What leaves with them.

A payroll lead retires after nineteen years. Her successor opens the shared drive and finds everything: procedures, checklists, the calendar of cut-offs. Three months later a site director asks why the night-shift allowance in one plant is calculated differently from everywhere else. Nobody knows. It was agreed with the works council in 2011, it was never written down, and the one person who remembered why has gone.

Everything that could be stored was stored. What left was the part that could not be.

The size of the problem

In October 2025 a survey of 1,000 professionals found that 92% of organisations do not consistently capture knowledge from employees who are about to retire. Fifty-eight per cent of executives described themselves as very worried about that loss, and 85% had not yet put AI to work on knowledge management at all. A 2023 systematic review of 91 empirical studies on turnover-induced knowledge loss reached the same place from the academic side: the loss is real, measurable and mostly unmanaged.

These are not new findings. Michael Polanyi wrote in 1966 that “we can know more than we can tell”, and Nonaka built a whole theory of the firm, in 1994, on the conversion of that tacit knowledge into explicit form and back. ISO even has a management-system standard for it, ISO 30401, published in 2018. What is new is that organisations are now putting AI assistants in front of their documents and expecting them to know what the documents never said.

What leaves is rarely a fact

After enough HR system programmes, a pattern becomes clear. The knowledge that walks out of the door falls into a few kinds, and none of them lives in a record.

1. The reason behind a rule. The policy says 24 months in grade before promotion. Nobody wrote down that the number was chosen to protect one site's apprenticeship scheme.
2. The exception that is always granted. The travel policy excludes one allowance, and every manager in one country approves it anyway, because of an agreement from before the merger.
3. Which field is wrong. The HR system has a cost-centre field that has been unreliable since a migration, and a second custom field that everyone uses instead.
4. Who to ask. The person who knows which pay components count as salary for a given country's reporting is not the person the org chart says.
5. The order things actually happen in. The process document shows five steps. The real sequence has seven, two of them outside the system.

A March 2026 paper on expert-knowledge preservation in the energy sector describes the departure of subject-matter experts as “the irreversible loss of tacit knowledge” that documentation rarely captures. HR is no different, except that its tacit knowledge decides pay.

An assistant inherits the gap

Put an AI assistant on top of a shared drive and an HR system and it will answer from what is written. It will read the policy's 24 months and apply it to the apprentices. It will read the cost-centre field that is wrong. It will describe the five-step process. It will be fluent, sourced, and mistaken in exactly the ways the retired payroll lead would have caught in a second.

A June 2026 paper on knowledge and AI in software teams put it bluntly: the bottleneck is “not model capability but knowledge architecture”. The model is not the scarce part. The written-down reasons are.

Capturing the reasons while they are still here

The useful move is not to document everything. It is to find the places where practice differs from the record and write down why, with the people who know, before an assistant starts answering from the record alone. In our own work with HR teams, that is a structured interview, run early, with three questions for every policy and every system field that matters.

1. Where does what we actually do differ from what the document says?
2. Why? Who agreed it, when, and does the reason still hold?
3. What would a new colleague get wrong in their first month?

The answers become rules an assistant can carry, each with its author and its date, so that the next person can challenge them. That is a modest amount of work, and it is the difference between an assistant that knows the organisation and one that only knows its files.

What to do this quarter

1. List the ten people whose departure would hurt HR operations most, and the five rules each of them is the only one to understand.
2. For each rule, record the reason, the exceptions and the owner, in a form that can be read by a person and by an assistant.
3. Treat the record as versioned: when a reason changes, the old one stays visible with its date.
4. Test any assistant on those rules before you test it on anything else. A system that recites the policy and misses the practice has not learned the organisation.