Companies Could Know More and Understand Less

Companies have spent decades trying to solve a persistent problem: they know more than they can remember. Important decisions disappear into email threads. Experienced employees leave and take years of tacit knowledge with them. The reasoning behind old projects becomes difficult to reconstruct. Lessons learned in one part of an organisation fail to reach another, while documents accumulate in repositories that are technically searchable but practically forgotten.

Generative artificial intelligence appears to offer an answer. An organisation can increasingly make its collective history accessible through natural language. Meetings can be transcribed, correspondence summarised, decisions indexed and large internal archives queried without knowing where the relevant document is stored. Knowledge that once depended on finding the right person can potentially be retrieved from the organisation itself.

This could represent one of the most valuable applications of artificial intelligence inside companies. It could also create a less obvious problem. A company can become much better at remembering what it knows without necessarily becoming better at understanding it.

Understanding here means the ability to explain the assumptions behind a decision, recognise when evidence challenges it and adapt the response. It is not a count of facts employees can recall unaided. An organisation includes people, tools and routines, so reduced individual recall can coexist with better collective performance. Research on organisational learning treats these elements as interdependent rather than assuming that knowledge stored outside a person is necessarily lost to the organisation.^[1]

Consider how institutional knowledge might be acquired. A junior employee asks an experienced colleague why a client behaves in a particular way. The explanation may begin with a simple answer but expand into the history of the relationship, previous failures, personalities involved and exceptions that never made their way into formal documentation. Through conversation, disagreement and repetition, the junior gradually constructs a mental model of the situation.

That process can be inefficient, but the inefficiency is not its only characteristic. A generative system could offer an immediate explanation assembled from years of emails, documents and previous decisions. In many cases the answer might be more comprehensive than the colleague could produce from memory. The employee would nevertheless have obtained the information through a different process, without necessarily participating in the relationships through which the organisation interpreted it.

Batia Wiesenfeld and Katherine Kellogg examine generative organisational learning through 19 months of observations, interviews and collaborative research at one health system. Their account describes new ways of searching, creating and transferring knowledge, alongside potential risks to social networks and transactive memory, the understanding of who knows what. It supports an analysis of opportunities and mechanisms, not a broad finding that AI knowledge systems make companies understand less.^[1]

An organisation with a well-maintained machine-readable archive could retrieve more of its recorded experience: the documented history of a customer relationship, discussions around an acquisition or reasons recorded when a supplier was replaced. Missing exchanges, tacit context and uncertain interpretation would still limit what could be reconstructed. Easier access could deepen employees' understanding, or reduce their engagement with the underlying history; the archive's size alone does not settle which occurs.

There is nothing inherently wrong with placing memory outside the individual. An employee need not memorise a supplier database to make good purchasing decisions, and searching a reliable archive can be more useful than relying on an imperfect recollection. The organisational consequence becomes more complicated when the thing being delegated is not recall but interpretation.

Knowing that a company rejected a particular strategy five years ago is useful. Understanding why the strategy failed, which assumptions proved wrong, whether those assumptions still hold and how much confidence should be placed in the original conclusion requires something more than access to the record. An AI system may reconstruct much of that context, particularly if the archive is rich. But the more employees rely on it to perform the reconstruction, the fewer opportunities they may have to develop their own integrated understanding.

The risk is not that information has moved outside individual memory. It is that access to a persuasive explanation is treated as equivalent to having a reliable way to challenge it. A company could retain more records while becoming less able to identify an obsolete assumption, but that outcome would have to be established through its decisions and adaptability rather than inferred from reduced recall.

The potential benefits should not be understated. A system that makes specialist knowledge accessible across departments can reduce dependence on the availability of a particular colleague. It may help a new employee explore unfamiliar work, expose contradictions in an archive and explain concepts more patiently than a busy manager. Extracting knowledge from individuals and developing understanding within individuals can reinforce one another. They simply should not be assumed to be the same objective.

A 2025 study by Hao-Ping Lee and colleagues surveyed 319 knowledge workers about their use of generative AI. Higher confidence in the technology was associated with less self-reported critical thinking, while greater confidence in their own ability was associated with more. The study concerns reported behaviour and effort, not an experimentally established decline in intelligence or long-term expertise. It nevertheless offers a reason to distinguish easier access to an answer from active engagement with it.^[2]

At the organisational level, the question is whether easier access to answers changes the capacity to evaluate them. A system that explains the evidence and exposes disagreements could strengthen that capacity. A process that rewards accepting an answer without examining its limits could weaken it. Neither the individual survey associations nor an increase in stored information establishes the aggregate result.

The distinction matters most when the world changes. Institutional knowledge is valuable partly because experience can be reused. A company learns that a particular supplier is unreliable, that a certain type of project tends to fail or that customers respond badly to a pricing strategy. Capturing those lessons can prevent it from paying repeatedly for the same mistakes. Yet an old conclusion can become dangerous when the circumstances that justified it have disappeared.

A sophisticated knowledge system could make precedent easier to retrieve than ever before. It could find the previous decision instantly, summarise the supporting evidence and present the conclusion fluently. That convenience might encourage reuse without enough examination of whether the evidence is still relevant. Conversely, a well-designed system could draw attention to changed conditions and help employees challenge precisely the precedents that a human organisation might otherwise preserve unquestioningly.

A related risk arises when generated interpretations enter the archive itself. An AI-generated proposition might be copied into a report, repeated in subsequent briefings and eventually appear to be supported by several documents. Those documents would not constitute independent confirmation if they all inherited the same original assumption. The problem would be institutional as much as technical: repetition could acquire the appearance of evidence.

The boundary between retrieval and interpretation therefore becomes important. A conventional document records what somebody wrote at a particular time, leaving the reader to assess its relevance. A generative knowledge system can synthesise several documents, infer relationships and resolve apparent contradictions. Its answer is not simply a retrieved record but an interpretation of records, and should be evaluated accordingly.

As these systems become more deeply embedded in organisations, they may increasingly mediate the relationship between employees and institutional history. Workers might read fewer original exchanges or speak less often to the people involved, relying instead on a synthesis that explains what the material means. This could make companies dramatically more efficient at learning from themselves. It could also make the quality and transparency of that synthesis unusually consequential.

The effect on management could be particularly significant. Imagine a senior executive receiving a concise briefing drawn from fifty operational problems. That executive would have more informational coverage than one who received no briefing at all. It would not follow, however, that the summary conveyed everything that direct engagement with several of those problems might teach. Compression can preserve facts while reducing exposure to the ambiguity and disagreement through which judgement develops.

The relevant question is not whether summaries should exist, but what the organisation expects them to replace. A briefing that prepares a manager to ask better questions serves a different function from one that removes the perceived need to ask questions. The first can deepen understanding; the second can conceal how little understanding has actually been acquired.

This does not imply that companies should preserve inefficient knowledge practices to keep employees intellectually busy. Requiring people to search poorly organised archives or repeatedly rediscover information is not a virtue. Nor is interpersonal transfer automatically superior. Human organisations forget, distort history and allow expertise to disappear. There is substantial room for AI to improve how institutional experience is retained and used.

The challenge is to recognise that a capable memory system changes what people need to know for themselves. Some information can safely remain external if it can be retrieved and checked reliably. Other forms of understanding require deliberate preservation because the company depends on someone recognising when an apparently authoritative answer is incomplete or no longer applicable.

Training may therefore need to put greater emphasis on causal understanding and the limits of precedent rather than on recall alone. Instead of asking whether an employee can retrieve the previous decision, a manager could ask what would have to change for that decision to become wrong. Instead of requiring a summary of the archive, an exercise could require the employee to identify which conclusion has the weakest evidential basis.

This is compatible with the finding that generative AI can change rather than simply remove critical thinking. Lee and colleagues describe a shift towards verifying information, integrating responses and taking responsibility for the task. The practical implication is that organisations should cultivate these activities rather than assuming that familiarity with an AI interface supplies them automatically.^[2]

Companies may also need to preserve the provenance of organisational knowledge more carefully. A conclusion derived from a signed technical report should not automatically carry the same weight as one inferred from informal discussion. A decision validated across repeated outcomes deserves a different status from a proposition generated once and subsequently copied into several documents. The archive should make those differences visible.

Without such distinctions, machine-readable memory could create an illusion of certainty. The system might know that something has been said many times without establishing whether the repetition represents independent evidence. A fluent account of the company's history could then become harder to challenge precisely because it appears more complete than any individual's recollection.

The organisations that manage this well could retain more experience and make expertise more accessible while improving their ability to question the record. The relevant outcome would be better explanations, earlier recognition of changed conditions and more reliable revision of decisions. Those capabilities may be distributed across people and tools; they should not be confused with either unaided memory or the apparent completeness of an archive.

Artificial intelligence may help solve the problem companies have spent decades trying to solve: retaining and retrieving far more of their experience. The harder question is whether the people relying on that memory will still understand enough to recognise when it is leading them astray.