A few days ago, Apollo released an article arguing that AI has not had payoff outside of the tech sector, raising concerns around a bust should AI stay confined to a small part of the economy. The article supposes that the cause of this is AI’s lack of awareness around reputational risk, stating that business intelligence lies in human-exclusive “judgment.”
“Judgment” continues to do a lot of work in preserving the need for human oversight over, or in lieu of, AI. Software engineers have used it to make a case for themselves. Perhaps you’ve heard this analogy circulate: Smartphones didn’t replace photography, so agents won’t replace software engineering.
This is a false analogy. Smartphones provided a better-than-disposable-less-than-professional tool to take photos, but it didn’t provide the skill to use that tool well. Smartphones don’t find the right backdrop, angles, and lighting, and then capture and edit the photo automatically. Smartphones, no doubt, raised the bar for what professional photographers needed to provide to retain premium value; however, one could still clearly identify that higher bar.
When it comes to software engineering, agents provide both the tool and the skill to write quality code. This is why the comparison is disanalogous. It is well documented that agents allow moderately tech-savvy people to build software applications. Until agents can easily perform long-horizon tasks, there undoubtedly will need to be engineers to over see agent-driven development—but there will be far fewer people needed (and advancements are being made daly on the long-horizon problem).
Apollo’s article strikes me as a similar claim that covers the non-tech corporate sector. It claims, “AI may be a great calculator, but in business, intelligence is mostly judgment. Judgment depends on context, and the answer varies with which human or which model you ask.”
However, tech companies possess the same business functions as any other corporate business. They need sales, marketing, HR, legal, etc. teams. Like a venn diagram, there are some functions that do not overlap, but many of them do. Anthropic et al have demonstrated that accountability, trust, culture, unwritten rules, and ethical trade-offs—what Apollo cites as irreplaceable business intelligence—can be embedded in and carried out by AI workflows. Claude’s very constitution is an obvious example. Moreover, Anthropic, in recent days, has highlighted how it has used AI to sharply increase productivity in non-technical (or, perhaps, less-technical) functions/roles.
Apollo’s article, therefore, doesn’t seem to accurately reflect the potential of AI in pockets of corporate America requiring business intelligence; it reflects the lack of will to integrate AI’s potential.
To state that there are “no signs of AI payoff” is misleading; there is a privation of indicators that signal “we’re willing to replace our business intelligence (and stomach the side-effects of worker displacement).” Anthropic gestured at a similar concern within their company, and so, they resorted to financial incentives to drive AI integration. The obvious difficulty is that most companies in corporate America do not have the influx of investment to provide these financial incentives that surmount the immediate hurdle of fears of worker displacement.
AI can handle business intelligence (i.e., “judgment”). That judgment in the hands (or rather, the intellect) of business leaders, however, can thwart AI’s absorption judgment. The fate of the AI economy is far from settled, despite Apollo’s claims. It depends on a race to AI integration before social costs overthrows the project—and Apollo’s sentiments reflect one of those rising social costs.
Any hopes of AI integration, and its payoff, will require addressing social concerns, and that problem is not primarily technological.