AI AUTOMATION

AI automation is the use of software — ranging from rule-based scripts to reasoning AI models — to perform tasks that previously required continuous human effort.

Levels of Automation

Not all AI automation is the same. Robotic process automation (RPA) executes rigid, rule-based workflows: copy this field, trigger that system. ML-assisted automation adds a classification or prediction step, allowing the system to handle variability in inputs. Agentic automation goes further still — a model reasons across steps, uses tools, and adapts to novel situations mid-task. These AI agents are built on advances in machine learning and large language models. The economic implications differ sharply at each level, and conflating them produces bad forecasts.

What Does Not Automate Easily

Tasks resisting automation share recognizable traits: they require physical manipulation of the world (plumbing, surgery, construction), they carry catastrophic failure costs that make AI error unacceptable (nuclear safety, aircraft control), or the human relationship is itself the product (therapy, leadership, high-stakes negotiation). The disruption is real but uneven. The timeline is slower than peak hype suggests and faster than most incumbents are preparing for.

The Task Decomposition Frame

The most analytically useful question is not which jobs will disappear but which tasks within jobs will change. A radiologist's role contains tasks that automate well — flagging anomalies in high-volume scans — and tasks that don't — communicating uncertain diagnoses, integrating clinical context, exercising judgment in edge cases. The net employment effect in any 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. Understanding this reshaping is central to thinking clearly about the future of work.

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 where outputs can be evaluated cheaply and objectively — code that either 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. Embedding AI automation into core operations is as much an organizational challenge as a technical one, which is why it is inseparable from broader digital transformation. Workers who adapt fastest develop judgment about when to trust AI output and when to override it — a meta-skill that is becoming as valuable as domain expertise. Organizations investing in this capability now are building a durable advantage over those waiting to see how the technology matures.

Open Questions

• As agentic systems improve, will verification costs fall proportionally or remain the binding constraint?
• Which regulatory frameworks will govern liability when an automated system causes harm?
• Does elastic demand absorb productivity gains into more output, or do firms extract them as margin?
• How do organizations measure the "automatable task ratio" within a role before making workforce decisions?

Part of the knowledge graph at The Best Blog Ever — reference definitions for ideas that matter.