AI Recommends Termination. Who Owns the Decision?

The manager has two shortcuts in mind: accept the recommendation and avoid a difficult conversation. Neither tells us whether the productivity decline was measured fairly or why it happened. Long tenure is not an automatic defense against a performance decision, just as an AI flag is not an automatic case for termination. The question asks what HR should do before the organization crosses that line.

I call the failure mode algorithmic abdication: a person with decision authority treats a system's recommendation as if it already settled the judgment. A human signature on an approval screen is not meaningful oversight if no one can explain the evidence, challenge the inference, or choose a different outcome.

First, stop the automated notice and preserve the recommendation and inputs for review. Confirm the period, source, definitions, and consistency of the productivity measure. Was the employee compared with people doing comparable work? Did the metric capture output while missing rework, mentoring, safety, or changed assignments? These are questions to investigate, not facts supplied by the scenario.

Next, talk with the manager and examine the employee's actual circumstances under the organization's process. A shift change, equipment issue, workload change, protected leave, accommodation request, or inconsistent coaching might matter if the evidence shows it does. The EEOC explains that algorithmic tools used for employees can raise disability and accommodation issues in the United States. In a real case, involve the appropriate HR, legal, or employee-relations specialists for applicable obligations; this jurisdiction-neutral exam question does not establish that any particular law was violated.

Assess whether the model, the chosen metric, or its application could produce uneven results. Ask what the platform was designed to predict, how it was validated for this use, what alternatives it omitted, and whether similar cases have been handled consistently. Document who reviewed the recommendation, what was checked, what remained uncertain, and why the final action was accepted, modified, or rejected. A performance-improvement plan or further inquiry may be more appropriate than termination; the scenario does not give us enough evidence to decide the employee's fate.

Only after that review should the responsible people make the employment decision. If termination is supported, communicate it directly and through an appropriate accessible channel, with the required process and support. The point is not to ban automation from every administrative step. It is to prevent an automated recommendation or notice from replacing the human decision and its consequences.

The operating-model question is simple: can the system send a high-impact notice before a named person has authority, relevant evidence, and a real opportunity to disagree? If yes, the workflow is designed to turn a recommendation into an action too easily. Put a hold between the flag and the notice, assign a decision owner, make the supporting data inspectable, record overrides and escalation, and give the affected person a way to raise an error through the organization's process.

This is an application of the voluntary NIST AI Risk Management Framework, which calls for defined human-AI oversight roles, context-sensitive risk assessment, and documented decisions. It is not a claim that NIST prescribes this exact six-step employment workflow. For the broader principle, see our human accountability and contestability walkthrough; for a distinct pre-launch bias problem, see the global performance-system rollout scenario.

For the SHRM-CP question, the priority is the earliest unresolved risk: whether the recommendation deserves to become an employment decision at all. Know the concepts, understand the limits of the data, and apply the judgment under pressure.

Frequently asked questions