.

The Permanente Medical Group did decide. Across 7,260 physicians and more than 2.5 million encounters, its scribes saved an estimated 15,791 hours; 84 percent of physicians reported better patient communication, and 47 percent of patients said their doctor spent less time looking at the computer. UCSF then measured the payoff: across nearly 1.2 million encounters, adopters billed 1.81 more relative value units per week, a 5.8 percent increase worth roughly $3,044 per physician annually, while handling 2.8 percent more encounters with no rise in denied claims. That is about $2.1 million a year for 698 adopting physicians; at Permanente's scale, roughly $22 million.

Same technology, different management. Before approving any AI pilot, the practical test is one sentence in writing: when this role gets back X hours per week, those hours will be used for Y. If nobody can complete it, there is no ROI plan. The companies winning here are not using smarter models; they are making clearer choices about where to start, what friction to remove, and where the hours go.