Uber's chief operating officer, Andrew Macdonald, said something unusual for the tech world: that it's becoming increasingly difficult to justify the company's AI spending. In an interview with Rapid Response, he admitted that greater activity with these tools doesn't always translate into proportional gains. He said it plainly, without exaggeration.
Almost simultaneously, Duolingo chose to stop measuring employee performance based on their AI usage. An organization that had integrated these tools into much of its operations decided that tracking the use of a technology isn't the same as evaluating the actual quality of the work.
Two signals. Two different organizations. A questioning that's starting to become visible.
Uber runs one of the largest mobility platforms in the world, with the resources to attract top talent and access to massive amounts of operational data. If any company could demonstrate a clear return on AI, it was this one. And yet its own chief operating officer publicly acknowledges that the numbers don't quite add up.
This matters because it reveals how many of these investment decisions came about in the first place. They didn't always start from a defined problem, with a baseline measurement and a follow-up evaluation. In several cases, it was more of a race not to get left behind. Boards were asking about AI strategy. Investors were scanning reports for the term. The pressure came from fear, not from data.
Macdonald isn't claiming AI lacks value. He's pointing to something more uncomfortable: that it's not always clear which part of it works, to what degree, or whether the cost of finding out is worth it.
Duolingo represents the flip side of the same issue. Its decision to drop the AI usage metric suggests they noticed they were measuring the tool instead of the outcome. It's like evaluating a craftsman by how many times he swings the hammer rather than by how durable what he builds turns out to be.
Both decisions seem correct. They represent institutional maturity. The problem isn't that AI investments are being reviewed. What's telling is what this reveals about how they were made in the first place.
Technology tends to arrive before the metrics needed to evaluate it. It gets adopted first and justified later. When results don't show up on their own, the common response is to invest even more, because so much is already committed.
Model and compute providers have a clear incentive to keep spending flowing regardless of outcomes. Their business model runs on volume of use, on calls to their systems, on hours of processing. More activity means more revenue for them, even if that activity doesn't always create real value for whoever is footing the bill.
When Macdonald says that greater activity doesn't translate into proportional gains, he's describing exactly that mismatch. The metric that matters to the provider doesn't line up with the one that matters to the buyer. For a while, this worked out conveniently for everyone: sales went up, active adoption got reported, and investors saw what they wanted to see.
Now that enthusiasm is running into financial reporting cycles. CFOs are asking questions that technical teams can't always answer with precision.
Something similar happened with the internet in its early days. Many organizations poured significant sums into digital presence and e-commerce before real demand actually existed. Some of those investments never paid off. Even so, the internet transformed everything — just not on the timeline or in the ways that were promised back then. The record of previous technological shifts shows that these cycles of enthusiastic adoption followed by reassessment aren't new.
I'm still not sure how far this correction will go. It's more complicated than it looks at first glance.
Does real value only emerge once we stop measuring usage and start looking at its effects on the things that actually matter?
