Knowledge retention in manufacturing: the complete overview

A production line is not running as it should. An operator with thirty years of experience walks past, listens for a few seconds, and adjusts something. The line runs again. Nobody asks why, because it always works.

In two years that operator will retire. The question is not whether the company will lose something, but what exactly, and whether anyone captured it in time.

This article gives the overview: what knowledge retention is, why it is more urgent now than ten years ago, what knowledge loss costs a production environment, and which approach holds up in practice.

What knowledge retention is

Knowledge retention is the set of measures an organisation uses to prevent critical shop-floor knowledge from disappearing when people leave, change roles, or retire. It concerns the knowledge that keeps the work running: machine quirks, exceptions, customer-specific agreements, and the judgement that determines when to deviate from the standard.

Four terms are often used interchangeably, while they mean different things.

• Knowledge retention is the overarching goal: the knowledge remains available to the organisation, even when the person leaves.
• Knowledge capture is the act: getting knowledge out of someone's head and into a system.
• Knowledge preservation is the upkeep: keeping what has been captured accurate and reliable.
• Knowledge transfer is the moment of handover, usually to a colleague or successor.

More on the three components: knowledge capture, knowledge preservation and knowledge transfer in the separate articles in this overview.

Knowledge management is the broader discipline around it, with policy, roles, and systems. In production environments, knowledge management often breaks down at one point: it is designed for people behind a desk, not for people with tools in their hands. Anyone who takes knowledge retention in manufacturing seriously therefore starts not with the system, but with the moment when the knowledge arises.

Why knowledge loss is more urgent now

Knowledge loss is not new. The scale is.

The Dutch working population is ageing quickly. According to Statistics Netherlands (CBS), more than a quarter of all workers were aged 55 to 75 in 2023, against a fifth ten years earlier. CBS explicitly points to a strong increase in over-55s in industry. In provinces such as Limburg, nearly one in three workers was over 55 in 2023.

The European picture is the same. Eurostat calculated that the median age in the EU at the start of 2025 was 44.9 years, and that there are now just over three people of working age for every person aged 65 or older. The European Centre for the Development of Vocational Training, Cedefop, shows what that means for the labour market: more than nine out of ten future vacancies arise not from growth, but because departing workers must be replaced, and retirement is the most common reason. In an earlier analysis, Cedefop already pointed out that workforce ageing can lead to the loss of critical organisational knowledge and experience as people leave.

On top of that, pressure on the sector itself is increasing. In the report Without robotisation, Dutch manufacturing will disappear (TNO, 2026), researchers mention ageing, structural staff shortages, and high labour costs in the same breath, and state that industry productivity must rise by roughly fifty percent over the next ten years to remain internationally competitive. Automation is mentioned as one route. But whoever automates must first know how the process really works, including the exceptions. That knowledge is usually not in the system.

Which knowledge disappears exactly

Not all knowledge is equally vulnerable. Work instructions, drawings, and manuals are captured somewhere. What disappears is the layer beneath.

Philosopher Michael Polanyi called this tacit knowledge and summed it up in the phrase: "we know more than we can tell" (Polanyi, The Tacit Dimension, 1966). On the shop floor it is usually called tribal knowledge. It is knowledge you build by doing something a hundred times: the sound that changes before a fault appears, the setting that always needs to be slightly different on one machine, the customer for whom you deviate from the standard.

Research into tacit knowledge elicitation for shop-floor workers (Kernan Freire et al., TU Delft, CHI 2023) confirms the pattern: this knowledge is hard to put into words and hard to capture in procedures and documents, and is mainly built through experience and practice. By definition it is not in the systems where you would look for it.

What knowledge loss costs

Precise amounts circulate widely, but are rarely traceable to a verifiable source. It is more useful to look at the mechanisms through which costs arise, because those are recognisable in every production company and can be calculated internally.

• Longer onboarding time. New operators master basic operations within one to three months, but full independence in knowledge-intensive production takes six to twelve months. During that period the new employee runs below level and the experienced colleague is pulled out of production to mentor.
• Quality loss and rework. Mistakes an experienced colleague would never make cost material, machine time, and sometimes customer trust.
• Dependence on a few people. If one person has the answer, that person becomes a bottleneck. Leave, sickness, and retirement then become operational risks rather than HR matters.
• Loss of improvement capacity. Improvement starts with knowing how things really work. Without captured practical knowledge, every improvement or automation project starts from zero.
• Repeated problem-solving. The same problem is solved again and again because last time's solution cannot be found anywhere.

Read more about the cost and duration of onboarding new employees on the shop floor.

Whoever fills in these five items with their own figures has a business case within an hour that is sharper than any benchmark number.

Why earlier attempts failed

Most production companies have already tried. Usually more than once.

A knowledge base or wiki asks for writing on top of the real work, and maintenance that nobody owns. Videos are quick to make but not searchable and not easy to update. An external consultant captures knowledge at one moment, after which capture stops once the assignment ends. And keeping the senior worker on longer solves nothing; it moves the problem to a moment when there is less time left.

What these approaches have in common is that they separate capture from the work. The operator must stop, sit behind a screen, and formulate the knowledge in written language. Each of those three steps costs effort at exactly the wrong moment.

Why typing, forms, and wikis fail on the shop floor, and why spoken capture fits, is covered in knowledge capture on the shop floor.

The four building blocks of knowledge retention that holds up

Knowledge retention is not a project with an end date, but a chain. If one link is missing, the rest still leaks away.

• 1. Capture during the work, not afterwards. Knowledge you reconstruct an hour later has already lost half its detail. Speech is the only form here that requires no extra action: someone explains in words what is happening, something that already happens on the shop floor. The difference is that it is captured instead of disappearing.
• 2. Structure without burdening the craftsperson. Ordering is a task for the system. Transcription and processing turn a spoken fragment into a searchable knowledge item, with the context of the work environment included: which machine, which order, which moment. The operator does not need to fill in fields.
• 3. A human control step before something counts as knowledge. This is the link most often skipped. As long as nobody with craft knowledge confirms or corrects what was captured, a collection grows that nobody trusts. Incomplete but accurate beats complete and unreliable.
• 4. Keep it current and reusable. Captured knowledge becomes outdated. Machines are replaced, customers change their requirements, procedures are adjusted. Without a mechanism that signals outdated knowledge, the system becomes as unreliable after a year or two as the wiki that preceded it.

How to keep captured knowledge accurate and reliable is covered in the article on knowledge preservation on the shop floor.

These four steps are also why knowledge retention is not the same as recording. A recording is raw input. Knowledge is what remains after it has been structured, checked, and kept current.

Knowledge retention and AI

Since language models became widely available, the question often comes in one breath: can AI not just solve this?

Partly. A model can turn speech into text, add structure, and answer questions. What a model cannot do is know knowledge that exists nowhere. A language model is trained on publicly available information and knows nothing of the quirks of your machines, your customers, and your exceptions. As long as that knowledge is not captured and confirmed by a craftsperson, the model fills the gaps with plausible but incorrect answers. In a production environment that is more dangerous than no answer at all.

Read more about what AI does not know about your production environment and about the EU AI Act for manufacturers using AI in knowledge management.

Regulation adds to that. The EU AI Act sets requirements for transparency and human oversight in AI systems, and employee speech is also personal data under the GDPR. Traceability to the original source and demonstrable human approval are therefore not a luxury, but a condition for using AI responsibly in knowledge retention.

The practical conclusion: AI is useful as a processing layer on top of a reliable knowledge base from your own environment, not as a replacement for it.

How to get started

Knowledge retention often fails on the scale of the ambition. An organisation-wide rollout is not needed to begin.

• Name the vulnerable points. Which one to three people make things go wrong when they are not there? Which machines or processes do only they know?
• Choose one bounded area. One line, one machine park, or one type of fault. Small enough to see results within weeks.
• Let capture coincide with the work. No separate sessions, no forms. Capture in the moment, in spoken language.
• Assign the control step. Decide who judges whether captured knowledge is accurate before others rely on it.
• Measure on usage, not volume. The question is not how much was captured, but whether someone with a question finds a usable answer within a minute.

What a focused approach to onboarding new employees delivers once knowledge is searchable is covered in the separate article on shop-floor onboarding.

If after a few weeks a new colleague finds an answer without interrupting the experienced colleague, it works. If not, the fault is usually in step three.

If someone on your team will retire in a few years, also read the article on knowledge transfer before retirement: capturing knowledge spread over years produces a more complete picture than an intensive programme just before someone leaves.