A new MIT study suggests the official AI economy is stalling—but a parallel, “shadow AI economy” is thriving. The report, State of AI in Business 2025, from MIT’s Project NANDA, finds that while only 40% of companies have paid subscriptions for tools like ChatGPT or Claude, employees at more than 90% of firms are already using personal AI accounts at work, often without approval from IT.
This reveals what the researchers call the “GenAI divide.” Despite $30–40 billion in enterprise AI spending, only 5% of companies report meaningful financial returns. For the other 95%, big investments have yet to show up on profit and loss statements. Meanwhile, under the radar, workers are quietly integrating consumer AI tools into their daily routines—often with more impact than sanctioned corporate projects.
Why shadow AI is winning
Employees say personal AI accounts are flexible, easy to use, and deliver immediate value. Consumer tools like ChatGPT and Copilot adapt to workflows in ways that rigid, enterprise-level deployments don’t. With low barriers to entry, workers can experiment, iterate, and make AI work for them—no lengthy approval cycles required.
By contrast, official enterprise AI initiatives often get bogged down in integration problems, clunky interfaces, and tools that can’t learn or adapt. Many pilots stall before reaching production, leaving employees to fill the gap with personal subscriptions.
AI’s role in the workplace
The study notes that AI is quickly taking over “simple work.” Seventy percent of workers prefer AI for drafting emails, and 65% use it for basic analysis. But for mission-critical tasks, 90% still trust humans over machines.
This dynamic is shifting employee expectations: once they get used to flexible, consumer-grade AI, they become less tolerant of static, inflexible enterprise tools.
Breaking myths about enterprise AI
The MIT report challenges some common assumptions. It finds that:
-
Few jobs have been directly replaced by AI.
-
AI hasn’t yet transformed business models in any significant way.
-
Most companies have already invested heavily in AI pilots.
-
Failures are less about regulation or raw model power and more about poor adaptability and lack of memory in enterprise tools.
-
“Build” projects—where companies try to develop AI internally—fail about twice as often as “buy” projects using external tools.
The report comes against a backdrop of industry upheaval. Tech sector layoffs, declining wage premiums for college degrees, and stalled enterprise AI initiatives all point to a labor market in transition.
Some observers now wonder if AI is reaching a plateau—especially after the underwhelming launch of ChatGPT-5. The Federal Reserve has even studied the possibility, concluding that AI could boost productivity, but perhaps in a way more akin to the invention of the light bulb: transformative in daily life, but slower than expected to reshape corporate profits.