AI vs Productivity
AI increases productivity. That is one of the main selling points that AI vendors are hawking these days. AI can certainly do many things much faster than any human could. But does that necessarily mean that it improves productivity? Industrial productivity is straightforward; inputs plus process equals outputs. If you move two machines closer to each other on a factory floor, you can measure the time that change saves you for the same output. Measuring productivity in knowledge work is quite a bit murkier.
The challenge of tracking how production inputs lead to outputs in knowledge work existed as a problem before the current age of AI. Our engineering (or accounting) minds longed to apply models from the industrial age onto the information age. You cannot measure the inputs of knowledge work like you can assembly-line labor. Instead of shifting our focus to the productive output, which is after all what is most important, most measurements instead relied on proxy input measures. We can measure how much time people spend in meetings or how many emails they read. If we assume that this measurement is representative of the inputs that go into knowledge work output, we have solved the logical problem and can apply industrial era formulas.
When we see claims of AI boosting worker productivity, it is often these proxy measures that are the underlying justification. If I went from 20 hours a week of meetings pre-AI and now I can consume 40 hours of meetings post-AI, then I have doubled my productivity. I often see this type of thing in testimonials claiming to have automated 75 hours a week. Were those people really working an additional 75 hours every week before AI?
The real problem lies in the fact that these activity measures as proxy metrics do not translate into output that matters. In fact, I would argue that they are often inversely proportional in knowledge work. I doubt many innovators would credit additional email and meetings with helping them reach a breakthrough. With or without AI, we need to measure knowledge work productivity based on results.
I think this was summed up well by Uber’s CEO: “Maybe implicitly there’s more that is getting shipped, but it’s very hard to draw a line between one of those stats and ‘Okay now we’re actually producing like 25% more useful consumer features.” He has a clear goal for productivity measurement: useful customer features. AI and other tools can drive real productivity, but only if we measure what actually matters. When we push workers to use AI in ways that are easy to measure, we may actually be guiding them in the wrong direction. More activity doesn’t guarantee more productivity, but people will respond to the incentives of the system they operate in. Good ideas often arrive on a walk in the woods, not while grinding away through back-to-back meetings. That walk is difficult to measure and speed up with AI.
