skip to content

Analysis / Business automation and AI evaluation

Choose an AI abstention threshold

Choose an AI review threshold in Proforma using labeled outputs. Compare coverage, error costs and review volume without assuming confidence scores are calibrated.

A starting point for

AI risk owner

Bring this question into Chat

Your starting prompt.

Analyze [labeled AI outputs] under candidate [review thresholds]. Show coverage, error types, review volume and missed high-cost mistakes for each threshold, retaining class and subgroup sample counts where permitted. Test whether scores correspond to observed correctness before calling them calibrated confidence. Propose a review threshold and escalation criteria for owner review. Do not activate autonomous decisions or claim the threshold guarantees safety.

Copy this starting point, then replace the placeholders with your own inputs.

Method and review

Keep the assumptions visible.

Define inputs, outputs, permissions and failure conditions before implementing an Action. Evaluate AI against representative labeled examples and retain human approval for consequential steps.

For this task, check

The workflow owner should approve error trade-offs and verify score calibration on representative new cases.

This starting prompt has no published worked example yet. Begin with a small sample and check the calculations before relying on the result.

Before you begin

  • Reviewed output labels and permitted score data
  • Explicit error consequences and review capacity

Work with permitted files or queryable data. Verify the result against the original records.

Take the next step

Make it your investigation.

Copy the prompt above, create or open your Workspace, and paste it into Chat. Add the inputs you have permission to use and adapt the assumptions with your agent.

Start exploring

Your prompt is carried by copy and paste. Signing up does not submit it or start this work automatically.

Or start with your own agent

The setup prompt includes this public task. Your agent will help you connect when needed and use the access you approve.

Choose Claude, Codex, or Cursor—or copy the prompt to use anywhere.