On the Allure of Salvation

Easter is upon us, and humanity has always enjoyed a good salvation story—so here is one: Once upon a time, there was a prosperous little financial services company. Suddenly it stumbled on an unforeseeable patch of rough ground: cost increased, revenue declined, profitability dropped, and it seemed as if the end was near. But shortly before it hit rock bottom, salvation for the little company arrived in the form of AI and automation: No longer do expensive human employees toil with tedious spreadsheets, PowerPoint presentations, and customer service calls—AI agents are here to deal with all the boring, monotonous work! As a result, our little company’s numbers are back up, investors are gleeful, and everyone is going to have a field day at the little company’s big IPO. Needless to say, they all lived happily ever after…

Sounds good, right? The company we’re talking about, of course, is Klarna: Riding a wave of pandemic-driven e-commerce growth, the Swedish buy-now-pay-later financier saw its valuation soar to $46 billion in 2021 and its global headcount swell to more than 5,000. A year later, however, as consumers began to return to physical stores, Klarna’s sales and profits had shrunk and its valuation plummeted to less than $7bn. But since then, Klarna has made a seemingly miraculous recovery: While steadily growing its GMV and revenue, it drastically cut costs and posted an operating profit of $181m in 2024 (up from a loss of $49m in 2023).

This resurrection of a business that seemed pretty much dead in the water is a lighthouse success story used by OpenAI to market its enterprise products: “Klarna’s AI assistant, powered by OpenAI, […] is doing the equivalent work of 700 full-time agents,” or so they claim. The message we should take away from this is clear: the AI-induced surge in productivity has enabled Klarna to lay off a lot of flesh-and-blood employees and save a ton of money. And there’s more than a grain of truth in this: According to Klarna’s annual report, about 22% of the workforce, or about 1,200 people, were laid off in the course of just one year. You can, of course, read this in the gloomy light of economic reality: as a thorough debunking of the happy myth that AI will “not steal jobs but allow workers to focus on more meaningful tasks”; as a harbinger of a wave of what Karl Max would have called “factor substitution”, a shift in the dominant means of production from labor to capital. But that’s not where we’re going (today). Where we are going today is a much more interesting place: OpenAI itself, whether replacing human workers with AI is a sustainable business model for them, and what would happen if it wasn’t.

But before we dive down the rabbit hole that OpenAI’s has become, there are two quick points I need to make about Klarna: First, the rise and fall of many e-commerce companies before, during and after the pandemic (think Shopify and its infamous 2022 layoffs) should make it clear that Klarna was thoroughly overstaffed by 2022. Apparently, it required more than 5,400 people at the time just to run an online payday-lending scheme. Many of them were likely hired on the basis of overly optimistic growth expectations that didn’t materialize. The fact that Klarna started reducing its workforce (from 5,700 in 2021 to 5,400 in 2022) well before the advent of AI and automation should be a strong indicator of this. So OpenAI won’t get all the credit for Klarna’s cost savings, but the implementation of AI certainly played a role during the later stage of the company’s downsizing. Second, it’s interesting to observe that women were disproportionately affected by these layoffs: Nearly 60% of those affected were women and only 40% were men, while the ratio is roughly reversed across the entire workforce. This could be an indication of the gender imbalance in the different roles at Klarna (customer success, sales, and marketing vs. management and engineering) as well as a sign of the increased vulnerability of roles traditionally held by women to AI replacement—another intersting aspect that we won’t be discussing further today.

Now let’s finally turn our attention to OpenAI: Let’s assume that of the 1,200 people Klarna fired, 700 were replaced by AI (as OpenAI claims). Using a very rough approximation of headcount costs, we can assume that this move would have saved Klarna around $70 million per year (OpenAI estimates that it “drove a $40m profit improvment” for Klarna, so we are at least in the right ballpark.) Of course, OpenAI does not give away its services for free. And while we don’t know exactly what they’re charging Klarna, we do know that it has to be a lot less than $70 million to make the math work for them. The conventional wisdom is that if you can’t demonstrate a 10x increase in efficiency, potential customers won’t be persuaded to replace whatever they’re doing with your new solution, so OpenAI probably couldn’t charge Klarna more than $7 million per year. If we stick to OpenAI’s own estimtae, this figure would be closer to $4 million. To add another angle: “ChatGPT Enterprise” is currently priced at about $60 per user per month. Assuming that half of the remaining 4,200 people at Klarna would need such a license, this would put the value of the contract to $1.5 million—the truth is probably somewhere in the middle, probably at the lower end. The bottom line is: If “replacing” 700 workers at Klarna brings OpenAI $2 million in annual revenue, that works out to about $2,800 per person replaced. But what does it cost OpenAI to bring in that revenue?

OpenAI’s cost structure, on the other hand, is quite convoluted, given what it has to pay for data collection and cleaning, model training, and every single query a user makes—plus the inevitable overhead of research and development, administration, sales, marketing, and so on. In total, the cost of running OpenAI in 2024 will be about $8.5 billion¹, according to reporting by The Information and Ed Zitron. So, how many workers in total would OpenAI thus have to “replace” in order to pay its ongoing annual costs out of pocket? Answer: 3 million. Give or take.

For the sake of argument, let’s assume four things in OpenAI’s favor:

1. OpenAI’s annual cost is going to stay constant for the years to come, regardless of the load it puts on its systems, the complexity of its new models, or any additional use cases it unlocks (which it won’t)
2. It’s technically feasible to replace many different kinds of “customer service” jobs with a single, pre-trained AI model like ChatGPT (which it isn’t)
3. All the workers who are currently doing these jobs are paid first-world wages that make them susceptible to replacement in the first place (which they aren’t)
4. There are no competitors out there who are vying for the same market (which there are)

The billion-dollar question is: Under these, highly idealized, conditions: Is it conceivable that OpenAI is going to pull this off before they run out of cash?²

This feat sounds difficult, to say the least, but let’s not dismiss it out of hand. Salvation for OpenAI seems to have just arrived in the form of a middle-aged Japanese gentleman: SoftBank, ran by Masayoshi Son, announced it would invest an astonishing $40 billion. The bad news is that OpenAI also has to face skyrocketing costs—the net result being that all that fresh money (which is unlikely to materialize in full, given how convoluted the deal is structured) would only keep it afloat until maybe 2027. Then, if it manages to raise the same amount again (which is even less likely), it would survive for another two or three years. So, sticking with our optimistic scenario, let’s say OpenAI’s runway lasts until 2030 before someone wants their money back—either Microsoft, or SoftBank, or the Saudis, or whoever else has sunk their family fortune into this venture by then. What would the world of “customer service” have to look like for this to work?

Globally, we’re talking about 20 million people working in “customer service” roles that don’t involve physical interactions, and so could theoretically be “replaced” by AI. Is it realistic to expect more than 15% of these to be replaced by OpenAI alone over the course of just five years? That would be 600,000 jobs, worldwide, every year from now on, including in 2025. The Klarna case, which affected 700 people, would have to be repeated about 850 times. Is this a realistic scenario, on top of all the other optimistic assumptions we made above?

Well, color me skeptical.

But more importantly, here’s billion-dollar question number two: What if this scenario (or any other that leads to OpenAI becoming profitable in the foreseeable future) doesn’t work out? What if OpenAI wakes up one day to find that the gap between revenues and costs is exploding, that it can’t keep funding itself out of trouble, and that the debt collector is knocking on the door? Will a company that at its peak (right now) is valued at more than $300 billion go quietly into the night? I don’t think so. And I think that could happen then is what we should be concerned about.

Suppose you run an apple stand. You buy a kilo of apples for $3 and sell them for $2 a kilo. You borrowed $100 from wealthy friends to get your business off the ground. What can you do to become profitable before you run out of money?

One way is to keep borrowing, hoping to undercut your competitors, and then, once you’re the only apple vendor on the market suqare, raise your prices dramatically. Of course, this strategy works even better if you make your apples irresistible to your customers by, say, injecting them with something like… I don’t know, heroin, maybe? Then, of course, you could easily charge something like $20 a kilo, or $200 a kilo, whatever, and still make a fortune (after paying off your creditors, of course). If that doesn’t appeal to you, you can look for other ways to supplement your income: Why not put small ads on your apples and let the advertisers pay you? What if you give away your apples for free, but outfit them with tiny sensors that collect granular data on your customers’ apple eating habits? Who eats what kind of apple, at what time of day, for what purpose, and with whom? You could fit your apples with tiny microphones or, better yet, video cameras, and listen in to the conversations people have around your apples. Can you thus find ways to deliver specific apples with highly personalized ads to specific people? Can you sell the troves of data your apples are harvesting to other businesses? Once you start thinking in that direction, possibilities are endless!

Shoshana Zuboff calls this “surveillance capitalism, and unfortunately it has become the dominant monetization model for Big Tech: build an addictive product, give it away at a loss, then turn around and sell user data and targeted advertising. This is the game that Google, Meta, Tiktok, and many others are playing, to the detriment of people’s mental health, the structure of democracy, the sustainability of high-quality journalism, and many other things that society at large should care about. If we allow OpenAI and others in the field to adopt a similar model, however, it could be much worse than what we’ve seen so far: The behavioral data that search engines and social media platforms can collect is coarse-grained and imprecise—what ChatGPT knows about your thoughts, dreams, aspirations, and fears has much more texture and context. Ads on web pages and in search results can be easily detected and automatically blocked (or simply ignored)—when they’re woven into a lengthy conversation with a chatbot in the form of subtle nudges, there’s no way to counteract them. The persuasiveness of generic text, video, and banner ads is pretty low—but when your hyper-realistic AI friend asks you to vote for this guy or that party (or not to vote at all), it will be impossible for many to say no.

The picture I’ve painted here is admittedly pretty bleak. But it’s only one of many ways in which the future of OpenAI, its cohabitants in the AI space, and society at large can play out. Alternatives are possible: OpenAI can become profitable more quickly than we’d expect. Innovation can dramatically reduce the cost of training and operating LLMs, reducing the incentives for AI companies incentives to monetize data and advertising. Consumers can choose to use these tools primarily in ways that don’t generate tons of exploitable data. Regulators can step in to prevent the worst outcomes of AI-driven surveillance capitalism. Investors can persuade the companies in question to stay on the high ground. Many more scenarios are possible, but we should keep our eyes open for the worst—while actively working toward the best. (Spoiler alert: I have much more to say about using AI and automation for purposes that actively contribute to the betterment of society. Stay tuned!)