# Knowledge that evaporates: what your company stops learning when everyone solves problems alone with AI > Every problem solved alone with AI and never recorded is a lesson that will never exist. What your company stops learning, and the habit that fixes it. Article · By Élie Sader, Network Effect · Published on November 4, 2026 Someone at your company hits a problem this morning. A quote that doesn't add up, a contract clause they don't understand, a machine showing an unfamiliar fault code. They ask an artificial intelligence (AI), get an answer in ten minutes, and move on. The problem is gone. Yesterday, solving it would also have left a trace: a question to the colleague at the next desk, a note in the file, a trade-off explained out loud in a meeting. Today, that trace was never born. Multiply that by every move in a working day, and the company's memory simply stops being built. Nothing breaks, and the dashboards look better than ever. The gains show up immediately. The losses only show up on the day you need what was never kept. > The danger is not that knowledge gets destroyed. It is that it never gets created. ## Where did what we just learned go? Know-how lives on two things: practice and transmission. As long as someone writes down what they understood, formalises it, explains it to someone else, knowledge circulates and grows stronger. That is what we mean by making it explicit: putting into words what you know how to do, so that someone else can pick it up without you. Making things explicit has always had a cost, and that cost carried a hidden benefit: when you explain, you often discover what you didn't know you didn't know. AI does not attack that knowledge. It doesn't erase it or corrupt it. It short-circuits the effort that used to bring it into existence. When the answer arrives on its own, there is no reason left to ask a colleague, no reason to write the note, no reason to explain why you decided the way you did. This is not destruction. It is non-creation: the internal note, the transferable lesson, the documented trade-off are simply never produced. Strictly speaking, nothing was lost. Something just never came to be. And nobody did anything wrong. The person who solved it quickly did good work. It is the organisation that never asked them for anything more. ## Why don't we see it coming? Because the two sides of the ledger don't become visible at the same time. The gains from AI are immediate and measurable: less time for the same work, a document drafted, a decision made. The loss of knowledge only shows when you need it: a new challenge that calls for the depth you used to have, a key person leaving, a new hire who takes twice as long as expected to become useful, because there is nothing for them to read. That asymmetry creates a bias, and the bias is structural. You over-optimise for fast execution, because it is visible, and under-invest in transmission, because it isn't. It is a strategic mistake dressed up as a productivity gain. In the short run, everything speeds up. In the medium run, a quiet fragility sets in: innovation slows, onboarding takes longer, and the company leans ever harder on the few people who still know. None of it appears on the balance sheet or on any dashboard. The organisation learns at the speed of AI, and forgets at the speed of staff turnover. ## Has anyone measured this? Yes, at the scale of public knowledge. In 2026, three MIT economists, Daron Acemoglu, Dingwen Kong and Asuman Ozdaglar, modelled the mechanism in a National Bureau of Economic Research working paper. Their conclusion is short: an AI that gives good, personalised advice improves today's decisions, and at the same time weakens the incentive to learn and to share. Past a certain level of recommendation quality, the system can tip into a state where shared knowledge disappears, even though every individual piece of advice is excellent. The one remedy they consider unambiguous: a better capacity to aggregate human knowledge, that is, to collect it and bring it together. The signs are already here. On Stack Overflow, the large forum where developers have asked each other questions since 2008, the number of new questions has fallen back to its 2009 level, with a sharp drop from late 2022. In October 2025, the Wikimedia Foundation reported that Wikipedia had lost around 8% of its human page views in a year, a decline it attributes to AI-generated search summaries and to social video. Every question no longer asked in public is an answer the next person won't find. What holds for public knowledge holds for a company's knowledge too, on a smaller scale and faster. Its internal commons, what each person learns and everyone else can reuse, drains in exactly the same way. ## Is my company more exposed than a large one? Yes, because its own knowledge is precisely what makes it irreplaceable. The craft, the region, the customers, the local constraints, the history of each relationship: that specific knowledge is what a generic competitor doesn't have. If it erodes, the company becomes just another competitor, and the bigger players have better models. A small or mid-sized company doesn't win on computing power. It wins on what it knows that nobody else does. The erosion hits first where accumulated knowledge is the skill itself: design, advisory work, estimating, decision-making. That is exactly where AI is most useful, and therefore where it most quickly wears away whatever was never formalised. AI dissolves the premium on rare, codified knowledge, the kind that lived in books and procedures. It does not replace the knowledge that comes from repeated contact with the real world. And that is precisely the knowledge that disappears first once people stop writing it down. There are two different losses here. The first is the knowledge you no longer write down: the trace. The second is the knowledge you no longer practise: judgement. The philosopher Michael Polanyi put it in a sentence that stuck: "we know more than we can tell". Part of any skill cannot be put into words; it is kept alive by doing. An estimator who never prices a job by hand any more eventually stops noticing when a margin is absurd. Worse, they don't know they've stopped noticing. Skills aren't lost overnight; they decline gradually. That second loss can't be repaired by writing alone. It requires staying hands-on with whatever you delegate. ## Why does memory keep starting from zero? Because that's how things were organised, long before AI. Look at professional services. A firm can work with a client for ten years and start from scratch on every new engagement, not out of incompetence, but because the knowledge lives in people. The files stay; the memory leaves with the team. This episodic cycle was tolerated for a simple reason: rebuilding was billable, and its real cost was quietly absorbed by the client, whose own staff brought each newcomer up to speed. Nobody ever named it. The same cycle runs inside every company: between departments, between generations, between a workshop manager who retires and the one who takes over. Last year's file is buried in the archive. What mattered was the reasoning, the assumptions, what had been tried and why it was dropped, and none of that was ever written. AI changes two things. It makes it technically possible for memory to accumulate instead of being rebuilt. And it removes the excuse. The question is no longer "who has the best experts?" but "who has the deepest memory, and where does it build up?" Bitcoin offers an old perspective on this. Writing and accounting were not invented for literature. They were invented to keep track of exchanges nobody could watch over. Mesopotamian merchants sealed clay tokens inside envelopes, then inscribed tablets with the detail on the front and the total on the back, so that a mistake would show. Bitcoin is the latest link in that chain: a ledger where every block carries the fingerprint of the one before, where nothing is erased, where the history is public and anyone can check it. The lesson for a company is simple. A system's memory is not an extra: it is the condition of its existence. What the ledger doesn't keep never happened. Thousands of machines that don't trust each other have held together for more than fifteen years because the network never rebuilds its memory. It accumulates it, block after block. A company that rebuilds its own on every engagement deprives itself of exactly what makes the ledger strong. ## So what do we actually do? First, recognise that this is not a technology decision. It is an organisational design decision. The erosion is not inevitable; it is what happens in the absence of a rule. And the remedy is not a big programme, it is a habit: give every use of AI a dual mandate. Solve the problem, and record what was learned. In one place for everyone, in a form that both a person and an AI can read. The objection comes straight away, and it is a fair one. Documentation takes time. It goes out of date. Nobody reads it. Anyone who has ever launched a big procedures project knows how it ends: a binder that is up to date for six months, then forgotten. If the remedy were "document everything", it would fail like all the others. But three things have changed. First, what you record today won't only be read by busy people: it will be read by tomorrow's AI, which will pick it up, connect it and bring it back the moment someone needs it. Second, AI itself helps with the recording: the trace can be produced in the same move as the solution, a few lines the person reads and corrects, rather than a Friday-afternoon chore. Third, the commons we're talking about is internal. The argument "I'd be feeding my competitors", which holds back sharing in public, doesn't apply: what you record stays with you. What remains is the cost of not doing it, the invisible one described above. What the company gets is a memory that builds up. One you can question: what have we already tried with this customer, and why? One you can pass on: a newcomer reads the reasoning, not just the conclusions. And one that lasts: it survives departures because it no longer lives in a single head. That accumulating knowledge is the company's intangible capital itself: what it knows, its weak signals, its dealings with the real world, carefully recorded and worth protecting. In [Intangible capital](https://networkeffect.fr/en/intangible-capital/) we described what you lose when it drains into platform agents. It can also be lost without going anywhere at all, simply because it stops being written down. The corpus is where the company's memory is built up instead of rebuilt. Two precautions to finish. Recording is no excuse for not practising: keep doing a few things by hand, now and then, on whatever you delegate, so judgement stays alive. And keep that memory in-house, in files you own. Memory accumulated at a supplier, any supplier, is one more dependency. ## What your company learns today AI speeds up execution, and quietly slows down transmission. That isn't fate; it is the absence of a habit. Choosing the dual mandate means deciding that what your company learns today will still belong to it ten years from now. That is the work of [CORPUS](https://networkeffect.fr/en/services/corpus/): recording what has never been written down, then connecting the corpus to the company's everyday flows so it stays alive rather than becoming a snapshot. And if the next question is who will be able to read that memory, it is covered in [the corpus trilemma](https://networkeffect.fr/en/corpus-trilemma/). ## Further reading - [Keep your company's brain](https://networkeffect.fr/en/services/corpus/) - [Intangible capital: what makes your company singular, and how to keep it in-house](https://networkeffect.fr/en/intangible-capital/) - [All resources](https://networkeffect.fr/en/resources/)