I spent nearly a decade as an intrapreneur inside the world’s largest global holding companies. On paper, it looked a lot like entrepreneurship: validate an idea, conduct research, raise or allocate funds, build capabilities, codify processes, launch SaaS platforms, measure value creation, and implement a communication plan.
In practice, it was very different. Big organizations are optimized for productivity and predictability, not the full lifecycle of experimentation that product building requires. That law of nature creates a constant source of friction between innovation and day-to-day business.
A new MIT study puts numbers to what many of us have experienced: 95% of enterprise GenAI pilots fail to deliver measurable business impact, despite billions invested. The problem is less about model quality and more about the learning gap: Tools and organizations do not naturally adapt to one another, so in-house pilots never become production systems.
MIT and other researchers highlight consistent fault lines:
- Flawed integration: Pilots sit on the side and never embed into real workflows. The companies that do see impact redesign processes and roles around AI rather than sprinkling models on top.
- Learning gaps and culture: Organizations treat AI like a one‑off project, not an evolving capability, so teams do not learn with the tools.
- Misallocated budgets: Spending skews to sales and marketing experiments while the highest ROI is often in back‑office automation that reduces outsourced processes and eliminates manual work.
- Build versus buy: Buying from specialized vendors and partnering works about 67% of the time, compared to internal builds succeeding roughly one‑third as often.
- Shadow AI risk: employees use personal chatbots at most companies, which muddies impact measurement and raises compliance risk. Reports find widespread unsanctioned use.
These patterns are not unique to AI. I saw the same dynamics at play when launching products within corporations long before the AI wave became the center point of the software conversation. The code is never the blocker to success. It’s all about incentives. Billable hours and short‑term deliverables are naturally at odds with the patience, rework, and staged learning a product needs. Without a protected path from pilot to scale, even strong concepts suffocate in a productivity‑first culture.
Context from prior waves reinforces this current moment in time: an MIT Sloan–BCG study found only about 10% of organizations realized significant financial benefits from AI, with success tied to how well humans and AI learn together. A year later the research emphasized that organizations capture value when individual workers also feel empowered and gain competence and autonomy from the tools. Even now, adoption at scale remains limited: One recent, large CIO survey reported only 11% had fully implemented AI due to security and data readiness constraints.
What successful programs do differently
The efforts that work do not live as science projects. They integrate early, align incentives with outcomes, and earn trust on the front line. They move quickly from test to tool. The playbook looks like this:
- Start with a workflow, not a model. Redesign the process where the decision happens, then fit AI to it. Treat AI as infrastructure that changes who does what and when.
- Pick one painful, measurable problem. Scope narrowly, ship a useful tool, and iterate in place. Tie success to a business owner’s KPI. The MIT study notes that the winners execute against specific pain points rather than broad ambitions.
- Choose to build, buy, or partner with discipline. If time‑to‑value matters, lean into vendors with proven outcomes, then extend. The success gap between vendor solutions and internal builds is material.
- Shift investment to the quiet ROI. Target back‑office and operational automation where savings are concrete and compounding. Use those gains to fund the next wave.
- Make learning a first‑class objective. Pair tool learning with organizational learning: training, job design, accountability, and feedback loops.
- Bring shadow AI into the light. Set clear guardrails, offer approved tools, and measure use so value shows up in the P&L instead of slipping through side channels.
The takeaway here is not that AI is overhyped; it is that experimentation without integration rarely creates transformation. Leaders who treat AI like infrastructure, align incentives to outcomes, and build learning into the operating model will escape the pilot trap. The rest will keep adding to the graveyard.
James Chester is cofounder and CEO of WVN.