Automation Is Cheap Until You Have to Check the Work

Decisions about automating a task tend to be framed around capability: can the system produce an acceptable output? That framing is incomplete, and the missing half is usually where the economics actually live. The more useful question is what it costs to confirm the output is acceptable — and who is responsible when nobody checks.

Verification cost as the deciding variable

Some outputs are trivially checkable. A total either reconciles or it does not; a file either parses or it fails. Others require the same expertise that producing the work would have required, in which case automation has moved the labour rather than removed it.

This explains an otherwise puzzling pattern in adoption. Tasks that look similar in difficulty diverge sharply in practice: drafting, summarising and triaging spread quickly, while approvals, diagnoses and anything with legal consequence move slowly. The dividing line is rarely capability. It is the cost of being sure.

Questions that tend to settle the decision

• —What does it cost to check this output, in time and in expertise?
• —What happens if a wrong result goes through unnoticed?
• —Who is accountable for that outcome, by name?
• —Is the error rate knowable, or merely assumed to be low?
• —Would a sample-based check be sufficient, or is every case load-bearing?

Designing around it rather than hoping

Where verification is expensive, the workable pattern is narrowing scope until checking becomes cheap: constraining inputs, requiring structured outputs that can be validated automatically, and routing anything unusual to a person by default. That is less impressive than full autonomy and considerably more durable.

A system that is right most of the time and confident all of the time transfers the hard part of the job to whoever has to tell the difference. Naming that person, and resourcing them, is the part of automation that determines whether the saving is real.

Why this framing ages well

Models will keep improving, probably faster than most plans assume. Verification cost is a more stable lens precisely because it does not depend on forecasting capability: it asks what a specific error would cost here, and whether anyone would notice. Those answers change slowly, which makes them a better basis for a decision.