Enterprises love to talk about their AI dreams. But MIT’s new State of AI in Business 2025 report basically says: dreams are cheap, execution is rare.
Here’s the tea: after 153 exec interviews, 350 employee surveys, and an analysis of 300 deployments, MIT found that only 5% of generative AI pilots actually made it to production with real revenue growth. Five. Percent.
The other 95%? Stuck in pilot purgatory, collecting dust while someone in the boardroom insists it’s “strategic.
And the divide between winners and losers couldn’t be sharper:
On the winning side, you’ve got scrappy startups (yes, some run by teenagers 🤯) going from zero to $20M in a year. How? They pick one pain point, go all-in, and partner smart. No fuss, no overthinking.
On the other side (the losing side), are big enterprises running dozens of “innovation pilots” that never make it past PowerPoint.
And the kicker? The gap isn’t about model quality. The tools themselves are fine. The issue is the learning gap— most enterprises simply don’t know how to fold these systems into messy workflows, so they stall.
Another brutal finding:
Most companies are throwing the bulk of their genAI budgets on sales + marketing tools (the shiny stuff). Meanwhile MIT’s data shows the real gold mine is in the boring back-office automations: those are tools that cut outsourcing, eliminates agency costs, and streamlines operations.
Basically, all the unsexy plumbing work nobody brags about—but where the ROI is massive.
This one’s hilarious. Enterprises love to “own” their own AI systems, especially in finance and other regulated spaces. But the data says otherwise: purchased AI tools succeed twice as often as homegrown builds. Twice. Yet big firms keep choosing the harder, more failure-prone path of “we’ll build it ourselves.” Because apparently “reinventing the wheel” is a corporate love language. (anyone thinking of Apple??)
And as for how AI is affecting jobs and people, the report suggests that AI isn’t causing Hunger Games-style layoffs (yet). What’s happening is quieter:
Roles in admin and support aren’t getting replaced once people leave.
And in the background? Employees are sneaking ChatGPT into workflows—even if the company pretends it’s “unauthorized.”
Coming attraction:
The cool kids are already experimenting with agentic AI—basically AIs that don’t just answer questions but learn, remember, and act.
Our Take:
This report isn’t really about AI failing. It’s about enterprises being enterprises—slow, bloated, addicted to flashy projects, and allergic to simplicity.
Startups win because they stay scrappy. Enterprises lose because they complicate themselves to death.
The playbook’s obvious:
Focus on one real pain point
Automate the boring but lucrative stuff
Buy smart instead of building bad
Let managers (not labs) drive adoption.
Until then? Expect more press releases, more pilots, and not very much profit.
You should read the full report.
