AI in HR: the layer your HR stack cannot see

Four tabs are open on the HR manager's screen. The May employee survey shows a 7.2 on engagement, absence rates are below the sector average, this quarter's exit interviews have been entered neatly. And still, this afternoon she is sitting across from the third nurse in four months to hand in her notice, someone who never surfaced as a risk in any of those four tabs.

AI in HR has delivered a great deal in recent years, but not this. Your HR stack records what people have done, how they feel about something on average and how they position themselves on a standardised scale. What lies beneath those outcomes, why someone reacts differently this week than a year ago, and which decision they have been weighing at home for three months, ends up in no field at all. That is not a failing of your instruments but a limit to what an instrument can do, and that limit is precisely where the conversation starts.

What does your HR stack do well?

It is tempting to open a piece like this by claiming that measurement delivers nothing. That is untrue, and it helps no one.

An HRIS keeps the workforce records in order and makes visible how turnover, contract types and absence are distributed across departments. That is the basis on which every conversation with a board member rests. An employee survey gives you a comparable, repeatable measurement that shows whether something is moving and where the outliers sit. An assessment or a DISC or TMA profile gives language and structure to what people otherwise only sense; it makes it discussable that one person wants to consult first and another wants to decide first, without attaching a verdict. And AI applications in HR make work that used to take days feasible in minutes: grouping open answers, spotting patterns in turnover, summarising large volumes of text.

The professionals working with these instruments are therefore not doing shoddy work. They are working with tools that are good at comparing, and comparing requires fixed categories. That is exactly where the limit sits.

Where does the measurable end?

Each of these instruments shares one assumption: that what you want to know can be expressed as an answer to a question thought up in advance. As long as that holds, it works excellently. It breaks down on everything someone does not say.

In 2023, in the Annual Review of Organizational Psychology and Organizational Behavior, Elizabeth Morrison summarised a decade of research into what employees speak up about and what they withhold. Her central point is that silence is not the absence of an opinion but an active choice, made on the basis of an estimate of what speaking up would cost. Silence and voice are therefore two distinct behaviours with different causes, and they can occur at the same time: the same employee reports a rota problem and keeps his judgement about the reorganisation to himself (Morrison, 2023). Morrison also notes that the field still leans heavily on self-reported data from single-moment studies, which makes measuring this particular phenomenon difficult.

For an HR department that means something uncomfortable. An employee survey measures what people are willing to write down on a form sent by their employer. Those who stay silent do not disappear from the data; they appear in it as an average. The 7.2 in the opening scene is not a lie, but it is the average of people who are perfectly content and people who have concluded there is no point in filling in anything else. From the score you cannot tell which of those two groups is growing.

What exactly is the undercurrent?

The undercurrent is the whole of drives, experiences and considerations that explains why someone behaves the way they do, and that rarely fits a standardised question. It is the layer in which a nurse has come to see herself, since last year's reorganisation, as an executor rather than a professional. It is the reason a team leader pulls every escalation towards himself, which has to do with a previous manager under whom he raised something too late.

That layer is neither vague nor beyond reach. It surfaces the moment someone recounts an event instead of assessing themselves. The question of how satisfied you are with your work produces a number. The question of which week of the past year you would most like to live again, and why that one, produces a story containing what someone values, where they lost it and what they need to get it back. Same person, same moment, a fundamentally different kind of information.

Why is this urgent now?

Because the gap between what people do and what the organisation sees of it demonstrably widened this summer.

Researchers at the Dutch central bank published an analysis in the economics journal ESB in August 2026 on AI use in the Dutch workplace. In every sector, the share of employees who say they use AI applications is higher than the share who experience active support from their employer. In most sectors that difference exceeds fifteen percentage points; in government and education it runs up to nearly thirty. Of all employees surveyed, a quarter experience active support from their employer (ESB, 2026).

What is happening there, in practice, is that people are reinventing the way they work without anyone writing it down. A policy officer who now finishes his memos in a third of the time also changes what he thinks of his work, what he is proud of and what he still needs in order to keep learning. None of those shifts reaches the job architecture, the training catalogue or this autumn's employee survey, because that questionnaire was drawn up when the situation was different. The faster the work itself changes, the older the categories you measure it with become.

What does listening deliver that measuring does not?

A useful Dutch example arrived this month. For his master's research at Leiden University, Youri van der Velden studied how bottom-up change comes about in healthcare, distinguishing two kinds of leadership behaviour: listening to employees, and taking action. His results show listening to be more effective than taking action. Care professionals who notice that they are being listened to feel safer about initiating a change themselves or about voicing criticism of how things are run. Team climate strengthens that effect: in a team where people feel safe, they speak up sooner (Van der Velden, 2026).

That result is more striking than it sounds. A manager who receives a signal reflexively wants to do something, because doing feels like taking it seriously. The research suggests the order runs the other way: first hear what is actually going on, and only then act. Whoever jumps straight to the solution gets a slightly flatter answer next time, and after three rounds a 7.2.

We saw that same difference around absence. In early signs of absence your dashboard never sees we described how a dashboard records who has dropped out, while the run-up to that moment is audible months earlier in how someone talks about their work.

How do you place the story layer alongside your existing instruments?

Not by replacing anything. Your measurement instruments stay exactly where they are; the story layer sits underneath them and answers a different question.

In practice that means three things you can organise independently of each other. First, attach every outcome to a conversation in which the event, not the score, is central. A 7.2 on engagement becomes usable the moment you hear from five people which week explains the number. Second, frame the question so that someone can recount rather than assess, and leave the outcome of that conversation with the employee rather than in a manager's dashboard. What a manager may know is that the conversation took place, not what was said in it. Third, repeat it at a rhythm that matches how fast the work is changing, because a story from eighteen months ago describes a job that has since become a different one.

At NarraTyx we are building exactly that layer, using frameworks from the narrative tradition such as the Hero's Journey and Schwartz's theory of values. It is a reflection instrument and not a test: it produces no score, no ranking and no prediction, but anchors, tensions and the questions that go with them. Where that comes together at organisational level, we describe on the page about the connecting story. The score is the beginning, the story completes it.

Which question do you ask once the employee survey is in, who do you ask it to, and does a number belong to it?

Sources

• Morrison, E. W. (2023). Employee voice and silence: Taking stock a decade later. Annual Review of Organizational Psychology and Organizational Behavior, 10, 79-107. https://doi.org/10.1146/annurev-orgpsych-120920-054654
• Van der Velden, Y. (2026, 19 August). Luister eens: zorg dat medewerkers zich gehoord voelen. Platform O. https://platformoverheid.nl/luister-eens-zorg-dat-medewerkers-zich-gehoord-voelen/
• ESB (2026, August). Werknemers lopen voor op werkgevers bij AI-gebruik. Economisch Statistische Berichten. https://esb.nu/