When companies use AI in performance evaluations, disciplinary actions, or termination processes, managers usually retain final approval authority. However, adding an approval button to the interface does not mean the manager has conducted a substantive review. If managers can see only the system’s conclusion, have no time to verify the underlying evidence, and lack the authority to return or stop the process, human approval merely rubber-stamps an automated outcome.
What companies need to preserve is the manager’s ability to exercise independent judgment. This requires clearly defining who is responsible for the review, what evidence must be examined, which circumstances must prevent the process from moving forward, and how to maintain traceable records. AI can support the earlier stages of a process, but high-impact HR decisions should still be made by people with the appropriate authority and accountability, who must also be able to explain their reasoning.
Why might a manager’s approval still fall short of substantive review?
According to iThome, California signed SB 947 into law, becoming the first US state to prohibit employers from relying entirely on AI to make disciplinary or termination decisions. Companies may still use AI in workforce management, but people cannot simply accept the system’s output; they must conduct a substantive review. Passed during the same period, SB 951 requires companies to disclose information about mass layoffs, relocations, or shutdowns caused by AI systems. AB 1883 prohibits employers from using AI workplace monitoring systems to infer employees’ emotions or psychological states.
The message for Taiwanese companies is straightforward: a manager’s involvement in a case is not the same as actually exercising judgment. If managers see only scores and recommendations generated from attendance, performance, or customer complaint records, but cannot inspect the data sources, applicable rules, or the employee’s explanation, it will be difficult to demonstrate that they made an independent judgment. System output can serve as a reference, but it cannot replace verification of the facts and circumstances.
Companies do not need to subject every routine HR activity to the strictest procedures, as this would make processes difficult to carry out. HR, IT, and legal teams can begin by identifying performance evaluations, disciplinary actions, terminations, and other processes that materially affect employee rights. They should then mark where AI is involved and determine which stages must never proceed automatically. The first question should be: Which high-impact HR processes currently generate recommendations or move into the approval stage before managers can review all the underlying evidence?
What should managers review and do as part of substantive human review?
Companies should turn substantive review into an actionable procedure. It is not enough for a policy to state that “a manager must confirm” the decision. At a minimum, four points must be clearly defined.
Assign responsibility for the review. The company should specify which role is responsible for reviewing the case, whether sign-off from HR or another manager is required, and who will step in if the reviewer has a conflict of interest. Simply referring to the “responsible manager” is often insufficient, especially in cases spanning multiple sites or involving a manager’s direct reports, where accountability can easily become unclear.
List the required evidence. Reviewers should not look only at scores, summaries, or recommendations generated by AI. They must also examine the work rules the company actually applies, relevant records, and the employee’s explanation. Different decisions require different information, so companies should define the mandatory review items for each process. Managers should also be told which information has not yet been obtained so that missing data is not mistaken for the absence of a problem.
Establish stop conditions. The process should stop when information is incomplete, sources conflict with one another, the case falls outside existing policies, or the manager cannot explain the decision based on company rules and the facts of the case. Managers must have the authority to withhold approval. They must not be pressured by deadlines or hierarchy to approve a decision simply because the system has already issued a recommendation.
Define how cases must be returned. Returning a case should involve more than writing “Please provide additional information.” The reviewer should specify which records are missing, which conclusions require re-examination, and who will review the case again after corrections are made. Without clearly defined return requirements, managers may simply forward the same result up the chain until it is eventually approved without any substantive change.
Companies can begin by piloting these controls in one high-risk process. They can run an existing blank form or a simulated case through the process to verify whether reviewers can access the required evidence, stop or return the case, and provide reasoning clear enough for another manager to understand. Once these four elements have been built into one process, the same approach can be extended gradually to other high-impact HR decisions.
How can human approval genuinely stop a process and create an audit trail?
Even the clearest policy will not work if the system allows the process to bypass approval checkpoints, leaving managers to sign off only after the fact. Companies should therefore configure human approval as a mandatory gate in the workflow. Once the preceding stage is complete, the designated role must verify the required evidence and stop conditions. The system must not proceed to the next stage until that person has approved it. The approval interface should also require managers to document their reasoning. When returning a case, they should be required to specify what must be re-examined, rather than choosing only between approval and rejection.
Tasks in EgentWrX can be connected into workflows, with a human approval gate set before each handoff. The process pauses until a person authorizes it to continue. This mechanism is well suited to critical checkpoints in high-impact HR processes because it makes human review a mandatory step that cannot be skipped. However, companies must still define the approver, verification items, stop conditions, and return rules. Platform mechanisms cannot replace sound process design.
Companies also need to retain more than the final outcome. HR, IT, and legal teams should jointly define a record-retention checklist that can answer, at a minimum, which process activities involved AI, who approved or returned the case and when, and whether the system ever continued the process without approval. EgentWrX audit logs can be queried and exported, while a hash chain can be used to verify record integrity. The system also records when an administrator downloads audit records.
Implementation teams can sample high-impact HR processes every quarter to confirm that records before and after approval form a continuous trail, that reasons for returning cases are specific, and that no process continued without approval. Building a human approval gate and record-retention checklist for one process, then conducting a joint walkthrough with HR, IT, legal, and the managers responsible for approval, comes much closer to substantive human review than simply adding an approval button.
FAQ
Can AI be used in performance evaluations or disciplinary processes?
Yes, but AI output should be treated as a recommendation or reference rather than directly determining a high-impact employment action. Companies should first identify the checkpoints that require human review. Managers should then verify the required evidence, the employee’s explanation, and the applicable rules before making an independent decision and documenting their reasoning.
Does clicking “Approve” mean that a manager has completed human review?
No. The act of approval proves only that someone handled the case. To demonstrate that the review had substantive content, the company must also confirm that the manager examined the required evidence, could challenge the system’s output, had the authority to stop or return the process, and documented the reasoning behind the decision.
When should an HR decision-making process be stopped?
The process should stop when information is missing, sources conflict, the case falls outside existing policies, or the manager cannot explain the decision based on company rules and the facts of the case. Companies should define these stop conditions in advance so that managers do not simply accept system recommendations under deadline pressure.
What information should audit records contain at a minimum?
At a minimum, records should show how AI participated in the process, when and by whom a case was approved or returned, and whether the process stopped at the approval checkpoint as required. Companies should also retain the reviewer’s reasoning and the items identified for re-examination so that internal auditors or dispute handlers can reconstruct how the decision was made.
References
- California becomes the first US state to prohibit AI from making employee termination decisions on its own — iThome’s overview of California’s workplace AI requirements under SB 947, SB 951, and AB 1883.