The Shifting Sands of IT Authority

The most striking finding from MIT’s GenAI Divide study wasn’t that 95% of formal AI pilots fail to deliver measurable returns—though that’s concerning enough. It was that workers at over 90% of these organizations were already using personal AI tools for work, without approval or oversight.

This isn’t a security failure, though it is one aspect. It’s a deeper verdict on how technology value is being created and consumed. When employees bypass formal processes to adopt tools they find valuable, it suggests those tools address real needs that existing systems aren’t meeting.

The Defense of Caution

The common explanations for slow AI adoption—data sovereignty concerns, model drift risks, vendor lock-in worries—feel incomplete when viewed through this lens. These are legitimate issues, yet they don’t fully explain why organizations accept greater risks in areas where the potential upside is clear to executives.

Philip Tetlock’s research on accountability offers insight: when decision-makers know their audience (like an executive committee), they tend to conform rather than challenge—using what he called the “low-effort acceptability heuristic.” CIOs, incentivized to be defensible rather than necessarily right, often favor caution.

The cost of inaction is invisible in variance reports or attributed to past decisions, lacking a clear owner or consequence. Meanwhile, visible risks carry greater weight.

When Expertise Depreciates

For decades, the technology function’s authority rested on scarcity—the knowledge and skills the business couldn’t replicate. But AI is changing this fundamental dynamic. When non-technical staff can build prototypes, analyze data, or produce results once requiring specialized expertise, that foundation erodes.

This creates a defensive routine where caution feels like diligence rather than protectionism—a phenomenon Chris Argyris termed the “undiscussability of the undiscussable.” Few CIOs would consciously admit to slowing down adoption because it devalues their domain knowledge.

The Distribution Compression

The impact isn’t uniform. Studies show AI assistants boost productivity more for novice users than experienced ones—compressing the skill distribution and diminishing the value of deep specialization.

As Daniel Kahneman and Gary Klein observed, expert intuition is only reliable when environments are predictable and experts have sufficient feedback to learn those patterns. With technology rapidly evolving, we’re failing the first condition.

The same BCG study that highlighted AI’s benefits also warned that consultants using it were less likely to get correct answers on tasks outside the model’s capabilities—reminding us that human judgment remains essential.