AI AUTOMATION

AI automation is not a single phenomenon. It ranges from simple rule-based process automation augmented by ML classifiers (robotic process automation) to sophisticated reasoning agents capable of handling complex, multi-step knowledge work. The economic implications are different at each level.

The tasks most vulnerable to automation share identifiable characteristics: they involve pattern matching over large datasets, produce outputs that can be evaluated objectively, don't require physical manipulation of the world, and have high enough volume to justify the engineering investment. Software development, legal document review, customer support, and financial analysis all show significant early automation.

What doesn't automate easily: Tasks requiring real-world physical interaction (plumbing, surgery, construction), tasks where the cost of AI error is catastrophic (nuclear safety, aircraft control), and tasks where human relationship and judgment are themselves the product (therapy, leadership, negotiation). The economic disruption is real but uneven, and the timeline is slower than peak hype suggests and faster than most incumbents are preparing for.

The task decomposition frame: The most analytically useful approach is not asking which jobs will disappear but which tasks within jobs will change. A radiologist's job contains tasks that automate well (flagging anomalies in high-volume scans) and tasks that don't (communicating uncertain diagnoses to patients, integrating clinical context, making judgment calls in edge cases). The net effect on total employment in a given role depends on the ratio of automatable to non-automatable tasks and whether demand is elastic enough that productivity gains translate to more work rather than fewer workers.

The verification bottleneck: AI automation at scale runs into a consistent constraint: someone or something has to verify the output. In high-stakes domains, verification costs eat into the productivity gain. This is why the automation frontier advances fastest in domains where outputs can be evaluated cheaply and objectively — code that either compiles and passes tests or doesn't, documents that match templates, classifications that can be sampled and graded. The deeper constraint on AI automation is not generation capability but evaluation capability: the ability to reliably distinguish good AI output from bad at scale.

The workforce transition reality: Automation transitions historically play out over decades, not quarters, because they require capital investment, retraining, process redesign, and regulatory adaptation. The workers who adapt fastest are those who develop judgment about when to trust AI output and when to override it — a meta-skill that is becoming as valuable as any domain expertise. Organizations that invest in this capability now are building a durable advantage over those waiting to see how the technology matures.