https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
Earlier this month, an MIT-affiliated study on the “State of AI in Business 2025” sent shockwaves through the market with one bold claim:95% of enterprise AI pilots fail.Headlines amplified the number. Investors panicked. Commentators rushed to call AI a bubble.
But when you unpack the methodology and the findings, the picture is far less alarming and much more familiar to anyone who has lived through previous waves of enterprise technology adoption.
Methodology Matters: Thin Data, Big Claims
The report is built on:
153 survey responses from conference attendees.
52 structured interviews with “stakeholders” (not necessarily executives with budget authority).
300 “publicly disclosed AI initiatives” — mostly press releases.
That’s not a comprehensive dataset. It’s directional at best. And it’s important to remember: “stakeholders” can mean anything from analysts to project managers. The study doesn’t tell us which level of decision-makers were interviewed, how success was defined, or what KPIs were actually tracked.
In other words, it’s anecdotes wrapped in academic branding. Useful for spotting themes, but hardly the statistical foundation for a sweeping claim like95% failure.
The Reality of Enterprise IT: Slow, Risk-Averse, Process-Driven
The study’s big “discovery” is that most AI pilots stall before production. But anyone who has sold enterprise software knows this is the norm, not a shock.
Corporations move on quarterly budgets and annual planning cycles.
Even minor API fixes can take months.
Managers are rewarded for predictability, not risk-taking.
Reorgs and shifting priorities regularly derail initiatives.
Against that backdrop, it’s not surprising that most AI tools don’t make it to production quickly. Enterprise adoption curves are always messy. Remember cloud in the early 2010s, or mobile enterprise apps before them.
So yes, the failure rate is high — 60%, 70%, maybe even 75%. But95%?That feels exaggerated, engineered more for a headline than an accurate picture of the adoption curve.
What the Report Accidentally Confirms
Strip away the drama, and the findings actually line up with what thoughtful observers already see:
Consumer tools win attention.ChatGPT and Copilot are widely adopted because they’re simple and flexible.
Enterprise tools lag.Custom deployments stall because they don’t integrate well into workflows, lack memory, and require change management.
Shadow AI is real.Employees are using personal AI tools at work every day, even when their companies’ official pilots fail.
None of this means AI is a bubble. It means AI is running into the same enterprise adoption bottlenecks that every prior wave of technology faced.
The Path Forward: Vertical AI Will Win
The real story is not about failure. It’s about where success is emerging — and where it will scale.
The next generation of winners won’t be generic LLM wrappers. They will bevertical AI systemsthat combine:
Domain expertise— understanding how a specific industry works at a process level.
Workflow integration— embedding into the tools and systems employees already use.
Unique data access— leveraging proprietary, context-rich datasets that compound in value over time.
That’s where we’re already seeing ROI:
Back-office automation replacing BPO contracts.
Finance, procurement, and risk management tools reducing external agency spend.
Customer retention improvements in industries where AI augments, rather than replaces, human workflows.
These aren’t headline-grabbing demos. They’re the quiet, workflow-specific wins that create durable enterprise adoption.
Conclusion: Ignore the Noise, Watch the Fundamentals
The MIT study is useful for identifying friction points, but the headline number was clickbait. It spooked the market without adding much that operators didn’t already know.
The truth is we’re still early. Enterprise AI is in the messy middle of its adoption cycle. Just as cloud, mobile, and SaaS before it, the pattern is familiar: pilots everywhere, production slow, eventual breakthroughs where technology aligns with workflows and data.
The winners will be thevertical AI vendorswho know their industry cold, integrate deeply, and learn continuously. The losers will be the generic wrappers chasing top-line hype.
So no, AI isn’t a bubble. It’s a long game. And we’re only in the early innings.
Would you like me tolayer in a few hard stats(e.g. “80% of companies piloted ChatGPT, but only 5% of enterprise tools reached production”) so this feels like adata-driven rebuttalinstead of just a narrative pushback?
You said:
750 words ish with more data, professionally written and impressive.
ChatGPT said:
Got it — here’s a~750-word professional newsletter draft, fully fleshed out with data, structured logically, and written to impress LPs, founders, or senior operators.
Why the “95% of AI Pilots Fail” Report Is a Nothing Burger
Earlier this month, MIT-affiliated researchers released theirState of AI in Business 2025report with one headline-grabbing claim:95% of enterprise AI pilots fail.The market reacted instantly. AI stocks dipped, commentators declared we were in a bubble, and skeptics found fresh ammunition.
But when you peel back the methodology, the number falls apart. The report is not a comprehensive data analysis. It’s a collection of anecdotes, surveys, and interviews repackaged into a sweeping conclusion. And while the study highlights real challenges in AI adoption, the panic it triggered misses the broader truth: enterprise AI is simply following the same messy adoption curve as every major technology before it.
Thin Data, Big Claims
The report is built on:
153 survey responses gathered at industry conferences.
52 interviews with “stakeholders,” an undefined term that could mean anyone from an analyst to a VP.
300 “publicly disclosed” AI initiatives — press releases, vendor blogs, and announcements.
That’s not a representative dataset. It’s directional, not definitive. Nowhere does the study clarify which executives were interviewed, how success was measured, or what KPIs were actually tracked. Did interviewees own budget and P&L, or were they project participants reporting secondhand impressions? The report doesn’t say.
So while the 95% number makes for a striking headline, it’s not grounded in rigorous statistical evidence. It’s “crap in, crap out” data — weak inputs leading to exaggerated outputs.
To anyone who has sold enterprise software, the study’s “discovery” is not surprising. Of course most pilots fail. That’s how enterprise adoption works.
Budgeting cycles are slow.Most corporations review budgets once or twice a year. Big changes get caught in reorgs, approvals, and quarterly reviews.
Processes are entrenched.Even fixing a minor API can take months. Expecting AI to suddenly integrate into brittle legacy systems overnight is unrealistic.
Incentives discourage risk.Corporate managers — lifers with quarterly performance reviews — are rewarded for stability, not for pushing untested technologies into production.
Against that backdrop, it’s obvious why pilots stall. Adoption isn’t blocked by model quality or regulation, as the report implies. It’s blocked by organizational inertia and the lack of systems thatlearn, remember, and adapt.
Strip away the hyperbole, and the report actually reinforces what we already know about GenAI adoption:
High adoption, low transformation.Over 80% of organizations have piloted ChatGPT or Copilot. Nearly 40% report some deployment. But these tools primarily enhance individual productivity, not enterprise P&L.
Enterprise solutions lag.60% of firms evaluated custom or vendor AI systems. 20% reached pilot stage. Just 5% reached production. That’s not an AI problem — it’s the same pilot-to-production chasm that cloud, mobile, and SaaS faced in their early days.
Shadow AI is real.While only 40% of companies purchased an official LLM subscription, 90% of employees report using personal tools like ChatGPT or Claude for daily work. Real transformation is already happening — just outside of IT’s official channels.
These aren’t signs of a bubble. They’re signs of a technology entering its second phase of adoption.
The study paints a bleak picture, but the reality is more nuanced. Early failures don’t mean AI is a dead end. They mean the first wave of adoption generic tools and wrapper applications isn’t enough. The next wave will be led byVertical AI vendorswho solve real business problems with specificity.
The evidence is already emerging:
Sales & Marketing toolscaptured ~50% of GenAI budgets, but most returns came fromback-office automationeliminating BPO contracts, reducing external agency spend, and streamlining finance and procurement.
Best-in-class organizationsreport $2–10M in annual savings by replacing outsourced document processing and customer support with adaptive AI agents.
High performersmove faster. Mid-market firms went from pilot to production in 90 days on average. Large enterprises took nine months or longer.
The pattern is clear: the winners will be the vendors that combine:
Domain expertise— deep understanding of industry-specific workflows.
Workflow integration— embedding AI where people actually work.
Unique data access— proprietary datasets that compound in value over time.
That’s where the real ROI lies. Not in flashy demos or top-line productivity claims, but in targeted deployments that deliver measurable savings and durable adoption.
Big Desk Energy | Subscribe
startup insights, stories, and vibes sent to your inbox every Tuesday
mail.bigdeskenergy.com/subscribe?_bhba=a912eba6-7a35-4c1b-a9cb-9721b5c72389
Venture Scout
High-quality software startups delivered straight to your inbox, every Wednesday.
www.venturescout.io/subscribe?_bhba=a912eba6-7a35-4c1b-a9cb-9721b5c72389
There’s only one place where CS leaders at the cutting edge will gather to explore the incredible opportunities presented by AI Agents:Pioneer.
Pioneer is a summit for AI customer service leaders to come together and discuss the trends and trajectory of AI and customer service. You’ll hear from innovators at Anthropic, Toast, Rocket Money, Boston Consulting Group, and more—plus a special guest keynote delivered by Gary Vaynerchuk.
You’ll also get the chance to meet the team behindFin, the #1 AI Agent for customer service. The whole team will be on site, from Intercom’s PhD AI engineers, to product executives and leaders, and the solutions engineers deploying Fin in the market.
Join Pioneer in-person or online on October 9th, 2025. Register today.
Or copy and paste this link to others:{{rp_refer_url_no_params}}
© 2026 Trace Cohen's Vertical Ai Investor Newsletter
1 Unicorn Ranch
Port Washington, New York 11050, United States