AI Transformation · Glossary

AI Automation

Updated October 25, 2026

AI automation means using an artificial intelligence (AI) model to do a step of work, or a whole defined workflow, that a person used to do. The working test is whether the output is usable without edits. Below that line the person still does the job, only faster, which is augmentation rather than automation.

Not every automation needs AI, and not every AI automation is an agent. In a conference talk, Barry Zhang of Anthropic, a company that sells the computing power agents run on, put the test in one line, because explicit branches cost less and give more control: "If you can map the decision tree, build the decision tree." In his framing, a workflow runs model calls through a sequence someone wrote, and an AI agent decides its own path. An agentic workflow is the pattern of work such an agent runs; the Agentic workflows: when not to use them article covers where that freedom is the wrong choice.

Which work automates well?

Work whose answer can be checked automatically and cheaply. In a public talk, Andrej Karpathy, a founding member of OpenAI, separated answers that check themselves (an invoice reconciles or it does not, code passes its tests or it does not) from answers that need judgment, such as whether a summary is good or an email well pitched, and said models improve fastest on the first kind. The same line works for scoping: automate invoice reconciliation and code checks first, and keep a human in the loop, a person who checks the output before it takes effect, on summaries and outreach. As a working rule, a vague process resists automation and the same process with a clear pass or fail becomes tractable, so building the check is often the highest-value engineering.

MIT FutureTech had experienced workers grade AI output from 1 to 9 across 6,648 real, text-based occupational tasks and more than 60,000 evaluations: 4 to 6 was useful but needed revision (augmentation), and 7 was usable as delivered (the automation threshold). Across all models and tasks, 60% of tasks reached usable-without-edits quality, 49.7% reached average-worker quality, and 27% beat the average worker. The scenarios gave the model every piece of information it needed, so the authors call their projections an upper bound, and they name finding information, integrating systems and satisfying regulators as last-mile costs likely to slow delivery, especially for smaller firms. When someone quotes an automation percentage, ask which threshold it uses. Teams that measure AI work describe it in hours saved per week; sellers quote percentages. In an AI transformation, set the threshold before an AI pilot starts.