The AI recommendation that changed after a human challenge
Validating consequential workforce recommendations and testing human oversight
PAWorld27 Zurich · 3–4 March 2027 · Zürich
About this session
An AI-supported capability assessment recommends a group of employees for redeployment. In this worked scenario, data-quality and model checks uncover missing experience records and uneven coverage across teams, and a structured review adds work-sample evidence and employee correction that materially change the recommendation.
The session examines how to test whether data and models are fit for a decision that affects people, how to design human review that actually catches errors, and what limits, appeal routes and performance checks should come before wider use.
Learning outcomes
- Assess whether data and model performance are fit for a particular workforce decision.
- Design a review that can detect material errors and record justified overrides.
- Set limits, challenge routes and stop conditions before expanding model use.
Strategic questions
Which consequences require a higher evidence threshold or a human decision?
Context
Workforce recommendations can look plausible despite missing experience data or uneven model error. The consequences differ sharply between information support and decisions affecting a person. Human review must be designed and tested rather than assumed effective, with a way for employees to correct the evidence and challenge the result.
