--- title: "MIT AI Report: What the 95% Number Actually Measured" description: The MIT AI report's 95% figure is real, but it measured something narrower than 'AI projects fail.' See what it counted and how to check any AI claim. image: https://www.sharkitectdigital.com/hubfs/raw_assets/public/cleo/site-theme/images/sharkitect-icon.png --- [Skip to content](https://www.sharkitectdigital.com/blog/mit-ai-report-95-percent-what-it-measured#main-content) [![](https://www.sharkitectdigital.com/hs-fs/hubfs/raw_assets/public/cleo/site-theme/images/sharkitect-icon.png?width=36&height=36&name=sharkitect-icon.png) SHARKITECTDIGITAL](https://www.sharkitectdigital.com/) [Services](https://www.sharkitectdigital.com/services) [Industries](https://www.sharkitectdigital.com/industries) Free resources [Bottleneck FinderFind yours in three minutes](https://www.sharkitectdigital.com/bottleneck-finder) [GuidesHow-to guides for owners](https://www.sharkitectdigital.com/guides) [Operator NotesShort articles from the field](https://www.sharkitectdigital.com/blog) [Proof](https://www.sharkitectdigital.com/case-studies) [About](https://www.sharkitectdigital.com/about) [Contact](https://www.sharkitectdigital.com/contact) [Book the Assessment](https://www.sharkitectdigital.com/architecture-assessment) [Book the Assessment](https://www.sharkitectdigital.com/architecture-assessment) 1. [Home](https://www.sharkitectdigital.com/) 2. [Blog](https://www.sharkitectdigital.com/blog) 3. The MIT AI Report and Its 95% Number: What It Measured, and How to Check an AI Claim Guide # The MIT AI Report and Its 95% Number: What It Measured, and How to Check an AI Claim By Christopher Sharkey, Founder & CEO · September 29, 2026 You've probably heard it in a blog post or a sales pitch: "MIT research shows 95% of AI projects fail." The MIT AI report behind that line is real, and it says something narrower. Most of the organizations it studied hadn't yet seen generative AI move their profit. I think knowing what to believe is now a bigger problem for owners than the technology itself. So here's the habit worth building: before you believe an AI number, ask where it came from and what it counted. ## What does the MIT AI report actually say? The report is "The GenAI Divide: State of AI in Business 2025," from MIT's Project NANDA, dated July 2025 and written by Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari. GenAI means generative AI: AI that creates new content, like writing, summaries or images, from what you ask it. ChatGPT is the best-known example. The executive summary opens with this: > "Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return." ... "Just 5% of integrated AI pilots are extracting millions in value, while the vast majority remain stuck with no measurable P&L impact." A pilot is a trial run of a new system. P&L is the profit and loss statement. You can read it in an [archived copy of MIT's original link](https://web.archive.org/web/20250818115520/https://nanda.media.mit.edu/ai_report_2025.pdf). ## Do 95% of AI pilots fail? That depends on which 95% you mean. The exact phrase "AI projects fail" isn't in the report, though page 7 does use "failure rate" for one specific measure. The report attaches 95%, or its flip side of 5%, to at least three different things: - Organizations getting "zero return" on generative AI (page 3). - Integrated AI pilots, where "Just 5%" are "extracting millions in value" (page 3). - Enterprise-grade AI systems: 60% of organizations evaluated them, 20% reached a pilot, and 5% reached production (page 3). Page 7 calls this "The 95% failure rate for enterprise AI solutions." It defines success more than one way, too. Page 7 counts a tool as successful when users or executives describe "a marked and sustained productivity and/or P&L impact." Page 24 uses "deployment beyond pilot phase with measurable KPIs" (key performance indicators), with the return "measured 6 months post-pilot." The authors are candid about the limits. Page 6 says the figures are "directionally accurate based on individual interviews rather than official company reporting." Page 24 adds that six months "may be insufficient" for complex systems, "potentially understating success rates." The report's own headline wording is blunt. What got lost in the retelling was the scope: enterprise and mid-market companies, AI pilots, and a return measured six months after the pilot. A pilot that hasn't shown a return yet isn't necessarily a failure. ## How to check an AI statistic before you believe it Don't ask only, "Is this claim true?" Ask, "What evidence would have to exist for this claim to be credible?" When I hear a statistic like the MIT 95% figure, I want to know: - Where did the number come from, and can I trace it to the original? - How was it determined, and who did the research? - Who was studied, and how were they chosen? - What counted as success or failure, and what was left out? - Does the source actually support the conclusion being presented? Having "a source" doesn't make a claim credible. I could make a statement tomorrow, link to some random website, and technically say, "Here's my source." Hold AI claims to the standard you'd expect of a journalist. ## Real data can still tell the wrong story I spent about ten years as a healthcare Performance Improvement Specialist. That work taught me you can have real data and still reach a misleading conclusion. Say you're measuring room turnover time, the time to clean and reset a room between patients. The benchmark is 10 minutes, and your average is 20. You present the data, and people start deciding which cases shouldn't count. Sometimes those exclusions are fair. But if you remove data after seeing the result, because it doesn't fit the story you want, you've changed what you're measuring. The number looks better. The reality hasn't changed. I'm not saying the MIT authors did this. The risk sits with anyone who retells a number without its definitions. Good analysts ask what was counted, what wasn't, and why. Ask the same of any AI number. ## Checking the MIT number: source, method, cause Those five questions come down to three. ### Source: where did this claim come from? Real and traceable: a named MIT project, a dated report, page numbers you can check. ### Method: how was it determined? The research ran from January to June 2025 (page 2). It drew on a review of over 300 public AI initiatives, "structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences." Page 7 defines enterprises as "firms with over $100 million in annual revenue." The words "small business" never appear. "SMBs" shows up once, in a list of buyer types on page 3, and no finding is broken out for small companies. So it tells you little about a 10-person business. ### Cause: what does the report say went wrong? The retelling implies the cause is AI itself. The report names its own: "The core barrier to scaling is not infrastructure, regulation, or talent. It is learning." It says most enterprise-grade systems fail "due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations." A brittle workflow breaks as soon as real work stops matching the setup. ## Why AI gets stuck: the band-aid problem What I see in small businesses is related, and it's my observation, not MIT's: AI gets used as a band-aid on a broken or poorly defined process. If the root cause isn't fixed, the AI may seem to work for a while, and then the problem comes back. That's why [your records need one source of truth before AI touches them](https://www.sharkitectdigital.com/blog/source-of-truth-before-ai-automation). You hear the plug-and-play pitch everywhere. "Missed calls? Get this AI voice agent." "Upload your documents and your AI will know what to do." Uploading a folder doesn't teach AI your business. It works inside the instructions, rules, data and connections it's been given. When someone promises one tool will "solve your problem," slow down and ask questions. ## AI hype words that should trigger a question Some words should work like an alarm: "always," "guaranteed," "instant," "effortless," "the only solution," "#1," "zero risk." Any of them can be true, but each is your cue to ask for evidence. "Most" should make you ask: most according to what data? "Best": best by what measure? "The biggest ever": compared with what? Critics get the same test as vendors. ## What this checklist can't tell you These questions can't tell you whether a tool will work in your business. If you haven't measured AI in your own business yet, the honest verdict on a claim you can't check will be "unproven." Treat that as a reason to test small. For example, have an AI tool draft your next 20 estimate follow-ups, check each one before it goes out, and count how many you'd have sent unchanged. And the same questions apply to anyone helping you with AI, us included: ask how success will be measured and what happens when the system is wrong. ## Questions owners ask ### Did MIT say 95% of AI projects fail? Not in those words. The report uses 95% for organizations getting "zero return," for pilots with "no measurable P&L impact," and, on page 7, for enterprise AI tools that didn't reach successful implementation. "AI projects fail" is a retelling that merges all three. ### Does the MIT AI report apply to small businesses? Not directly. Its enterprises are firms with over $100 million in annual revenue, and it breaks out no finding for small businesses. ### Where can I read the MIT AI report? As of September 2026, MIT's original link leads to the project's page, which offers its reports through a request form. An [archived copy of the original link](https://web.archive.org/web/20250818115520/https://nanda.media.mit.edu/ai_report_2025.pdf) is on the Internet Archive. You might also like [What is AI washing?](https://www.sharkitectdigital.com/blog/what-is-ai-washing), which covers the other side of the same problem: products made to sound smarter than they are, and the one question that exposes it. ![](https://www.sharkitectdigital.com/hs-fs/hubfs/raw_assets/public/cleo/site-theme/images/sharkitect-icon.png?width=34&height=34&name=sharkitect-icon.png) SHARKITECTDIGITAL The AI Transformation Agency, an agency that refuses to upsell. Kansas City, Missouri · serving the KC metro in person and the US remotely. [![Sharkitect Digital BBB Business Review](https://seal-nebraska.bbb.org/logo/sehzbum/bbb-1000086813.png)](https://www.bbb.org/us/mo/raytown/profile/artificial-intelligence/sharkitect-digital-0714-1000086813/#sealclick) ## Talk to us [(816) 443-0826](tel:+18164430826) [solutions@sharkitectdigital.com](mailto:solutions@sharkitectdigital.com) [Book the Architecture Assessment](https://www.sharkitectdigital.com/architecture-assessment) ## Services - [Lead response automation](https://www.sharkitectdigital.com/services/lead-response-automation) - [System integration](https://www.sharkitectdigital.com/services/system-integration) - [Local marketing engine](https://www.sharkitectdigital.com/services/local-marketing-engine) - [Social content engine](https://www.sharkitectdigital.com/services/social-content-engine) - [All services](https://www.sharkitectdigital.com/services) ## Free to use - [All guides](https://www.sharkitectdigital.com/guides) - [Lead follow-up guide](https://www.sharkitectdigital.com/guides/lead-follow-up) - [Connecting your tools](https://www.sharkitectdigital.com/guides/connect-your-business-tools) - [The owner's bottleneck](https://www.sharkitectdigital.com/bottleneck-sorter) - [The Bottleneck Finder](https://www.sharkitectdigital.com/bottleneck-finder) ## Company - [About](https://www.sharkitectdigital.com/about) - [Case studies](https://www.sharkitectdigital.com/case-studies) - [FAQ](https://www.sharkitectdigital.com/faq) - [Industries](https://www.sharkitectdigital.com/industries) - [Kansas City](https://www.sharkitectdigital.com/locations/kansas-city) - [Contact](https://www.sharkitectdigital.com/contact) © 2026 Sharkitect Digital. All rights reserved. [Privacy](https://www.sharkitectdigital.com/privacy-policy) [Terms](https://www.sharkitectdigital.com/terms-of-use) [Text messages](https://www.sharkitectdigital.com/sms-opt-in) Cookie settings [LinkedIn](https://www.linkedin.com/company/sharkitect-digital) [Instagram](https://www.instagram.com/sharkitect_digital) [Facebook](https://www.facebook.com/SharkitectDigital) **Cookies.** Analytics so we know which pages actually help. [What we collect](https://www.sharkitectdigital.com/privacy-policy) Accept Essential only ```json { "@context" : "https://schema.org", "@id" : "https://www.sharkitectdigital.com/#organization", "@type" : "Organization", "areaServed" : [ { "@type" : "City", "name" : "Kansas City" } ], "description" : "An AI Transformation Agency for contractors and trade businesses. We diagnose broken systems and build what actually fixes them.", "email" : "solutions@sharkitectdigital.com", "founder" : { "@id" : "https://www.sharkitectdigital.com/#chris" }, "legalName" : "Sharkitect Digital LLC", "name" : "Sharkitect Digital", "sameAs" : [ "https://www.google.com/maps?cid=11278465346363269515", "https://www.linkedin.com/company/sharkitect-digital", "https://www.facebook.com/SharkitectDigital", "https://www.instagram.com/sharkitect_digital", "https://www.trustpilot.com/review/sharkitectdigital.com", "https://nextdoor.com/pages/sharkitect-digital-independence-mo/" ], "telephone" : "+1-816-443-0826", "url" : "https://www.sharkitectdigital.com/" } ``` ```json { "@context" : "https://schema.org", "@id" : "https://www.sharkitectdigital.com/#chris", "@type" : "Person", "description" : "Founder of Sharkitect Digital. Previously spent about ten years as a healthcare Performance Improvement Specialist, finding what was broken inside working operations.", "familyName" : "Sharkey", "givenName" : "Christopher", "image" : "", "jobTitle" : "Founder & CEO", "knowsAbout" : [ "Business process automation", "System integration", "QuickBooks Online integrations", "Construction and trades software", "Lead response automation" ], "name" : "Christopher Sharkey", "sameAs" : [ "https://www.linkedin.com/in/christopher-sharkey-aitp" ], "url" : "https://www.sharkitectdigital.com/about", "worksFor" : { "@id" : "https://www.sharkitectdigital.com/#organization" } } ``` ```json { "@context" : "https://schema.org", "@id" : "https://www.sharkitectdigital.com/#website", "@type" : "WebSite", "name" : "Sharkitect Digital", "publisher" : { "@id" : "https://www.sharkitectdigital.com/#organization" }, "url" : "https://www.sharkitectdigital.com/" } ``` ```json { "@context" : "https://schema.org", "@type" : "BreadcrumbList", "itemListElement" : [ { "@type" : "ListItem", "item" : "https://www.sharkitectdigital.com/", "name" : "Home", "position" : 1 }, { "@type" : "ListItem", "item" : "https://www.sharkitectdigital.com/blog", "name" : "Blog", "position" : 2 }, { "@type" : "ListItem", "item" : "https://www.sharkitectdigital.com/blog/mit-ai-report-95-percent-what-it-measured", "name" : "The MIT AI Report and Its 95% Number: What It Measured, and How to Check an AI Claim", "position" : 3 } ] } ``` ```json { "@context" : "https://schema.org", "@type" : "Article", "articleSection" : "Guide", "author" : { "@id" : "https://www.sharkitectdigital.com/#chris" }, "dateModified" : "2026-09-30", "datePublished" : "2026-09-29", "description" : "The MIT AI report's 95% figure is real, but it measured something narrower than 'AI projects fail.' See what it counted and how to check any AI claim.", "headline" : "