AI failure rates in 2026: why every source quotes a different number

Ask how often AI fails and you will get an answer somewhere between 10% and 95%, depending entirely on who you ask. The AI failure rate is quoted as 95% by people citing MIT, as 80% by people citing RAND, as 42% by people citing S&P Global, and as roughly 10% by us. Those numbers are all defensible and none of them is a correction of the others. They measure four different things, and almost every article that stacks them together is comparing a product census to a survey of enterprise pilots as though the two were rival estimates of one quantity.

This piece separates them. What each number counts, whose question it answers, where it gets misused, and which one to reach for depending on what you are actually trying to decide in 2026.

The four numbers, and what each one actually counts

Read down that denominator column and the disagreement mostly evaporates. One counts products, one counts companies, and two count projects inside companies. A tool can be perfectly alive — not in our 10.3% — and still be the subject of a pilot that produced nothing, which puts it inside MIT's 95%. Those are not contradictory findings. They are orthogonal ones.

The product number: 10.3% of AI tools are gone

This is ours, and it is the narrowest of the four. As of September 3, 2026, 285 of the 2,756 AI tools in our catalogue have shut down or been acquired — a 10.3% product-level failure rate, the first time it has crossed 10%.

What it counts is deliberately literal: a product that was real enough to list, that had a website and users, and that is no longer independently available. It does not count companies, funding rounds, pilots, or disappointment. Of the 285, 135 shut down outright and 150 were acquired — and of the acquisitions, 84 still ship under their own name, which is why we track "acquired" and "dead" as separate states rather than collapsing them into one scary total.

The reason this number is so much lower than the others is not that AI is doing better than people think. It is that "the product still exists" is a very low bar. Our AI graveyard is a census of products that failed to clear it. Most of the failure the other three numbers describe happens well above that line, inside companies that are still paying for tools that still exist.

The company number: how many AI startups shut down

This is the bucket where the sourcing gets weakest, and it is worth saying so plainly. Figures like "40% of AI startups founded in 2024 have already shut down" circulate widely in 2026 and are frequently repeated without any traceable methodology behind them. We are not going to launder one by citing it here.

What can be said from our own data is narrower and checkable. Retirements in our registry ran 22 in 2023, 55 in 2024 and 106 in 2025 — multiples of 2.5× and then 1.9×. Measured like-for-like from January 1 to September 3, 2026 is running 84 against 69 in the same window of 2025, about 22% ahead. Growing, but nothing like the doubling of the previous two years. We wrote up what that means for the broader argument in when will the AI bubble burst.

The project numbers: 42% abandoned, "more than 80%" failing

These two are about enterprises, not vendors, and they are the ones most often misattributed.

S&P Global Market Intelligence found that the share of companies abandoning most of their AI initiatives before production rose from 17% to 42%, with the average organization scrapping about 46% of proofs of concept before they reached production. That is a real measurement of a real thing: enterprise projects dying between pilot and deployment.

RAND's widely-quoted 80% deserves more care than it usually gets. The line people cite — that more than 80% of AI projects fail, roughly twice the rate of non-AI IT projects — appears in RAND's work as an estimate RAND is citing, not a figure RAND measured. RAND's own contribution was qualitative: interviews with 65 experienced data scientists and engineers about the root causes of failure. That is valuable research, and it is not a survey that produced an 80% number. "RAND found that 80% of AI projects fail" is a sentence that should not be written, and it is written constantly.

The pilot number: MIT's 95%

The most-quoted statistic in AI in 2026 comes from the MIT Media Lab's NANDA initiative, in a report titled The GenAI Divide: State of AI in Business 2025. Its finding is that roughly 95% of enterprise generative-AI pilots produce no measurable return to the profit-and-loss statement, against $30–40 billion of enterprise spending. The method was 150 interviews with business leaders, around 350 employee surveys, and analysis of 300 public implementations.

Three qualifications travel badly and so usually get dropped:

• It measures pilots, not deployments. A pilot is by design a cheap experiment with a high expected failure rate. Reporting that most experiments do not immediately move the P&L is closer to a description of experimentation than an indictment of the technology.
• It measures measurable P&L impact, which is a demanding and specific test. A tool that saves a team six hours a week and was never instrumented shows up as a failure here.
• MIT's own diagnosis is organizational, not technical — what the report calls a learning gap, meaning companies not integrating models into workflows. The statistic is routinely deployed as evidence that the models do not work, which is close to the opposite of what its authors concluded.

Which number answers your question

• "Will the tool I am about to buy still exist in two years?" Use the product number. 10.3% of tracked tools are gone, and the risk is not evenly spread. Across the 38 categories we measure it runs from about 2% to 16% — a sevenfold spread as of September 2026, with Engineering & Simulation safest and Automotive riskiest. The live ranking is in our category failure rates report.
• "If my vendor gets acquired, am I in trouble?" Neither of the enterprise numbers helps. Use the acquisition split: of 150 acquisitions we track, 84 are still shipping under their own brand. Acquisition is the likelier ending and it is usually survivable, at least at first.
• "Will our AI project deliver?" This is where S&P's 42% and MIT's 95% belong, and where our number tells you nothing at all. The determining factors in both studies are organizational — integration, workflow, ownership — not vendor selection.
• "Is the whole thing a bubble?" None of these settle it. They are inputs. Our read on the capital side is in the AI capex bubble.

Where these numbers get misused

Three patterns recur often enough to be worth naming.

Denominator swapping. "95% of AI fails" is the most common formulation, and it silently converts a statistic about enterprise pilots into a statistic about AI generally. The 95% figure says nothing whatsoever about whether a given product works.

Estimate laundering. A number appears as a cited estimate in a reputable institution's report, gets attributed to that institution, and then hardens into a fact with a prestigious name attached. The RAND 80% is the clearest current example.

Undated totals. Cumulative counts grow. Our own registry went from 219 entries on August 2 to 285 on September 3, and most of that jump was verification catching up with older events rather than a month of carnage. A count quoted without an "as of" date will be wrong shortly and will not announce it — which is why every figure we publish carries one.

How we know

Our product-level figures come from a registry we compile and review by hand rather than scrape, because a scraper cannot tell the difference between a dead product and a live marketing site. Roughly half the retirements we confirmed in a recent batch could not have been found by checking whether a website still loads: one retired the consumer app 1.5 million people used while keeping its site up, another reads as healthy on GitHub until you notice the repository is archived.

Counts are live and dated on the AI graveyard, with the full statistical treatment in our AI tool mortality report. The methodology, including how we distinguish shutdowns from acquisitions and from acquired-but-still-operating products, is published alongside it. The third-party figures above are attributed to their originating studies, and where a number is an estimate rather than a measurement we have said so.

Frequently asked questions

What is the AI failure rate in 2026? There is no single figure, because four different things get called the AI failure rate. At the product level, 10.3% of the 2,756 AI tools we track have shut down or been acquired as of September 3, 2026. At the enterprise level, S&P Global found 42% of companies abandon most AI initiatives before production, and MIT's NANDA report found about 95% of generative-AI pilots show no measurable P&L return. These measure products, projects and pilots respectively and are not alternative estimates of one number.

Is it true that 95% of AI projects fail? Not as usually stated. The 95% figure comes from MIT NANDA's The GenAI Divide: State of AI in Business 2025 and refers specifically to enterprise generative-AI pilots showing no measurable profit-and-loss impact. It is not a claim that 95% of AI products do not work, and the report attributes the failures primarily to organizational integration rather than to the technology.

Did RAND find that 80% of AI projects fail? No. The "more than 80%" figure appears in RAND's work as an estimate RAND is citing from elsewhere, not one it measured. RAND's own research on the question was qualitative — interviews with 65 experienced data scientists and engineers about why AI projects fail. Attributing the 80% number to RAND as a finding is a common misattribution.

How many AI tools have actually shut down? As of September 3, 2026, 285 of the 2,756 AI tools in our catalogue have shut down or been acquired: 135 shut down outright, 66 were acquired and sunset, and 84 were acquired but still operate under their own brand. The live count is maintained on our AI graveyard.

Which failure rate should I use when evaluating a vendor? The product-level one, and the per-category version of it rather than the headline. Category failure rates in our data span roughly 2% to 16% across 38 categories — about a sevenfold spread — so the category a tool sits in tells you considerably more than the 10.3% overall average does. The enterprise pilot and project statistics measure your organization's implementation capacity, not your vendor's durability.

Why do AI failure statistics vary so much? Because they use different denominators and different definitions of failure. A product census asks whether a tool still exists. A company survey asks whether a startup survived. A project study asks whether an initiative reached production. A pilot study asks whether an experiment moved the P&L. A tool can pass the first test and be involved in failing every other one, which is why the numbers can all be simultaneously true.

— The ToolDirectory.AI editorial team