AI ROI debates are often hiding a capability measurement problem

A lot of AI investment reviews are being run with the wrong test. If a CHRO, CPO or CLO is being asked to justify AI spend with a standard ROI model, I would slow the conversation down and separate two decisions: are we funding a tactical efficiency move, or are we building a capability that changes how work gets done over time? My read is narrow but useful. The current push to connect AI usage to business value may help leaders get more disciplined, but only if they stop treating all AI spend as one financial category.

OpenAI has set out a simple definition of incremental business value: the net financial gain created by an AI system after subtracting the full cost of operating it. That is a sensible finance anchor. Alongside that, a second source in the record argues that AI investments fall into five types, with two aimed at maintaining position and three aimed at building durable advantage, and that traditional ROI tools do not fit all of them well. I would not take that as proof that old metrics are obsolete across the board. I would take it as a prompt to tighten the operating question before approving or rejecting spend.

For workforce and learning leaders, the mistake to avoid is measuring an AI-enabled capability investment as if it were a short-cycle automation purchase. If the investment is really about reducing time to proficiency, improving manager decision quality, increasing the consistency of frontline execution or shifting work from knowledge retrieval to judgment and action, the value chain is longer. It runs through application in the work, repeated practice, more reliable behavior and then business performance. If an organization adopts AI in those settings, usage data alone will not tell you whether the investment is working. You need evidence at the task and workflow level.

That changes ownership. The CFO still needs a financial view, but the CHRO, CPO, CLO or transformation lead should insist on an investment map that links each AI use case to one of three things: direct productivity gain, risk reduction or capability build. Then set the smallest evidence threshold that could change the decision. For a tactical use case, that may be cycle time, error reduction and operating cost. For a capability build, I would want to know whether people can perform higher value tasks sooner, whether manager load rises or falls, and whether the surrounding rules, approvals and role design still fit the new level of human capability.

Speed without diagnosis scales waste.

The limitation is straightforward. This is still early, and the record here does not establish that new measurement frameworks are being adopted widely or that they improve investment quality in practice. There is also a real warning in the counterevidence: many AI investments may fail to produce material value. That is exactly why leaders should be stricter about classification and evidence, not looser. If I were making the call now, I would not approve broad AI value claims from learning, HR or transformation teams unless they can show which investment type they are funding, what operating condition is expected to change, and what evidence would stop, improve or scale the effort.