No KPIs: Treat AI Like R&D

_{IN ONE SENTENCE}

Imposing KPIs on your AI projects today means killing innovation before it has a chance to emerge.

_{THE OBSERVATION}

This is one of Ethan Mollick's strongest convictions: in the current phase of AI adoption, KPIs are dangerous. Every company wants to measure the ROI of its AI investments. That's understandable. But Mollick observes that this obsession with measurement systematically pushes organizations toward the same mediocre results: 30% efficiency gain in customer service, followed by 30% headcount reduction. It's a vicious cycle that kills innovation.

The fundamental problem: we don't yet know what AI can really do for organizations. Optimizing for predefined metrics is like searching for your keys under the streetlight: you measure what you know how to measure, not what creates the most value.

_{WHAT YOU NEED TO UNDERSTAND}

You're optimizing for the wrong thing

If you put a KPI on document productivity, you'll get more documents. But do you want 300 PowerPoints a week? If you measure lines of code, you'll get more code. But is it better code? Mollick emphasizes that traditional metrics: output volume, processing time, cost per unit: don't capture the real value of AI. Worse: they steer teams toward efficiency rather than transformation.

The cost-cutting trap

AI KPIs almost systematically converge toward cost-reduction objectives. And cost reduction means layoffs. Which destroys exactly the ecosystem you need for AI to work: employees who experiment, share their discoveries, and collaborate with AI tools without fear.

Adopting an R&D mindset

Mollick recommends treating this phase as R&D: an exploration budget, not an optimization budget. Productivity gains will come naturally and quickly, especially in software development where results are already clear. But for transformational use cases, experimentation needs time and space without the pressure of quarterly metrics.

_{WHAT THIS CHANGES FOR YOU}

• Resist the temptation to put productivity KPIs on your AI projects in the first 6-12 months
• Allocate an AI budget labeled as R&D with tolerance for failure; not an IT budget with expected ROI
• Beware of KPIs that converge toward headcount reduction: that's the signal you're optimizing for the wrong thing
• Quick wins (coding, summaries) can be measured, but transformational use cases require patience and exploration