Great Quarter

My Sunday night musing about AI: The majority of companies aren't seeing productivity gains from AI because the predominant use case of AI is to speed up existing human tasks, not to redesign how work gets done. Using AI isn't going to magically increase large-organization revenues by 10%. The gate on that isn't "my employees aren't doing enough." Revenue gains come from a combination of ideas, implementation and iteration. Much of that takes time, because the humans on the other side (the consumers) have to react and change their decisions. The revenue will come, but it lags. No, the vast majority of near-term productivity gains come from making your employees more efficient at their existing jobs. At least initially, those gains go to the employees, who spend less time working because AI is making their job easier. That's why it doesn't show up in the data. Productivity is output per paid hour. If an employee finishes in 6 hours what used to take 8 and takes the other 2 as slack, paid hours don't fall and output doesn't rise. Measured productivity stays flat. The gain is real, but the worker captures it, not the firm. It also doesn't add up neatly. Saving 60 minutes a day for six people doesn't free up a seventh person. Over time, companies will reduce headcount and get the same amount of work done by fewer people. That process takes time because AI is currently pretty good at tasks but not great at entire jobs or workflows. (That’s what the next $500b of training is supposed to solve!) Laying people off due to AI efficiency is also not a good signal right now. So unless you're Microsoft, the likely mechanism is to stop or slow backfilling attrition. Typical white-collar attrition runs 10-15% a year. Leave a quarter of those seats empty and headcount falls a few percent a year with no layoffs. This may only be the first wave. In his 1990 paper "The Dynamo and the Computer," economist Paul David showed that the first electrified factories simply swapped the steam engine for an electric one and kept everything else the same, and productivity barely moved. The big gains came only when factories were redesigned around electricity, with a motor on each machine. That took decades, because it meant rebuilding factories and power grids. AI won't take that long, since it runs on infrastructure that already exists. But the order will be the same: the efficiency gains come first, and the workflow redesign gains come later and are bigger. Concluding "AI doesn't increase productivity" from "the data doesn't show productivity gains" is being overly reliant on trailing, imperfect indicators instead of looking at what's fundamentally happening at the worker level. It's also early. Most companies have only been using AI in earnest for the last 6-12 months, and what they spend on AI models is still a rounding error next to what they spend on people. A tool that costs a fraction of a percent of your overhead, adopted a year ago, isn't going to show up in the national productivity numbers yet. If I'm right, here's what to watch over the next 2-3 years: 1) Revenue per employee rising at companies with flat or shrinking headcount. 2) Entry-level hiring slowing in the roles most exposed to AI. It’s crazy because the entire industry is moving so fast, but it’s worth remembering there’s still “real-world bottlenecks” (people, organizations, customers) that don’t behave as nicely as scaling laws.

$RWT @greatquarter