Complement, don't compete: designing work where humans and models each do what they are for
The productive question is not "can the model do this job?" but "which parts of this job should it do, and how do we design the handover?" A short field guide to division of labour.
When teams first bring a capable model into a workflow, they tend to ask a binary question: automate or not. The better results come from a finer-grained one. Every job is a bundle of tasks, and the tasks differ in how well-specified they are, how costly an error is, and how much the outcome depends on context a model cannot see.
A simple sorting rule
Give the model tasks that are well-specified, high-volume and cheap to check. Keep for humans tasks where the goal is ambiguous, the error is expensive or irreversible, or the relevant context lives in a relationship rather than a document. Design the handover deliberately: what the model passes to the human, in what form, with what confidence, and what the human is expected to do with it.
The handover is where it goes wrong
Most failures we investigate are not model failures. They are handover failures: a recommendation presented as a decision, a confidence score nobody was trained to interpret, an override path so cumbersome that reviewers rubber-stamp. Automation bias is a design outcome, not a character flaw. If the interface makes agreeing effortless and disagreeing laborious, people will agree.
Design the disagreement path as carefully as the agreement path.
Measuring complementarity
Track the override rate and its direction. A reviewer who never overrides is not reviewing. A reviewer who always overrides has a model that should not be there. Track outcomes by who made the final call — human alone, model alone, both — and let the data move the boundary. Complementarity is not a fixed allocation; it is a setting you tune.
Done well, this produces work that is both faster and more humane: the model absorbs the volume, the human absorbs the judgement, and each is measured on what it is actually for.
