The Problem Is Not Technology But ''AI That Cannot Learn''
An MIT research team''s July report "The GenAI Divide: STATE OF AI IN BUSINESS 2025" reveals that despite companies worldwide pouring billions into AI, only 5% are achieving actual results. Researchers called this phenomenon the "GenAI Divide" — most companies have adopted AI but are stuck at the pilot stage without creating change.
Over 80% of companies have tried general-purpose AI like ChatGPT or Copilot, but cases achieving results visible in profit and loss statements are extremely rare. Large companies have many pilots but low actual adoption rates. Medium companies make fast decisions and often complete from pilot to full adoption in an average of 3 months. An interesting finding: while company-level AI projects stagnate, employees are already actively using AI in daily work — corporate official LLM subscription rate only ~40%, but individual employee ChatGPT/Claude usage exceeds 90%. Researchers named this the "Shadow AI Economy."
Why doesn''t corporate AI adoption lead to results? The conclusion: AI cannot remember context and cannot learn from feedback — the biggest problem. A lawyer who bought an AI contract analysis tool for tens of thousands of dollars still uses ChatGPT in actual work: "The specialized tool was rigid and inflexible. ChatGPT gives answers the way I want. But I can''t use it for important contracts — it repeats the same mistakes." AI is welcomed for short-term tasks like simple summaries or email drafting, but humans are overwhelmingly preferred for complex projects or important decisions.
What made successful companies different: buying rather than building in-house had twice the success probability; starting with small successes (narrow tasks like contract summaries and call center log organization) spread quickly; real results came from back-office work not visible marketing — some companies reduced customer center outsourcing costs by millions, cut advertising agency costs by 30%. Researchers emphasize the next 18 months will be a critical watershed. "Agentic AI" that remembers and learns — like MS Copilot or ChatGPT memory features — is already emerging. The "Agentic Web" where AIs connect and cooperate could transform business processes from human-coordinated to AI self-negotiating/connecting/executing network economies. The key differentiator between 95% failure and 5% success was not technology size but learning ability and adaptability.
