In my essay The AI Dividend Fallacy, I argue that AI productivity gains associated with generative AI (GenAI) do not automatically translate into reduced headcount and lower costs. That naturally leads to another question I hear from executives:
How should we measure the ROI of enterprise GenAI investments?
My answer may seem surprising.
For general-purpose GenAI assistants such as ChatGPT, Claude, or Gemini being used for broad knowledge work, organizations often spend unnecessary effort trying to measure something that isn’t even worth capturing precisely (if that’s even possible).
The economics are already favorable
In one organization I know well, the enterprise GenAI tools made available for broad knowledge work would have paid for themselves by increasing employee productivity by less than 0.5%.
Think about that for a moment.
How would you reliably measure a 0.5% productivity improvement across thousands of employees performing dozens of different knowledge tasks?
Building instrumentation capable of detecting a change that small could easily cost more than the uncertainty you are trying to eliminate.
Knowledge work is difficult to measure
This is really a reflection of a common challenge in most organizations – measuring the productivity of day-to-day knowledge work, and the impact of tools on that productivity.
Consider a simple thought experiment. What productivity gain does your organization achieve by using Microsoft Office? Do you calculate ROI each quarter? Where would you even start?
Most organizations now simply accept that Microsoft Office is enabling infrastructure for knowledge work rather than an investment with an objectively measurable ROI. And I would argue that GenAI is rapidly becoming the same type of capability.
Good governance still matters
I’m not suggesting that GenAI for knowledge work should become a free-for-all. Organizations still need employee training, spending controls (including token caps), data protection/management, robust governance, and risk management to use GenAI safely, effectively, and responsibly. And the additional costs associated with GenAI should be included when considering the broader economics. But again, even if these things increase costs by a factor of four, how effectively could most organizations measure a 2% gain in efficiency across broad knowledge tasks?
Measure where it matters
Measuring ROI rigorously becomes far more meaningful when AI is embedded within repeatable, high-volume business processes and workflows. Such processes often have existing checkpoints or review gates that can serve as reference points, making them easier to instrument. They’re often measured using existing metrics, making them easier to benchmark and easier to compare to historical performance.
And most importantly, they’re often key to the way the organization generates value. That means that productivity gains can drive substantially larger and more measurable business impacts than those associated with broad knowledge work.
So my advice for measuring GenAI ROI? Focus measurement efforts where they can produce meaningful insight rather than chasing statistical noise. Measure and make important investment decisions around value creation. And when it comes to business enablement, focus on managing costs and delivering capabilities as efficiently as possible.