Matt Busigin

I expect AI to make execution dramatically cheaper across many industries, but not necessarily in a way that appears as a standalone “AI revolution” in economic accounts. A pharmaceutical company might produce cheaper medicines, or an energy company might extract oil more efficiently; the gain shows up in those industries’ output. That remains different from comprehensive physical automation, where AI can reliably act throughout the real world. Cheaper execution also does not eliminate expertise. Agents can compress work such as infrastructure changes or server migrations, but someone still needs enough systems knowledge to define the task, inspect the result, and catch dangerous mistakes. The value shifts upward from manually performing every step toward specifying, supervising, and integrating the work. The key engineering issue is control. An action-biased agent that eagerly changes files, runs commands, or alters systems is useful precisely because it can act—but that same trait makes unsupervised deployment risky, especially when operations are irreversible. I think the practical path is bounded, executable workflows: structured plans, explicit permissions, review points, and smaller decision models embedded inside larger control systems. The future depends less on a model producing impressive text than on building reliable machinery around its decisions.

The biggest update would come from systems that can execute long, consequential workflows—including physical-world tasks—reliably without deep expert supervision. That would challenge my view that expertise mainly shifts upward into specification, review, and integration, and it would make a distinct, economy-wide automation transition more plausible. The opposite update would be persistent evidence that agent reliability does not improve once tasks involve irreversible actions, changing environments, or long chains of decisions. If progress remained concentrated in cheap generation while verification and supervision costs stayed high, I would expect AI’s impact to remain substantial but mostly absorbed into existing industries and bounded workflows rather than becoming comprehensive automation.