AI efficiency gains are real but nearly impossible to precisely quantify because saved time gets absorbed by higher-value work, so companies must extract ROI through top-down headcount constraints rather than bottom-up measurement.
After Uber's CTO publicly said they blew through their annual AI budget in 4 months, Macdonald explains they are seeing tangible efficiency gains (e.g., cutting a 15-hour capital allocation process to 2 hours), but the ROI can't be formulaically measured — so Uber's approach is to hold headcount constraints tighter and let trusted teams allocate between compute and people budgets. ✦ AI generated
Andrew Macdonald · 20VC · 2026-08-17 · original ↗
starts at this moment · 36:47
“is that evidence of incredibly effective tools, or is that evidence of a desperate need for guardrails?”
the point around like ROI, for me it's a couple things. One is at the end of the day, we do want to get efficiency, or we want to get new and cool stuff built, and we are seeing examples of that every single day. We have stood up a pod of 30 of our best AI engineers that are partnered with business people, or partnered with folks in the G&A functions, to go in and go process by process, and start sort of ground up with AI. How do you improve that process? And if you can take like a capital allocation process, like every week we're allocating pricing dollars across thousands of markets globally, and I can take that from being a 15-hour process to a 2-hour process, which is what we've done. That is tremendous tangible ROI, cuz now you get 2 days of someone's time back... the way companies ultimately have to extract AI efficiency, at least from like a pure OPEX perspective, is just in your target setting, hold the constraints tighter. Like, if we really believe that AI is making our employees 10% or 20% or 30% more efficient, then next year we should just not increase head count. Or we should increase it by 2% instead of 10%.
verbatim transcript · starts at 36:47
36:47want to get new and cool stuff built, and we are seeing examples of that every single day. We have stood up a pod of 30 of our best AI engineers that are partnered with business people, or partnered with folks in the G&A functions, to go in and go process by process, and start sort of ground up with AI. How do you improve that process? And if you can take like a
37:11capital allocation process, like every week we're allocating pricing dollars across thousands of markets globally, and I can take that from being a 15-hour process to a 2-hour process, which is what we've done. That is tremendous tangible ROI, cuz now you get 2 days of someone's time back. Um if you're able to take a forecasting process, which our finance team is constantly reforecasting every inch of our business,
37:40uh and you're able to turn that from uh 8 hours of work into 2 hours of work, you're able to now do that not only with more precision, cuz you can put an additional layer of nuance into those forecasts, but you're just able to have your folks do other stuff. There's clear ROI there. If you're able to take marketing QA from 2 weeks to 2 days, like there's so many examples of that um
38:03that we see. And and the way we've done that again is by pairing the business folks with the AI engineers. The the second thing I think that >> Are you actually seeing that today? Because Alex Karp came on CNBC or CNN and and said like, "No, the the ROI question is still there." Like, to to validate what you said, to be clear. Outside of coding and customer support
38:23with the greatest of respect, I think anyone who runs a budget in a large enterprise state would say, "Yes, it's still not material at best." >> I think it's just hard to know. Like, these things are just hard to quantify and so you do have to be a bit top-down um and belief-based about it, right? I I think three examples I just gave there. Assume there are dozens of more of
38:44those. The natural question is, "Okay, great. Like, how many of those people can I take out of my organization so that I get the cost back and that flows through to the bottom line or I can put it into other things?" But formulaically doing that is really hard cuz guess what? The The 8 hours of value that was created or the 8 hours of excess time gets filled with some other
39:04activity, which is also like presumably high value. And maybe before wouldn't have got done to or wouldn't have been done to a level of precision. So, it's just very hard. So, I think the way companies ultimately have to extract AI efficiency, at least from like a pure OPEX perspective, is just in your target setting, hold the constraints tighter. Like, if we really believe that AI is making our employees 10% or 20% or 30%
39:28more efficient, then next year we should just not increase head count. Or we should increase it by 2% instead of 10% or we should decrease it by 5% and say, "You all should be getting more done with less." And here are all these sub examples of of people doing that. But drawing the direct line between I transform this process and therefore like I need two less operations analysts.