This week, the talk of the town has been Ford Motor Company’s decision to begin hiring senior, or “graybeard,” engineers after having laid off many in favor of AI automation.
The details are indeed delicious, with the company’s vice president of vehicle hardware engineering stating openly that “Mistakenly, we thought that by just introducing artificial intelligence and ingesting the design requirements that we had, that that would produce a high-quality product.”
The issue seems to have been that human engineers were simply better at spotting flaws before they hit the plant floor, resulting in significantly lower costs for warranty fulfillment and even recalls. Ford CEO Jim Farley said that bringing back competent human engineers led to “literally hundreds and hundreds of millions of dollars” in savings.
That would imply that AI had previously been costing “hundreds and hundreds of millions of dollars” more; that’s perhaps not surprising, given that Ford has already been associated with billion-scale losses due to high warranty costs. And it’s worth remembering that Ford needs to worry not just about recall costs but also about loss of human life and the lawsuits that follow.
Credit: Rykoff Collection/Getty Images
Still, Ford’s decision seems to be part of a larger trend. In recent months, both IBM and Commonwealth Bank of Australia have directly reversed AI-driven cuts. IBM, specifically, cited not just problems with the quality of AI outputs, but the need to invest in the next generation of talent; the company eventually announced it would triple its entry-level hiring.
The issue isn’t just work quality; costs to employers are skyrocketing as AI service companies switch to new billing methods. Not only is usage-based billing causing some corporate AI bills to spike by multiples, but more advanced LLMs are also consuming more tokens per request.
Not-so-coincidentally, the same companies that recently began encouraging employees to use AI as aggressively as possible are now enforcing usage maximums to avoid going over budget on tokens. The “unlimited” model was always going to be temporary, it seems.
Beyond that, companies are only now beginning to face up to the fact that AI costs more than human workers across a wide range of areas. It’s not just true of lower-paid workers, either; higher-paid positions tend to come with more complex tasks, which means more AI tokens consumed and a greater chance of a subpar contribution from the AI.
Even Nvidia’s VP of applied deep learning admits that, at Nvidia itself, the “cost of compute is far beyond the costs of the employees.”
There are attempts to keep costs down, such as the OpenRouter service, which routes requests to the lowest-cost model it thinks can handle the task, but these can’t solve the fundamental problem. By some estimates, even usage-based billing hasn’t made the biggest LLMs profitable per token of work delivered.
Credit: Xia Yuan/Getty Images
This is not a problem that OpenAI or Anthropic will be able to solve anytime soon. Costs to companies need to come down significantly, just as customers become savvier about LLM output quality and begin demanding far more token-intensive models.
It’s worth asking: Just how primitive would an LLM become if it had to cost less than it could get away with charging? It’s a tortured question, since the worse the model, the less it could realistically charge, meaning there’s a chance there simply isn’t a directly profitable option at any level of quality.
As we see in Ford’s comments, even the most advanced AI models lag behind skilled human labor in the areas that matter most. Yet, as seen in the failure of AI voice-bots to eliminate human call centers, cost considerations mean AI struggles to succeed even where it does deliver decent work.
With every passing week, it seems more as though the AI job-pocalypse will fizzle out before it gets started; the only problem is that, if the failure of AI automation causes the overall AI bubble to burst, just how many jobs will disappear for humans and AI, alike?
