I see quite a few retirement posts on LinkedIn. After thirty or forty years with a company, someone signs off for the last time. They thank the people they have worked with, mention the friendships, reflect on a few things they achieved. They are always good posts to read.
Lately I have found myself thinking about something else while reading them: What happens to everything that person knows?
Not the obvious things. Current projects, outstanding work, key contacts, whatever their replacement needs in order to pick up on Monday morning – all of that gets covered in a handover, and handovers are generally done competently.
I mean the other material. The note somebody made against a policy document fifteen years ago because there was something unusual about it. The email explaining why a particular decision was taken. The reason they know to look in one place rather than another when something does not quite add up. The dozens of small things learned over three decades that were never written into any formal process, because nobody thought they needed to be.
This is not a marginal concern. The proportion of the UK population aged 16 and over who are 50 or older was projected to reach 46.4% in 2025, up from 44.6% a decade earlier, and to keep climbing for the next twenty years [1]. Whatever happens when a long-serving person leaves, it is going to keep happening, at scale, for some time.
Two different losses, with two different answers
It is worth separating what actually walks out of the door, because the usual framing treats it as one thing when it is really two, and only one of them is genuinely unrecoverable.
Some of it only ever existed in that person’s head. Judgement built from repetition. A sense of when a set of numbers feels wrong before anyone can say why. The instinct that a particular counterparty will be difficult in a particular way. That is tacit knowledge, and it is real. You cannot interview thirty years of experience out of somebody in a fortnight, and pretending a handover document will capture it is the kind of thing everyone agrees to and nobody believes.
But a great deal of what that person knows was written down at the time – just not by them, and not anywhere designed to be found again.
Consider what the long-serving employee actually knows. They know a particular client relationship nearly collapsed in 2011 and why it recovered. They know which supplier contract contains an unusual indemnity clause, because they were in the room when it was negotiated. They know a policy everyone now follows exists because of a single incident fifteen years ago that went badly.
Almost all of that was documented. The correspondence exists. The contract exists. The incident report, the claim file, the board paper, the email chain in which somebody argued for a different approach – these were written down, because organisations of any size document what they do.
What that person carries is not the content. It is the index. They know the thing happened, roughly when, and roughly where to look. Without them the records are still there, but nobody knows they are relevant, or that they exist at all.
That is a materially different problem from knowledge loss, and it has a materially different solution. The tacit half you largely have to accept. The indexed half is tractable.
Current systems record what is true, records record why
The reason this matters more now than it used to is that organisations tend to assume their current systems already contain their organisational knowledge. They contain something, certainly, but it is worth being precise about what.
A modern operational system records state. It tells you what the current policy is, which customers you have, what the contract terms are, what the process requires. It is an accurate picture of the organisation as it stands today.
What it does not generally record is how the organisation arrived at that state. The rule is there; the incident that produced the rule is not. The contract terms are there; the negotiation that explains why they are unusual is not. The approved supplier list is there; the supplier removed from it in 2014, and the reason, is not.
Institutional knowledge is not really a collection of facts. It is decisions, the reasoning behind them, and what happened as a result. Current systems reliably hold the first of those three. The second and third, where they exist at all, sit in the historical estate – in correspondence, case files, claims, minutes, reports and the accumulated paperwork of things going right and wrong over decades.
Having records and having knowledge are not the same thing. You can keep every policy, every claim, every email and every document and still lose the understanding that sits behind them.
Organisations learn most from what went wrong
One category deserves singling out, because it is both the most valuable and the most neglected.
Organisations learn far more from failure than from success. The claim that was mishandled, the project that overran, the contract that produced consequences nobody anticipated, the complaint that exposed a gap in a process – these generate the sharpest organisational learning, and unusually rich documentation, because things that go wrong get investigated and written up.
They also tend to be exactly the records that get archived quickly, classified defensively and kept at arm’s length. Sensitive, closed, retained because policy says so, and no longer part of any live system.
So an organisation’s most useful accumulated experience is often the part furthest from anything a modern system could reach.
The coping mechanism does not transfer
Perhaps we have never worried much about this, because organisations have always found a way to cope. Someone else remembers part of it. A colleague knows roughly what happened. People search through old emails. Eventually enough of the story gets pieced together to carry on.
That workaround has a hidden dependency: it requires a human being doing the reconstructing, someone who can ask around, interpret a half-remembered account and recognise the relevant document when they finally open it. It is inefficient – APQC’s survey of 982 knowledge workers found that recreating information which already exists is among the standard weekly productivity drains, alongside the 2.8 hours a week spent looking for or requesting information [2] – but it works well enough that the underlying problem never has to be solved.
We are now asking technology to do something quite different. The premise of applying AI to organisational information is that it can work across decades of material, find patterns, answer questions and help us understand what we already know.
It can only work with what was actually captured, and only with what it can reach. It cannot ask the person who retired three years ago why that decision was made.
Which produces a specific problem. An AI system given access only to current systems inherits the organisation’s conclusions without its reasoning. It can tell you what the policy says. It cannot tell you why the policy exists – which means it cannot tell you when the circumstances that produced the policy no longer apply, or when an apparent exception is a repeat of something that has happened before.
That is not a hypothetical concern about model accuracy. It is a straightforward consequence of the information available. For many of the questions people most want to ask – should we accept this risk, have we seen this before, why do we do it this way – the answer depends on precisely the material that has been archived out of reach.
Making the archive readable
None of this argues for indiscriminately exposing decades of documents to an AI platform. Much historical information is sensitive, superseded, or genuinely should not be used, and a great deal of it has no continuing value at all.
It argues for a different question than the one most organisations are asking. Not “how long must we keep this?” but “which parts of this represent knowledge the organisation would want to keep using?”
Answering that means knowing what exists and in what condition, deciding which categories carry real institutional value, and then doing the unglamorous work: making records searchable, capturing the context and relationships that give them meaning, establishing provenance, and applying governance appropriate to material created for a different purpose in a different era.
That last point matters more than it sounds. The UK government’s guidance on preparing datasets for AI records The National Archives’ concern about digitising historical records without metadata, on the grounds that doing so creates governance and risk gaps [3]. Converting an archive to digital images does not make it findable. It makes it digital.
So when I next see one of those retirement posts, I will still congratulate them on a good career. But I will also wonder how much of what that person knew is written down somewhere, and how much simply left the building with them.
Some of it did leave, and that part is gone. The rest has been sitting in storage for decades, paid for and preserved, waiting for something other than a long-serving employee to be able to find it. Those employees are retiring.
Which brings me back to the question I keep asking. Can you really have an AI strategy without a historical data strategy?
How Dajon helps
Most of our work sits in the recoverable half of this problem. We help organisations establish what their historical estate actually contains, then make it findable: digitising with metadata rather than without it, capturing the context and relationships that give records meaning, and applying classification and access controls suited to material created decades ago. If you want to know what your archive would be worth to an AI system, that is a question we can answer concretely.
References