top of page

AI Certainty Plus Outcome Feedback Makes Delegation to AI Work

2 hours ago
3 min read

Image via Wikimedia Commons (public domain).

When employees can choose, task by task, whether to hand work to AI, what information helps them choose well? Taudien, Fügener, Gupta and Ketter find in Information Systems Research that showing AI certainty or outcome feedback alone does not help, but showing both raises delegation and lets human-AI teams beat either working alone.

Two signals, one better decision

Much of the debate on AI at work assumes a simple choice between human and machine. In many real settings the arrangement is subtler. In visual inspection, content moderation or applicant screening, the AI sits next to an employee who keeps the decision right: for each case, the person decides whether to handle it personally or pass it to the system. The promise of this setup depends on people delegating the right cases, and prior work shows they often do not. Anna Taudien, Andreas Fügener and Wolfgang Ketter (University of Cologne) and Alok Gupta (University of Minnesota) ask which kind of system information helps users delegate better.

The team compared two types of information. The first is the AI's certainty, shown before the user decides whether to delegate a case. The second is outcome information, shown afterwards, telling users whether the AI's decision turned out to be correct. In an earlier conference version of the project (ICIS 2022), the authors tested these two features alone and in combination in a 2x2 between-subject experiment with 560 participants.

According to Taudien, Fügener, Gupta and Ketter (Information Systems Research, 2026), each type of information on its own had no positive effect and in some cases even reduced effective delegation. Combined, they increased how often people delegated and helped them send to the AI precisely the tasks where it held a performance advantage. With both signals in place, human-AI teams outperformed both the AI working alone and humans working alone.

The logic is intuitive once stated. A certainty score tells users where the AI thinks it is strong, but without feedback they cannot learn whether to believe it. Outcome feedback tells users when the AI was right, but without a certainty signal they cannot map that track record onto the next case. Together, the two signals let users build a usable picture of where the system's error boundaries lie. The study's scope is the setting it names: employees who retain decision rights over AI use on repeated, case-by-case tasks. The published abstract does not report effect sizes, so readers who want magnitudes should consult the full article.

Why it matters: For researchers who study human-AI collaboration, the paper is a clean demonstration that design features can interact: testing them one at a time could wrongly conclude that neither works. For researchers who use AI themselves, for example to code qualitative data, screen literature or classify text, the same lesson applies at the desk: ask for the model's confidence and keep checking its outputs against ground truth, because neither habit alone tells you when to trust it. And for organizations, the finding offers a reassuring counterpoint to automation narratives, since well-informed humans still add value once AI enters the decision.

Taudien, A., Fügener, A., Gupta, A., & Ketter, W. (2026). Enhancing AI Use: How Complementary System Information Drives Delegation Frequency and Effectiveness. *Information Systems Research*. Read the original. Independent editorial summary of third-party research.

Comments


This initiative is supported by the following organizations:

  • Twitter
  • LinkedIn
  • YouTube
logo_edited.png
bottom of page