Is AI worth it? What the AI ROI research actually shows
The widely shared claim that 95 percent of AI pilots fail comes from a narrow sample. Here is what the research measured and what successful implementations do differently.
The short answer
AI is far more likely to produce a measurable return when a business treats it as an implementation: one defined problem, a redesigned workflow, trained people, an accountable owner, and results measured against a baseline. The widely shared 95 percent figure comes from an MIT NANDA report that counted a pilot as successful only when it showed a marked and sustained productivity or P&L impact, based on public initiatives, 52 interviews, and 153 conference surveys. A separate survey of nearly 6,000 executives found they personally used AI about 1.5 hours a week. Neither finding measures what AI delivers when a business implements it with discipline.
Where the 95 percent figure comes from
The statistic comes from MIT NANDA's 2025 report, The GenAI Divide: State of AI in Business. Its research base was a review of more than 300 publicly disclosed AI initiatives, 52 structured interviews, and 153 survey responses gathered at four industry conferences. That is useful directional research, but a conference sample and public announcements are a narrow window into how thousands of businesses use AI. The report counted a task-specific AI tool as successfully implemented only when users or executives described a marked and sustained productivity or P&L impact, and it found about 5 percent of integrated pilots met that bar. That is a demanding test for early pilots, and the headline became a general verdict on AI that the sample cannot support.
What the larger executive survey measured
A 2026 working paper published through the National Bureau of Economic Research, Firm Data on AI, surveyed nearly 6,000 senior executives in the United States, the United Kingdom, Germany, and Australia. Nine in ten reported no impact on employment or productivity over the prior three years. The same executives reported using AI for an average of 1.5 hours a week, and they expected AI to raise productivity at their firms by 1.4 percent, raise output by 0.8 percent, and reduce employment by 0.7 percent over the next three years. The same study found that 69 percent of firms actively use AI. Broad adoption combined with light personal use by leaders suggests that access has spread faster than workflow redesign and measurement.
Why so many AI efforts look like failures
Many companies buy licenses, announce access, and wait for results. A subscription with no process design, training, or measurement is a purchase order, not an implementation plan. Meanwhile, employees often adopt their own AI tools because they need relief from workload now, which means the most productive use in a company is frequently happening outside any official plan, unmeasured and ungoverned.
What successful implementations do differently
Businesses that see a return pick one costly problem and solve it well. They redesign the workflow before they deploy the tool, configure the system for that specific process, and train the people who will use it. They judge success by business outcomes such as response time, cycle time, error rates, capacity, and revenue rather than technical benchmarks. They measure continuously with leading and lagging indicators, and they give one person clear accountability for the result.
How to measure the return in your own business
Start with a baseline before anything is built: how long the process takes, how often it fails, how many people touch it, and what it costs each month. After the build, measure the same things the same way. Include software, usage, maintenance, and human review in the cost, and count time saved as value only when the business can redeploy that capacity or avoid a real expense.
A practical first step
Choose a process where information is currently lost or delayed and where the result is easy to see. AI meeting notes are a simple example: a background notetaker preserves decisions, commitments, and customer details from every conversation. Lead follow-up, intake, and scheduling are common next candidates because their results show up quickly in response time and completed next actions.
Sources
Related service
Find the first AI workflow with a return you can measure.
Not ready to book? Get a free diagnostic report from The Advisor