Writing / Essay
The AI Value Gap
Why 95% of enterprise AI produces nothing, and it is not the technology.
MIT’s 2025 “State of AI in Business” report puts a number on the question boards keep asking: 95 percent of enterprise AI initiatives produce no tangible bottom-line impact.
95 no tangible bottom-line impact 5 moved the number
I have spent years inside companies that sit inside that 95 percent. The pattern is always the same: a seven-figure contract, a steering committee, a launch email, and then quiet abandonment. The report calls this the “GenAI Divide”. Most companies are stuck on the wrong side of it, running a playbook that does not work, while a small 5 percent on the other side follow a clear, replicable approach. The divide is not about the technology. It is about deployment.
A tale of two AI economies
The 95 percent is not one number. It applies specifically to the rigid, custom-built, top-down enterprise AI tools, the large expensive projects meant to plug directly into core processes. The implementation funnel for these systems is brutal. MIT’s data shows that while 60 percent of organizations evaluated such tools, only 20 percent reached a pilot, and 5 percent achieved full production with a demonstrable P&L impact.
- 60% of organizations evaluated a custom, top-down enterprise AI tool
- 20% reached a pilot
- 5% reached full production with a demonstrable P&L impact
Inside those same organizations, a second economy is delivering real value. 78 percent of AI users bring their own tools to work. Employees at over 90 percent of companies use personal AI for their jobs even where the company has no official subscription, and studies put the individual productivity gain from these general-purpose tools at 14 to 56 percent.
That is the central business value gap. Most organizations run two parallel AI economies: a failing, top-down enterprise initiative and a succeeding, bottom-up shadow-AI movement. The strategic question is why the formal model keeps losing to a $20-a-month chatbot.
Anatomy of failure: the learning gap and the build trap
The research pins down two reasons for the 95 percent.
The first is what MIT calls the learning gap. Most enterprise tools are static: they are deployed and never learn from user feedback. Users describe them as “brittle” and complain of “too much manual context required each time”. A system that makes the same mistake twice is not intelligent, which is why 90 percent of users still prefer a junior human, who learns, over a typical AI tool, which does not. I have sat in those user interviews. The tools are not hated. They are simply not used.
The second is the build trap: the belief that a proprietary model is a competitive advantage. The data flatly contradicts it. Internal builds fail twice as often as projects that leverage external partnerships. The success rate for a buy-and-integrate strategy is 66 percent, against 33 percent for building from scratch. Companies burn capital reinventing a wheel that specialist vendors already perfected.
Lessons from the field
Colgate-Palmolive shows the buy-and-adapt route. The company partnered with Microsoft to apply an AI platform to decades of internal research. It did not build a new model; it aimed an existing one at its own proprietary data, and cut insight-generation time from weeks or months down to minutes.
Klarna shows the cost of speed. The fintech went all-in on an AI assistant doing the work of 700 customer service agents, projecting a $40 million profit improvement. The first numbers looked impressive. Then customer satisfaction dropped 22 percent, and CEO Sebastian Siemiatkowski admitted the error: “cost... seems to have been a too predominant evaluation factor.” Klarna is now re-hiring for a hybrid model. Moving too fast in human-centric roles destroys value. It is the one case where I have seen a CEO say the quiet part out loud.
A pragmatic playbook
The 5 percent that succeed are not lucky. They follow a different playbook, pragmatic, decentralized, focused on integration rather than invention. Four mandates come out of their success.
- Buy and partner, don’t buildMandate 01
- Partnering with an expert vendor doubles the chance of success, and the financial case matches: a custom build can top $8 million, an enterprise license starts around $100,000. Treat AI vendors as strategic service partners, not software sellers, and hold them accountable for business outcomes.
- Empower the edge, not the centerMandate 02
- The best AI ideas are not coming from centralized innovation labs. They come from frontline managers and power users already solving real problems with consumer tools. Find those bottom-up solutions, govern them, and fund them for scale.
- Mandate learning systemsMandate 03
- Make it a non-negotiable technical requirement. An AI tool that cannot learn from user feedback is not fit for a dynamic business. Prioritize flexibility, an 80 percent concern among successful firms, over raw performance: demand memory, feedback loops, and access to current, proprietary data.
- Redesign roles for career accelerationMandate 04
- A 2025 Harvard study provides the frame: AI adoption slows hiring for junior roles but increases promotions for existing junior staff, because the entry-level grunt work is automated and current team members build senior skills faster. Reframe the transformation from job replacement to career acceleration, and build the internal paths to promotion.
The 95 percent are not failing because AI is a bad bet. They are failing on strategy. Success is not the most advanced algorithm. It is the smartest integration of a learning capability into the organization.
Most numbers and quotes here come from the MIT study, The GenAI Divide: State of AI in Business 2025. It is worth reading in full.
