How Far Can You Trust AI-Recommended Internal Moves?

Adoption is still early, but the regulation arrived first

Not many companies in Korea have yet built AI-based internal mobility matching into their actual assignment process. In SHRM's 2026 survey, about 39% of organizations had adopted AI in the HR function, and more than half had not started at all. Narrowed to Korea, that share is likely lower still.

The regulation, however, is already in force. With the Framework Act on the Development of Artificial Intelligence and the Establishment of a Foundation for Trust (the AI Basic Act) taking effect on January 22, 2026, and its enforcement decree applying in full from July 21, HR functions including recruiting were classified as high-impact AI. Providers and adopting organizations of high-impact AI carry obligations for risk management, securing explainability, human oversight, and documentation related to safety and trust.

The reason the regulation singled out recruiting is simple. AI is taking part in decisions that change people's careers. Internal mobility is no different. Candidates are narrowed, ranked, one is chosen, and the rest are not. That outcome changes a person's job, their reporting line, and the career path that follows, through a formal assignment. Regardless of how far the regulation reaches, the requirement to be able to explain why it was this person does not change.

If anything, that requirement arrives from closer by. An external applicant who is turned down mostly leaves, but the person passed over for an internal move comes into the same office next week and can ask the person in charge directly. Reversing it works differently too. A bad hire can be ended quietly, while a bad assignment can only be undone by another assignment, and that record stays in the individual's HR history.

So in internal mobility, "can you explain why you accepted that recommendation" matters more than "did you get a good recommendation." The four points below are the conditions that make that explanation possible, and they are also requirements to confirm in advance, back when you are choosing the tool.

Verification point 1. What was the recommendation based on

When AI recommends an internal mobility candidate, that judgment comes out of a particular combination of data. Skill tags, past project experience, performance history, and collaboration data within the organization are typical inputs. The problem is that this combination works differently from one match to the next. One recommendation may be driven mainly by skill fit, another mainly by performance history.

The first question for the person in charge is a simple one. Can you explain why this candidate came out first, and which data items weighed most heavily? Proceeding with an assignment without being able to explain it leaves you with no grounds to respond when an objection comes in later. Overseas HR risk material makes the same point repeatedly: if a decision was made on the basis of an AI output, you must be able to explain why you accepted that decision, and a decision you cannot explain is hard to defend in an audit or a challenge.

Here is the scene that plays out often in practice. Three candidates are listed in order with match scores of 87, 82, and 79, and the person in charge proceeds with an assignment for the top candidate without knowing whether that gap comes from skills, years of experience, or recent project performance. When the second-ranked candidate raises an objection, that person is left holding a score and no basis to explain it.

In practical terms, check it this way. On the recommendation screen, look first at whether the basis for each candidate, skill fit, career similarity, performance data and the rest, is broken out item by item, or whether it comes as a single score or ranking only. If the tool shows a score without the reasoning behind it, that recommendation is safer used as reference material than carried straight through as the basis for an assignment.

Verification point 2. Is the data itself already skewed

The second thing to check is the structure of the data used in the matching. Because AI works by learning from past data and finding patterns in it, a skew that already existed in your historical HR data can carry straight through into the recommendations. Candidates from one particular division showing up unusually often, or only those above a certain tenure reaching the top of the list, are typical patterns.

If past fast-track promotions were concentrated among people from one division, for instance, an AI trained on that data is likely to keep recommending people from the same division at the top. This is not the AI malfunctioning. It is closer to an honest reproduction of a bias that already existed. The problem is that the more that reproduction repeats, the more the bias accumulates as data again and gets reinforced in the next round of recommendations.

Gartner emphasizes, as a leadership-level principle for AI-assisted HR judgment, that human decision authority must be made explicit, that bias and explainability standards must be enforced, and that changes in the model, meaning drift, must be monitored continuously. In other words, this is not something you verify once and close out. It is something you check periodically, to see whether matching results keep tilting toward a particular group.

There are two practical ways to check. First, gather the results of the last several recommendation rounds and look at the distribution by division, tenure, and job family. If candidates cluster under certain conditions, treat data bias as a possibility. Second, check in reverse for anyone excluded from the recommendations who would in fact have been a good fit. Looking only at who was recommended makes bias hard to spot. It is by looking at who was not that the missing pattern becomes visible.

Verification point 3. What is it that AI cannot see

The third point is context that never becomes data. Information that is not recorded in any system, such as relationships within a team, a recent reorganization, or an informal division of work, often decides whether an assignment succeeds or fails. AIHR explains that even when AI is applied to internal mobility, managers must be involved at the key decision points, not because AI gets things wrong but because there is information AI cannot reach in the first place.

A match can be perfect on the data and still miss that the candidate was recently at the center of a conflict within the team, or that the receiving department is about to be reorganized. That kind of information usually exists only in the heads of the person in charge and the manager on the ground, and this is exactly where human judgment has to complement the AI's recommendation.

There is one more thing to check here. However carefully you work through the first three, if final approval authority effectively sits with the AI's ranking, the verification is a formality. Even if the person in charge reviewed the matching rationale and checked for possible bias, proceeding with the top-ranked candidate as-is leaves no trace of their judgment between the verification steps and the assignment outcome. Verification means something only when, faced with an objection, the person in charge can explain for themselves why they accepted that ranking and on what basis they finalized it.

The checklist below organizes those four verification points so they can be applied directly in practice. Use it as a self-check before carrying matching results into a proposed assignment.

Carrying a recommendation ranking straight into a proposed assignment without checking these four is the same as executing an unverified judgment in the organization's name. Conversely, once the practice of checking all four every time takes hold, AI recommendations settle in not as a tool that replaces the judgment of the person in charge but as one that raises the speed and accuracy of that judgment.

People make the judgment, elizax makes the basis checkable

Working through this checklist by hand every time is a realistic burden. Extracting the underlying data separately each time a match comes out, and gathering past recommendation history to calculate distributions, is not something one person can keep repeating.

The Internal Mobility feature of the elizax People Intelligence Agent resolves the transparency problem covered in verification point 1 right at the matching screen. Rather than showing a candidate as a single score or ranking, it presents where their capabilities and career matched an internal position in a form the person in charge can inspect item by item.

The data-bias problem covered in verification point 2 is likewise eased, because the structure lets the person in charge reopen recommendation history item by item without tallying separately whether recommendations are tilting toward a particular organization or tenure band, which reduces the burden of checking for bias. The organizational-context problem covered in verification point 3 also holds, because the Internal Mobility feature is designed to place matching results in front of the judgment of the person in charge and the manager as evidence rather than to make the final decision for them, so the principle that people still fill in what the system does not hold remains intact.

elizax sits on the same foundation as hunel's assignment data. Recommendation, review, final confirmation, and assignment remain as one unbroken history. It is a structure that turns "a person made this judgment" from a claim into a record you can look up, and that becomes a practical foundation for meeting the documentation requirements of the AI Basic Act.

If you want to know how the verification points covered in this guide could be applied to internal mobility in your own organization, talk with HCG directly.