Coordinate and act
The Mixed-Initiative Seam
Build the handoff so the model raises its hand on hard cases, instead of a person checking every case equally.
Developed by Superadditive in our own review of the research. The term "mixed-initiative" comes from Eric Horvitz's research on systems where people and computers share control.
Changes a rule or role · An hour or more to set up · Anyone on the team can start it
Try this first
Start with two triggers. The model flags when it is unsure, and when the decision is hard to reverse.
Use it when
People are either checking every AI output by hand or checking none of them.
Skip it when
You need a setup that can signal its own confidence and read the stakes. Without that, fall back to human judgment about what to check rather than trusting the model to escalate.
How to introduce it
Let the model decide when to bring in a person. It escalates when its confidence drops, and when the decision is high-stakes or hard to reverse. The alternative, a person checking every output the same amount, spends attention where it is not needed and starves the cases that are.
How to show up
Work with the team on the escalation triggers, low confidence and high stakes. Resist both extremes, check everything and check nothing. Setting the triggers well is the craft.
How long it takes
A design session to set the triggers. Ongoing tuning against real outcomes.
What makes it hard
Confidence signals are imperfect, so tune the triggers against what goes wrong and do not over-trust the model's self-assessment. The stakes threshold is a judgment the team owns, not a number the tool hands you.
What it looks like when it's working
Attention lands on the risky and uncertain cases while safe ones flow through. People still checking everything, or dangerous cases not escalating, means the triggers need work.
How long until it sticks
Several rounds of adjusting the triggers before the escalation is well-calibrated.
How you know it stuck
The team designs handoffs by asking when the model should pull in a person.
The idea behind it
Uniform checking wastes attention on safe cases and under-attends dangerous ones. Let the risky cases pull attention toward themselves.
Where it comes from
our review. Uniform verification wastes scarce human attention on the safe cases and under-attends the dangerous ones. Design the handoff so the AI escalates to a human when its confidence is low or the stakes and irreversibility are high, rather than the human uniformly checking everything.
Evidence
No outside source. This move came out of our own review of the research, so treat it as reasoned judgment rather than tested practice.
How Galgo scaled fraud control with exception-based AI review, Zapier Blog A lending company had AI check delivery photos against set rules and sent only the flagged cases to people. Manual review fell from about 150 hours a month to about 60 photos a month. These figures come from a vendor case study.