Hype Alert: "95% of GenAI Projects Fail" Is a Fake Headline

Well.. and here we go again. LinkedIn just had one of its signature collective meltdowns over a single statistic: "95% of GenAI projects fail." You know the drill—someone posts a scary number, everyone shares it without reading the fine print, and suddenly we're all living in an AI apocalypse.

I did what apparently nobody else bothered to do: I actually read the MIT report that started this whole frenzy. Plot twist—it doesn't say what everyone thinks it says. Not even close.

This isn't your typical "the sky is falling" AI story. It's a valuable example of how social media can turn nuanced research into viral doom-scrolling. But let me walk you through what really happened, and more importantly, how to actually make AI work in your organization.

What the Report Actually Says (Upon closer reading)

Here's where things get interesting: that hyperbolic "95%" number is not what some interpretations want you to believe.

The real story buried in the MIT-NANDA report is way more specific. They looked at custom, task-specific enterprise AI tools—think bespoke chatbots built for your customer service team, not someone using ChatGPT to write emails. Of those custom tools, only about 5% made it to production and showed measurable business impact within six months.

But here's the kicker: the report explicitly calls this data "directionally accurate" rather than gospel truth. It's based on 52 interviews and 153 surveys—a helpful signal, but hardly the definitive census of AI failure that LinkedIn seems to think it is.

The report also does something fascinating that got utterly lost in the headlines: it separates general tools (like ChatGPT and Copilot) from custom enterprise solutions. About 80% of organizations tried the general tools, and 40% actually deployed them. That's... not exactly a failure story.

The six-month measurement window is particularly telling. Even the researchers admit this might be too short to judge complex organizational change. It's like declaring a fitness program useless because you don't see abs after two weeks of going to the gym.

And then there's my favourite finding that nobody's talking about: "shadow AI." The report found that about 90% of employees are using personal AI tools (ChatGPT, Claude, etc.) for work, while only 40% of companies have official AI subscriptions. That gap isn't failure—it's a massive hint about where value already lives.

Why the Headlines Got It So Wrong

This is classic telephone game meets social media amplification. Here's how "5% of custom enterprise tools reached production within six months" became "95% of all AI projects fail":

• Category confusion: Lumping everything together makes for better clickbait but worse understanding. A low success rate for custom, embedded tools doesn't mean your marketing team isn't getting value from ChatGPT.
• Time compression: Six months is nothing in enterprise-change time. I've seen email migrations take longer than that.
• ROI tunnel vision: The report equates "no sustained P&L impact" with failure. But anyone who's worked in operations knows that's not how transformation works. The early wins show up as faster cycle times, fewer errors, and reduced external spending—all before they hit the P&L statements.

It's like judging a construction project by whether the walls are painted while the foundation is still being poured.

What I See in the Real World

I run workshops with public sector teams and digital leaders, and I've built my own AI tools (including a "Public Sector Career Navigator" that actually helps people). Here's what I observe over and over:

• Pilot theatre is easy. Integration is hard. Everyone loves a good demo. Getting AI actually to stick in daily workflows? That's where things get interesting. Most organizations try AI where it looks flashy (customer-facing stuff) but find success where it's frequent (back-office tasks).
• The shadow AI phenomenon is real—and it's your friend. People are already using ChatGPT and Claude for work, often getting better results than the "official" tools their companies bought. Instead of fighting this, smart organizations study what their power users are doing and build on those patterns.
• Boring wins. Every time. The sexiest AI implementations aren't in sales pitches or marketing copy. They hide in boring organizational corners, such as claims processing, document routing, and data reconciliation. You know, the stuff that happens so many times a day; even a slight improvement brings real value.

The Actually-Useful Playbook for AI Success

If you want results that survive the next budget cycle, here's what actually works:

• Start Where Work Repeats: Forget the flashy stuff. Look for high-frequency, measurable workflows: claims, reconciliation, routing, document prep. These tasks show up constantly, which means compounding value over time. It's not glamorous, but neither is profit.
• Measure Before You Automate: Baseline your current performance—cycle time, error rates, external spend. Then roll out AI to a matched team and compare. "Feels faster" is not a business case.
• Stage Rollouts Like Science Experiments: Use A/B testing or staggered launches. Keep a control group. Measure at month one (usability) and month six (business impact). The report's six-month window is tight—plan for 6-12 months for real process change.
• Buy First, Build Maybe: Start with existing tools that can integrate into your workflow. The report shows higher success rates for external partnerships versus internal builds. Take the hint. You can always build custom solutions later when you actually understand what works.
• Legitimize Shadow Usage: Ask your power users how they're already using AI. Document those patterns, then productize them with proper data controls. That 90/40 split between personal usage and company subscriptions? That's your free R&D department.
• Prioritize Learning Systems: Choose tools that retain context, adapt to feedback, and improve over time. The report identifies this as the key differentiator between pilots that stall and tools that stick.
• Chase External Spending, Not Headcount: You'll find cleaner ROI by reducing contractor and agency costs than by promising layoffs. The report's most precise dollar figures are in BPO savings and reduced external spend.

Read the Fine Print, Not Just the Headlines

The MIT report is actually quite helpful if you read beyond the viral soundbites:

• The pilot-to-production chart shows the 5% figure and flags the research limitations right next to it.

• The shadow AI data reveals where employees are already finding value.

• The functional allocation visuals show that most money goes to sales and marketing, while the ROI examples come from back-office operations.

• The appendix explicitly warns that six months may understate success for complex systems.

• This isn't a failure story—it's a roadmap disguised as a cautionary tale.

Anti-Hype, Pro-Progress

So yes, most organizations are stuck in the messy middle between pilot theatre and real integration. But the solution isn't to declare AI a failure and move on. The goal is to improve at the process of proving value.

GenAI isn't failing—our approach to implementing it is just immature. We're still figuring out how to measure success, where to focus effort, and how to turn experiments into operations.

The path forward is practical, not mystical. Start where work repeats. Measure what matters. Learn from shadow usage. Partner your way past the first mile.

And maybe, just maybe, read the actual reports before sharing the scary headlines. Your LinkedIn feed will thank you.

If this resonates, follow along for more field notes from the trenches of practical AI implementation—no snake oil, no magical thinking—just honest exploration of what actually works.

Source: "State of AI in Business 2025" (MIT Project NANDA). I actually read it, so you don't have to suffer through the academic prose.

Originally published on my blog [no paywall] where you can also download a PDF copy of the original MIT report.