What should actually be automated?

What should actually be automated?

Do not begin with jobs or departments. Begin with units of work. A role combines routine execution, exception handling, relationship, accountability and learning; treating it as one automation target usually destroys the context needed to make sound design choices.

Decompose the flow and score each activity on frequency, variability, judgement, consequence, reversibility and observability. Stable, frequent work with explicit rules belongs in conventional automation. AI assistance fits variable tasks where a person can cheaply verify the output. Supervised or autonomous execution requires bounded actions, reliable detection and consequences the organisation can absorb.

The evidence supports selectivity. Stanford�s 2026 AI Index finds the largest productivity gains in structured, measurable work: roughly 14�15% in customer support and 26% in software development. Yet agents still failed about one in three attempts on a structured computer-use benchmark. Capability and control must be evaluated at task level, not inferred from a general model score.

Redesign before automating. Remove unnecessary approvals, clarify ownership, standardise inputs and decide how exceptions should travel. Otherwise technology accelerates queues, rework and low-value controls. Calculate economics across the full process, including review effort, integration, errors, recovery and work pushed to customers or downstream teams.

Finally, protect the capability to learn. If automation removes the cases through which junior employees develop judgement, create deliberate practice and escalation pathways. The best boundary is dynamic: automate what is understood and safely observable, augment where context remains human, and revisit the division as evidence�not enthusiasm�changes.