Mental Fast Food™ · 04

The 95% Myth

Interference topic · The statistic isn't about the technology

What if a statistic is not about the technology at all — but about where the money went?

01 · The claim you swallowed

"95% of AI projects fail."
A real headline. Not the verdict everyone read into it.

95%

The MIT number is solid: 95% of enterprise AI pilots show no measurable return. The internet read that as "AI is hype." The study says almost the opposite. One click shows why.

Source

MIT NANDA · State of AI in Business 2025

Behind it

EQUORA Institute · measured, not estimated

Updated

June 2026

02 · The question you didn't ask

That 95% answers a question. Almost nobody checked which one. Here are two. The first is what everyone assumed. The second is what the data actually measured. Same data. The question decides what you see in it.

The question the number answers

Does AI actually work?

The question that matters

Where did the 95% get deployed?

Try this interaction

The axes flip. The failure relocates. ↓ Same data. Different question. Different answer.

03 · Why the question matters

Ask "does it work?" and you
blame the tool for the aim.

MIT was blunt: the failures aren't driven by model quality. They're driven by approach — where the money went and how it was deployed. Over half of AI budgets pour into sales and marketing, where returns are weakest, while the quiet back-office wins stay underfunded. Ask "does AI work?" and you keep funding the visible, low-value zone.

This is the skill the page is really about: when a number lands, the first move isn't "true or false?" It's "what question is this the answer to?" A real statistic, aimed at the wrong question, becomes a false conclusion.

04 · The current answer, in full

Where the money goes vs. where the value is

300+ enterprise AI deployments · MIT NANDA 2025 · bubble size = share of budget

Axes right now: does AI "work"? (yes / no)

50–70% of the budget goes to sales & marketing AI — the visible, easy-to-measure stuff. That's where most of the 95% "failures" sit.

5% that succeeds sits mostly in back-office automation — unglamorous, underfunded, and where the real ROI hides.

2× the failure isn't the model. Buying from specialised partners succeeds ~67% of the time; internal builds ~33%. It's about where and how, not whether AI works.

↑ Ask the right question above to flip the axes.

05 · So what do you do differently

So if you're deciding where to use AI, the question isn't "is it good?" — it's "where does it actually create value?"

Two axes sort it fast: does the task create real value, and can AI carry it without a human in the loop? The order of what you ask matters more than how much you spend.

1 Start from value, not from hypePick the task by the value it creates, not by how impressive an AI demo looks. The flashiest use cases (marketing copy, chatbots) are often the lowest-value ones.

2 Prefer where humans aren't the bottleneckAI compounds value where it runs end-to-end — back-office automation, document processing — not where it just drafts something a human still has to finish and check.

3 Buy focused, don't build broadSpecialised partner systems succeed about twice as often as internal builds. "Let's build our own AI platform" is where a lot of the 95% quietly died.

Four things the headline skips

"95% fail, so AI doesn't deliver."

The 5% that succeed extract millions in real value. The failure is concentrated where AI was aimed badly, not spread evenly. It's a targeting problem, not a technology one.

"Put the budget where AI is most visible."

50–70% of budgets go to sales & marketing — the lowest-ROI zone. The biggest returns hide in back-office automation: cut outsourcing, document review, risk checks. Visible and valuable are not the same axis.

"Build your own AI platform to own it."

Internal builds succeed ~33% of the time; buying from specialised vendors ~67%. Twice the success rate — because partners bring workflow fit and adoption, the parts that actually decide ROI.

"The official corporate AI tool is where the value is."

A "shadow AI economy" runs underneath: estimates suggest ~90% of workers use personal AI tools daily, often outperforming the sanctioned systems. The real adoption already happened — just not where the budget looked. Note: this figure varies widely by study and definition — nobody has measured it cleanly yet, which is itself the point.

06 · The next question

So if "95% fail" really means "95% were aimed wrong," how would you measure whether your own AI use sits in the value zone — before you've spent a year finding out?

Here's the gap: the MIT study measured this for big enterprises, after the fact. For a small team, a school, a solo maker, nobody has built a simple, measured way to tell value from hype before you commit. Most "AI readiness" advice is vibes. The Institute's line holds here too: measure, don't estimate.

Proposed research · EQUORA Institute · measured, not estimated

Help build it → A measured value-vs-hype test small teams can actually use. Researchers and contributors welcome, alongside the AI-first team. No fee, no catch.

Scientific background

AI productivity and the 95% figure — the research landscape

MIT NANDA / MLQ.ai (2025) — The GenAI Divide: State of AI in Business 2025. 300+ deployments, 52 interviews, 153 leaders. Finding: 95% of enterprise GenAI pilots show no measurable P&L impact; 5% of integrated (not pilot-stage) deployments extract significant value. The key distinction the headline obscures. PDF →
Brynjolfsson, E., Li, D. & Raymond, L.R. (2023) — Generative AI at Work. NBER Working Paper 31161. RCT with 5,179 customer service agents: 14% average productivity increase; largest gains for novice workers, smallest for experienced workers — consistent with AI as a skill-equaliser, not a skill-multiplier. nber.org →
Acemoglu, D. (2024) — The Simple Macroeconomics of AI. NBER Working Paper 32487. Skeptical counter to optimistic productivity forecasts: estimates only 4.6% of tasks currently exposed to AI will be cost-effectively automated within 10 years, implying GDP impact of ~0.5% over a decade. nber.org →
Eloundou, T. et al. (2023) — GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. OpenAI / OpenResearch. 80% of US workers have at least 10% of their tasks exposed to LLMs; 19% have 50%+ exposed. Exposure ≠ displacement — but maps the surface area. arXiv:2303.10130
Papp, L. (2026) — Why Trust Scores Fail. Zenodo preprint. Argues that any scalar trust score violates Arrow's Impossibility Theorem (no aggregation rule can satisfy all rational preference axioms simultaneously) and is subject to Goodhart's Law (when a measure becomes a target, it ceases to be a good measure). Proposes proof-based verification as the only epistemically defensible alternative. doi:10.5281/zenodo.20772213 · ORCID: 0009-0005-6329-5808

Research provenance

This page comes out of research at the EQUORA Institute and captures one state of that work rather than a settled institutional position. That state rests on the findings available at the time of publication; later findings appear here only where the page has been updated, which the date shows. AI takes part throughout the research process as a thinking partner; responsibility for interpretation and publication remains human.

Kiadva: 2026. augusztus 1.Published: 1 August 2026

Papp László · EQUORA InstituteHow We Research →