Do 95% of AI Pilots Really Fail? The Truth Behind the MIT Study

“MIT says 95% of AI pilots fail” made a great headline and a misleading summary. The MIT 95% AI failure figure comes from The GenAI Divide: State of AI in Business 2025, a July 2025 report from MIT’s Project NANDA. It found that about 95% of organizations were getting zero measurable return on their generative AI pilots, and that only about 5% of custom enterprise AI tools reached production. It did not find that AI doesn’t work. It found that results depend on how a company goes about it.

When Fortune reported that MIT found 95% of AI pilots fail, the headline spread fast. For many executives, it sounded like confirmation that AI is all hype and no results. But the study wasn’t saying AI doesn’t deliver ROI. It was pointing out that most companies jump into pilots without strategy, integration or accountability, and that’s why they stall.

The 5% that do succeed show AI can create measurable value, and quickly. What they do lines up with what we see at Daylight when we help clients get from experiments to working systems.

What did the MIT study actually say?

MIT’s “State of AI in Business 2025” report was led by Aditya Challapally with co-authors from Project NANDA, a research initiative at the MIT Media Lab. It combined a review of over 300 publicly disclosed enterprise AI initiatives, structured interviews with people at 52 organizations and survey responses from 153 senior leaders, gathered between January and June 2025.

Its main finding was blunt. Adoption is widespread, and impact is rare. Despite an estimated $30–40 billion in enterprise spending on generative AI, about 95% of organizations reported no measurable effect on profit and loss. The drop-off was steepest for enterprise-grade tools. About 60% of organizations evaluated custom or vendor-built systems, about 20% reached a pilot, and only about 5% reached production. General-purpose tools like ChatGPT and Copilot did better. More than 80% of organizations had explored or piloted them and nearly 40% had deployed them, but mainly to boost individual productivity rather than the P&L.

Look past the headline number and the pattern is more useful:

• Pilots fail when they lack fit. Tools break in workflows, don’t retain context or require too much manual input.
• Success comes from customization and outcomes. The 5% that scale focus on process-specific use cases, integration and measurable results.
• Partnerships matter. External partnerships reached successful deployment about twice as often as internal builds, roughly 67% of the time compared with about 33%.

The report’s message is that approach determines results. “AI doesn’t work” is the wrong takeaway.

Is it true that 95% of AI projects fail?

Not in the way the headline suggests. The 95% describes organizations seeing no measurable P&L impact from generative AI pilots during the research window. It doesn’t mean 95% of AI tools are broken, that employees aren’t getting value from them, or that the projects were abandoned.

The report is also open about its limits. It calls itself preliminary findings, describes its deployment figures as “directionally accurate” because they come from interviews rather than official company reporting, and notes that success definitions vary between organizations and that a six-month observation period may be too short to judge complex rollouts. Critics, including Futuriom and the Marketing AI Institute, questioned the sample sizes, the definition of success and how the 95% was derived. Some coverage even reported different interview and survey counts from the ones in the report.

The fair reading is in between. The exact number is soft, but the pattern it points to, lots of pilots and little scaled impact, shows up in nearly every other major survey.

What has happened since the MIT report?

Other research published around and after the MIT report tells a consistent story. Almost everyone is using AI, and a minority are getting enterprise-wide value from it.

• Abandoned initiatives rose. Earlier in 2025, an S&P Global Market Intelligence survey of more than 1,000 companies found that 42% had abandoned most of their AI initiatives, up from 17% in 2024, and the average organization scrapped 46% of its proofs of concept before production (CIO Dive).
• Agent projects face the same test. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls.
• Leaders who measure report returns. In October 2025, Wharton’s AI adoption report found that three in four leaders saw positive returns on their generative AI investments and 72% were formally measuring ROI.
• Use is broad, scale is not. McKinsey’s State of AI survey, published in November 2025, found 88% of organizations using AI in at least one business function, but only about a third had begun scaling it and 39% reported any enterprise-level EBIT impact.

The numbers differ because each study defines success differently. The practical lesson is to decide what success means before the pilot starts, or you won’t be able to tell which group you’re in.

Why do AI pilots stall? Missing strategy, missing fit

MIT calls the core problem a learning gap. Most generative AI systems don’t retain feedback, adapt to context or improve over time, so they work for one-off tasks and break down in the workflows that matter. The researchers found that the barrier wasn’t model quality, regulation or infrastructure. It was fit.

We see the same thing in organizations that come to us after a failed AI experiment. Teams buy licenses or launch pilots without answering the basic questions: Where does AI actually add value? What ROI will we measure? How will it fit into existing systems? Our post on common AI implementation pitfalls covers the warning signs in more detail.

The result is shadow adoption. MIT found that only 40% of companies had bought an official AI subscription, yet workers at more than 90% of the companies surveyed regularly used personal AI tools for work. Employees use personal ChatGPT accounts daily for quick tasks while official pilots sit unused. So the potential is there. What’s missing is a structured plan for fitting AI into how people actually work.

At Daylight, we start every engagement by turning those experiments into a clear roadmap, tied to KPIs and a realistic integration plan. AI without a plan is just trial and error.

Where does AI ROI really show up?

Many leaders overlook where the biggest gains come from. In MIT’s data, it’s often the back office. Budgets over-index on sales and marketing tools, which took at least half of the AI budget in MIT’s survey, because their results are easy to measure. But the clearest returns often appear in operations and finance, where top performers reported savings from less outsourcing and lower external agency spend.

We see this too. The highest-ROI use cases tend to be the less flashy ones:

• Faster decisions, less guesswork. Real-time insights that turn data into action without adding analysts.
• Operational efficiency without burnout. Automation of admin tasks, documentation or service requests that saves hours per employee.
• Smarter support at scale. AI assistants and knowledge bases that improve customer satisfaction while reducing costs.
• Faster launches. Content and campaign workflows that move faster, reducing time to market.

Most of that ROI comes from taking repetitive work off people’s plates, not from replacing them. Teams spend their time on higher-value tasks, and the business saves time and money.

What are the 5% doing differently?

The organizations on the right side of MIT’s “GenAI divide” share a few habits. As Challapally told Fortune, they “pick one pain point, execute well, and partner smartly.” According to the report, they:

• Start narrow. They pick one high-value workflow instead of a broad rollout.
• Buy or partner before they build. Partnerships succeeded about twice as often as internal builds.
• Demand fit. They insist on customization for their own processes and tools that integrate with existing systems.
• Judge tools on outcomes. They evaluate business results rather than software benchmarks or demos.
• Decentralize with accountability. Managers close to the work drive adoption, with clear ownership of results.
• Choose systems that learn. Tools that retain feedback and improve over time keep delivering after launch.

How to avoid the 95%: Daylight’s AI implementation playbook

So how do you avoid being in the 95%? Daylight is a digital product and experience agency, and our AI strategy services are built around a phased approach that carries strategy through to execution:

1. AI readiness and use case assessment. We audit workflows, identify opportunities and prioritize AI use cases based on feasibility and ROI.
2. AI strategy and implementation roadmap. Together we create a practical plan with ROI models, resource planning and success metrics.
3. In-workflow execution. Instead of adding more tools, we embed AI into the systems you already use, such as CRMs, ERPs and ticketing platforms, for smarter automation and fewer mistakes.
4. Ongoing AI advisory and support. We stay engaged to oversee implementation, optimize performance and evaluate emerging technologies.
5. Continuous improvement planning. AI isn’t a one-time rollout. We measure results against KPIs, refine workflows and expand into new areas once impact is proven.

The point is to keep projects moving, reduce risk and get to measurable business outcomes instead of a string of pilots. It sits alongside our responsible AI principles, which guide how we use and recommend AI.

What sets Daylight apart

Plenty of firms talk about AI strategy. What we do differently is pair business-first thinking with hands-on implementation:

• Practical over theoretical. We skip the hype and focus on solutions you can use and measure.
• Strategy plus execution. We don’t stop at a roadmap. We help you deliver and optimize.
• Continuous value creation. AI adoption is ongoing, so our support and insights grow with you.

The MIT study calls out the gap between static tools and systems that learn and integrate. That gap is where we do most of our work, building systems that change with your workflows and keep delivering ROI.

AI Strategy

Not sure where AI fits in your business?

We’ll help you find the one workflow worth starting with.

Let’s Talk

How long does it take to move an AI pilot to production?

Successful companies don’t sit in pilot mode for long. They move quickly, with a structured rollout. MIT found that top-performing mid-market companies averaged about 90 days from pilot to full implementation, while large enterprises took nine months or longer. The exact timeline varies, but the winning pattern looks like this:

• Phase 1: Baseline and readiness. Define KPIs and high-value opportunities.
• Phase 2: Pilot and validate. Focus on one workflow and measure impact.
• Phase 3: Expand and optimize. Scale adoption, enable teams and track KPIs.

Some organizations see measurable results within months. Others take longer, depending on complexity. What matters is momentum. A clear, repeatable process keeps you from stalling in pilots and builds toward impact that lasts.

AI isn’t failing, ad hoc pilots are

MIT’s report shows that AI without strategy, integration and accountability fails. That’s a different finding from “AI fails,” and a more useful one. With the right partner, businesses can cross the AI divide and get ROI that compounds over time.

At Daylight, we help clients ignore the noise, pick the high-value opportunities and build AI systems that work inside real workflows. That work usually starts with an AI strategy consult. The goal is faster decisions, less wasted effort and growth you can measure.