AInvasion’s Research desk traces AI claims back to their primary sources (studies, regulator texts, pricing pages) and grades each finding as established, probable, speculative, or framework, so you know how much weight to give it before you repeat it, budget around it, or build on it.
Most AI coverage repeats a number until it becomes a fact. This desk asks where the number came from, what was actually measured, and whether the headline survives contact with the source.
A statistic without a denominator is a rumor.
Last reviewed: 7 October 2026.
| Grade | What it means | How much weight to give it |
|---|---|---|
| Established | Traced to a primary source and consistent across independent evidence | Safe to cite and plan around |
| Probable | Supported by credible evidence with a known gap or limitation | Use it, but state the caveat. |
| Speculative | Plausible, but the evidence is thin or indirect. | Treat it as a hypothesis, not a fact. |
| Framework | Our own analytical model, a way of thinking rather than a finding | Use it to structure decisions, not to prove them. |
Major articles also carry a visible update ledger, so we log corrections publicly. The full process is on our Methodology page.
It is the most repeated number in enterprise AI. We read the report itself (MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025) and ran the figure through seven questions. Use the same seven on any AI statistic before it goes into a slide, a post, or a budget.
| # | The question | What the source shows | Verdict |
|---|---|---|---|
| 1 | Who measured it? | Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari of MIT’s Project NANDA. The report labels itself preliminary findings. | Credible source, early evidence |
| 2 | What was counted? | A review of 300+ public AI initiatives, interviews with representatives of 52 organizations, and a survey of 153 senior leaders collected at four industry conferences. The authors say the figures are directional, drawn from interviews rather than company reporting, and that success definitions differ across organizations. | Small, self-selected sample |
| 3 | What is the denominator? | The same 5% appears in three framings: “95% of organizations are getting zero return,” just 5% of “integrated AI pilots” extracting millions in value, and 5% of enterprise-grade, task-specific tools reaching production after 60% were evaluated and 20% piloted. These are different populations. | The headline merges three denominators. |
| 4 | Original or retelling? | Several retellings headline it as “95% of pilots fail.” At least one notes it is a preliminary finding, not a universal rate. The report also has loose ends of its own: its summary and exhibit cover eight sectors, while its body text says nine. | Read the report, not the headline. |
| 5 | What does “fail” mean here? | Success means deployment beyond pilot with measurable KPIs, with ROI measured six months after the pilot. The report uses both “zero return” and “failure rate” for the same figure and concedes six months may understate success for complex systems. | “Not yet measurable” is not “failed.” |
| 6 | How old is it? | Research window January to June 2025, published July 2025: about 15 months old. | A snapshot, not a current rate |
| 7 | Who has a stake? | The report was produced in collaboration with Project NANDA, which it later names, alongside MCP and A2A, as infrastructure that will shape which vendors succeed. It states the views are the authors’ own. | Disclose it; don’t dismiss it. |
Our grade: Probable for the direction of the finding: in this sample, adoption is high and measurable P&L impact is rare. Speculative for “95%” as a general failure rate for AI projects.
What the report also found, and most headlines skip: over 80% of organizations had explored or piloted general tools like ChatGPT and Copilot, and nearly 40% reported deployment, but these mainly lifted individual productivity rather than P&L. Workers at over 90% of the surveyed companies reported regular use of personal AI tools for work, while only about 40% of the companies said they had bought an official LLM subscription.
Copy this table into your next AI strategy review. Read the full breakdown: Do 95% of AI Pilots Really Fail?
Deferred, not cancelled. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It moved obligations for stand-alone high-risk systems (Annex III) to 2 December 2027 and for AI embedded in regulated products (Annex I) to 2 August 2028, and it added further prohibited practices. Article 50 transparency obligations began applying on 2 August 2026; the content-marking requirement follows on 2 December 2026. Grade: Probable, since the dates are corroborated across multiple law-firm summaries of the Official Journal text. Read this if you build, buy, or publish AI-assisted work that touches the EU market.
AI APIs bill per token, with input and output priced separately, so the same feature can cost very different amounts depending on how much it reads versus writes. Published rate cards also carry levers that move the bill: Anthropic’s pricing documentation, for example, halves token costs for batch jobs and prices cached input reads at a tenth of the base input rate. Read this if you are estimating the real cost of an AI feature. Pair it with the AI Cost Calculator.
The case file above, in full: the primary source, the three denominators, and what the number can and cannot support. Read this if you have ever quoted that figure in a deck, a pitch, or an argument.
See what happens when AI claims meet reality in our AI postmortems, or follow the rules as they change in Regulation Watch.
It is a real finding, but a narrower one than the headline. MIT’s Project NANDA reported in a preliminary July 2025 study that 95% of the organizations in its sample were getting zero measurable return from generative AI. The sample was small and partly drawn from conference attendees, and “zero return” is not the same as “failed.” See the case file above.
Partly. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026, pushed stand-alone high-risk obligations (Annex III) to 2 December 2027 and product-embedded high-risk obligations (Annex I) to 2 August 2028. General-purpose AI obligations have applied since 2 August 2025, and Article 50 transparency obligations began applying on 2 August 2026.
We trace each figure to its primary source and label it “Established,” “Probable,” “Speculative,” or “Framework.” If a number only appears in secondary coverage or vendor marketing, we say so and grade it accordingly. See our methodology.
Major articles carry a visible update ledger that records what changed and why. We correct in the open rather than silently.
Sources read for this page: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025; hosted copy); Regulation (EU) 2026/1744 as reported by White & Case, Cuatrecasas, and EU Law Live; and Anthropic API pricing documentation.
Articles appear under the editorial byline Tom Morgan, an Invasion editorial byline rather than an individual reporter, and are held to one standard: no fabricated statistics, no invented quotes.
AInvasion: Independent AI News, Analysis & Tools. Evidence before certainty.
AI Research — Field Notes Tom Morgan September 8, 2026 · ~11 min read AI conferences are receiving too many submissions for human reviewers to handle, so venues are turning to large language models to draft or supplement reviews — including a conference-wide pilot on nearly 23,000 papers at AAAI-26. Those models agree with each … Read more