AI & Productivity
Back in 2010, P&G owned a prestige fragrance business. They had their own floor in the Geneva HQ building, which they would redecorate a few times a year to reflect their latest marketing campaigns, at a cost of tens if not hundreds of thousands of dollars.
At around the same time, I worked for P&G’s factories in China, where engineers and I evaluated projects to reduce manufacturing costs. We squeezed every dime of excess we could find: one memorable proposal from a site manager was to limit the supply of toilet paper — a suggestion that the HR lead angrily and correctly shut down (but which the Chinese government adopted).
It seems absurd that one business unit would spend serious money on shiny posters and display cabinets, while another part of the business would closely monitor its air conditioning bills, but it shows that in most domains, bad business decisions are rarely the outcome of imperfect information or inadequate technology, but the result of poor incentives, org structures, and managerial systems.
All this is to say that at the micro level, there is a lot of misplaced optimism in AI’s potential to directly improve productivity and efficiency. But there is a great deal of nuance in this statement.
On paper, AI will increase productivity, in the sense that companies will be able to reduce headcount. Already it is trivial for companies to connect AI agents to their data sources, ask questions in plain English, and get thorough analysis that is surprisingly accurate. In the past, companies employed armies of analysts to answer questions such as: why did our business unit’s average revenue per unit go up? Should we run a 50% price reduction promotion, or a buy-one-get-one-free promotion? In which city do we have the highest market share? Should we offer annual subscriptions? Etc. Now you can have an AI agent answer these in minutes.
However! First of all, companies never needed armies of analysts to answer these questions in the first place. In my experience, the most productive analysts were always able to produce 10x the output of the average analysts. But in the past, many managers were rewarded by growing the size of their teams, and so they inflated headcount. Now we’re moving to a market that incentivises AI adoption, so managers are encouraged to show token usage and reduce headcount. Many will find that they don’t even need to use AI all that much: they can cut the size of their teams and attribute the efficiency to AI. I suspect many will do that.
So AI itself is unlikely to be the real reason for efficiency gains in terms of headcount reduction. Second, it’s even more unlikely to lead to better decision-making in most companies. There are some companies that are hyper-optimised, and there AI may be able to add value by helping managers identify ever-diminishing marginal impacts. But in most companies I’ve worked for, there’ve been big fat juicy low-hanging fruit just hanging there waiting to be plucked. The reasons they hadn’t yet been are almost always political. Let me give another example: one of my biggest successes at Google was working with GTM to establish sales teams to sell the company’s Universal App Campaigns. These were ads for clients’ apps (as opposed to their websites). It didn’t take a genius to make the case that consumers use apps at ever-increasing rates; the difficulty wasn’t in pulling the numbers together, it was in doing roadshows and convincing market VPs to re-allocate resources, which would (as always) ruffle feathers. Our job in pulling numbers together wasn’t to inform; it was to protect — to give people making decisions the arguments to push back against naysayers. You don’t need AI to uncover such insights. AI won’t make your company produce better goods and services. Cutting red tape and politics might; which, yes, easy to say hard to do — but if you can’t cut politics anyway, why do you think AI will help?
I said earlier that there is misplaced optimism at the micro level, by which I mean within individual companies. Let’s now turn to the macro-level, meaning the economy as a whole. Here there is a real possibility that AI will be beneficial, if it leads to breakthrough innovation. Pioneering companies might use AI to discover new drugs, build functional quantum computers, invent new materials, produce cheaper energy, and reduce wasteful friction and coordination costs. All this may well happen.
But until then, I worry that many people will mistake productivity growth for increase in wealth. Historically, the two go hand-in-hand: the more productive we are, the more things we can produce. But the two do not necessarily go together. If the increase in productivity comes from eliminating latent bloat, and that bloat cannot be put to other productive use, then we’ll just end up with more unemployed people and the same amount of wealth. There won’t be more things to go around; only fewer people producing them, and able to afford them.
In other words, like many others, I worry about inequality. Unless we bank on AI’s transformative impacts outlined above, I disagree with the optimists who think that the jobs lost because of AI will be offset by other, new, AI-enabled roles. Sure, this will happen to an extent: for every SAAS company that dies because its clients can now vibe-code internal tools, there will be new engineer roles created in those clients. But since most of the jobs displacement will come from the elimination of bloat, and because that bloat was engaged in white-collar jobs supporting the provision of services, and because I doubt we need more services, there won’t be much for most of the people who lose their jobs to do. Most of us don’t need more apps (since the advent of AI we’ve seen more apps than ever being published, but fewer being used); we need better houses, furniture, food, clothes. The people who’ll be fired because AI took their jobs don’t have the skills to produce these goods.
But I also have nits to pick with my fellow inequality worriers. For one thing, unlike many of them, I don’t think inequality is morally reprehensible. If some people are 10x as productive as others, it seems fair to me that they earn 10x as much. That said, pragmatically, something has to be done to anticipate the effects of inequality and unemployment, otherwise we’ll see more polarisation, division, and hatred. The obvious answer proposed by many is tax. Dario Amodei is calling for universal basic income to account for the impact AI will have on the job market, but it will be practically impossible to introduce universal income equal to the salaries of the people who’ll be laid off. A data scientist on $200k won’t feel grateful for $30k in free government money.
What we need instead is government effort to produce more of the things we want. First, make it easier for people to make things: de-regulate, open up job markets by coming down hard on licensing, re-open trade, stop NIMBYs, cut down red tape. Second, instead of universal income, offer training: help data scientists learn how to install pipes, make furniture, ferment yoghurt, cut cloth, plant tomatoes, whatever.
Companies are using AI as an excuse for realising efficiency gains they could have delivered decades ago. Maybe governments can do the same.