AI Readiness10 min read

Why 95% of AI Projects Fail (MIT, 2025) — and What It Means for SME Adoption

MIT Project NANDA's State of AI in Business 2025 report found almost all enterprise GenAI pilots deliver zero P&L impact. Here's the actual failure pattern — and why process fit predicts success better than which model you pick.

Haojun See

Founder & Director, On The Ground

Updated 11 July 2026

The headline number, and what it actually measured

In July 2025, MIT Project NANDA — a research initiative at the MIT Media Lab — published "The GenAI Divide: State of AI in Business 2025," built from a review of more than 300 publicly disclosed enterprise generative AI initiatives, 52 structured interviews, and 153 survey responses collected between January and June 2025. The headline finding: despite an estimated US$30–40 billion in enterprise GenAI spending, roughly 95% of organisations reported no measurable impact on profit or loss. Only about 5% of initiatives — described in the report as narrow, workflow-embedded deployments — were extracting real financial value. That's a striking number, and it's been widely cited since. It's also worth reading with the appropriate caveats: this is an industry research report, not peer-reviewed academic work, its full dataset hasn't been published, and some commentators have pushed back on both the methodology and the incentive a report titled around "failure" has to generate attention. We're citing it because its direction is corroborated elsewhere — McKinsey's 2025 State of AI survey separately found that while nearly 90% of organisations report regular AI use, only about 39% report any EBIT impact, and just 5.5% attribute more than 5% of EBIT to AI — not because the exact "95%" figure is beyond question.

The real failure pattern: learning, not infrastructure

The report's more useful contribution isn't the headline percentage — it's the diagnosis. MIT's researchers concluded the core barrier to scaling GenAI wasn't infrastructure, regulation, or even talent shortages. It was learning: most GenAI deployments don't retain feedback, adapt to context, or improve the way a new hire would over their first few months on the job. They get deployed once, in a form disconnected from how the actual workflow runs, and then stay static while the business around them keeps changing. That maps directly onto something we see constantly in SME AI engagements: a tool gets switched on, produces mediocre first-week output because nobody adapted it to house-specific context, and gets quietly abandoned rather than iterated on. The tool wasn't necessarily bad. It was never actually integrated into the workflow it was meant to replace.

Buy vs. build: why vendor tools succeeded three times more often

One of the report's more actionable findings: initiatives that purchased tools from specialised vendors and built implementation partnerships succeeded roughly 67% of the time — about three times the success rate of internally built tools. The reason isn't that vendor-built AI is inherently smarter. It's that a vendor selling a workflow-specific tool has usually already done the integration and iteration work that internal pilots tend to skip, because internal teams treat "get the model working" as the finish line rather than the starting point. For an SME with no in-house AI team, this cuts a particular way: the fastest path to real value is rarely "build something general-purpose ourselves." It's picking (or commissioning) a narrow tool built for one specific workflow, and treating the first version as a starting point to iterate from — not a finished deliverable.

The "GenAI Divide": what the 5% who won actually did differently

The report frames the gap between the 5% who saw real returns and the 95% who didn't as a "divide" — not a matter of budget or model choice, but of approach. The initiatives that worked shared a pattern: they were narrow rather than broad (one workflow, not "AI everywhere"), embedded inside how the work already happened rather than bolted on as a separate tool, and treated as something to keep adjusting rather than switch on and leave alone. None of that requires an enterprise budget. It requires picking the right first workflow and being willing to keep tuning it — which is precisely why our AI Readiness Audit weights the Processes dimension as heavily as it does: how repetitive and well-understood a workflow already is predicts whether an AI tool will actually get embedded into it, or quietly abandoned in month two.

What this means if you're a 10–50 person Singapore business

Three practical takeaways, independent of whether you find the exact "95%" figure convincing: Start narrower than feels ambitious. One workflow, chosen for how repetitive and painful it already is — not a company-wide "AI initiative." Prefer tools built for your workflow over general-purpose ones you'll have to adapt yourself. The buy-vs-build gap in the report reflects integration effort, not model quality — something an SME without an internal AI team should weight heavily. Budget for iteration, not just deployment. The report's core diagnosis is that static tools fail. Plan a review cadence — even a monthly 30-minute check on whether the tool is actually being used and where it's falling short — rather than treating "go live" as the end of the project. If you want to see how this shows up in a structured self-assessment, our AI Readiness Audit scores exactly this — process repetition and fit — as one of its four readiness dimensions, alongside Data, Team, and Governance.

Frequently asked questions

What exactly did the MIT report find?

Reviewing 300+ publicly disclosed enterprise GenAI initiatives, 52 structured interviews, and 153 survey responses, MIT Project NANDA's July 2025 report found that despite an estimated US$30–40 billion in enterprise generative AI spending, roughly 95% of organisations saw no measurable profit-and-loss impact. Only about 5% of initiatives — mostly narrow, workflow-embedded deployments bought from specialised vendors — showed clear financial return.

Is the MIT report peer-reviewed or fully verified?

No — it was released as an industry research report by an MIT Media Lab initiative, not as peer-reviewed academic research, and the underlying dataset has not been fully published. Some commentators have questioned the methodology and noted an incentive for a report about AI failure to attract attention. Treat the specific '95%' figure as a widely cited industry data point rather than an incontrovertible fact, while noting that its broad direction — most GenAI pilots underdeliver — is consistent with other 2025 surveys, including McKinsey's, which found only about 39% of organisations report EBIT impact from AI despite near-universal adoption.

Does buying an AI tool really work better than building one in-house?

The report found tools purchased from specialised vendors and implemented through partnerships succeeded roughly 67% of the time, versus about one-third that rate for internally built tools — largely because vendor tools arrive with the workflow integration and iteration built in, which is exactly what internal pilots tend to skip. That said, 'buy' only outperforms 'build' when the bought tool is actually integrated into a real workflow, not just switched on and left alone.

What should a Singapore SME actually do differently because of this?

Pick one narrow, painful, repetitive workflow rather than a broad 'AI transformation.' Prefer a tool that plugs into how the work already happens over one that requires the team to change how they work to use it. Plan for iteration — the report's core finding is that static tools that don't adapt to feedback are the ones that stall, not the ones using an inferior model. And measure the one workflow you automated, rather than measuring 'AI adoption' as an abstract goal.

PreviousAI Governance for Singapore SMEs: NIST AI RMF, ISO 42001 & IMDA's Framework, Explained Next How Much Does AI Software Development Cost in Singapore? (2026)

Want to Apply This to Your Business?

We're a Singapore AI development and automation agency. Let's discuss how we can help solve your specific challenges.

Get in Touch View our services