Zusammenarbeit zwischen Mensch und KI

A useful arrangement specifies what the system can propose or do, what evidence a person sees, and when the person can correct, stop, or override it.

Wichtige Erkenntnisse

• Make proposals and completed actions visibly different.
• Give reviewers evidence and authority.
• Measure the combined human-system outcome.

Begin with a task analysis. Identify repetitive work the system can support and judgments that require context, accountability, or expertise. Adding a human approval button is not enough if the reviewer lacks time or information to evaluate the proposal. Design the handoff carefully. Show the relevant source, uncertainty, action consequences, and meaningful alternatives. A recommendation should be distinguishable from an action already taken. Keep cancellation and escalation available at the moment they matter. Evaluate the team rather than only the model. A suggestion that is usually correct may still reduce overall performance if people become less attentive or must spend excessive time checking it. Measure completion quality, review burden, and error recovery with realistic users and tasks. Assign responsibility for maintaining the workflow. People need to understand the system’s limits, and reported mistakes should reach someone who can change the product. Preserve a usable manual path when automation fails or when a task falls outside the evaluated conditions.

1. Imagine an assistant suggesting a refund after reading a support conversation.
2. Show the request, applicable policy passage, amount, and proposed action before approval. Do not require the reviewer to reconstruct those facts from separate screens.
3. Test whether reviewers catch deliberately incorrect suggestions under realistic time pressure.

This constructed workflow measures whether the review step actually helps prevent mistakes.