The cheap draft has hidden costs

A manager celebrates that a report now takes minutes to generate. A colleague spends the afternoon checking its references, correcting assumptions, and reconciling it with existing records. The production time fell; the total work may have moved somewhere less visible.

Generating a report quickly can move work to the colleague who must check it. The saved minutes are meaningful only after the whole task, including repair and integration, is considered.

Efficiency should be measured across the workflow. Include preparation, verification, revision, integration, and the cost of errors. More output can create more demand for attention. A team that produces ten times as many drafts may not be able to review ten times as much work responsibly.

There are genuine gains when assistance removes low-value repetition or helps a person complete a task better. Those gains need not be denied. But they should be demonstrated with an appropriate comparison rather than inferred from the speed of a single generated response.

More output also creates more demand for attention. Sometimes the useful improvement is fewer documents with clearer decisions, not a larger stream that nobody has time to examine responsibly.

Teams can run bounded experiments with quality measures chosen in advance. Ask reviewers what new burdens appeared. Decide when not to generate another document. Sometimes the most productive intervention is fewer outputs with clearer decisions.

A cheap draft is useful when it makes the whole task better. It is less useful when it invoices someone else in attention. The economics of AI need to count human checking as work, especially when the person doing it has little power to make the invoice visible.

But they should be demonstrated with an appropriate comparison rather than inferred from the speed of a single generated response.

Has the work disappeared, or moved to someone less visible?