What Organisations Teach Their Machines

When an organisation connects artificial intelligence to its internal knowledge, it is tempting to imagine that the machine has suddenly gained access to institutional memory.

But organisations rarely possess a single coherent memory. They possess fragments: documents, tickets, policies, decisions, meeting notes, architecture diagrams, old project plans, informal expertise and abandoned systems whose relationships are often poorly understood.

A decision may be recorded in one place, while the reason for it sits somewhere else. The final architecture may be documented, but not the alternatives that were rejected. A workaround may exist in a service ticket long after the person who understood why it was needed has left. A policy may still be searchable years after the operating reality has moved on.

This matters because retrieval does not repair those gaps. It makes whatever has survived easier to find.

A machine does not inherit what an organisation knew. It inherits what the organisation managed to leave behind.

Retrieval inherits the archive.

Retrieval-augmented generation, or RAG, was developed partly to overcome a limitation of language models: their knowledge is constrained by what was encoded during training. Instead of asking the model to rely entirely on that internal knowledge, a retrieval system searches an external corpus and supplies relevant material as context for the answer. The original RAG work by Lewis and colleagues demonstrated that combining retrieval with generation could improve performance on knowledge-intensive tasks.[1]

In an enterprise, the attraction is obvious. The model does not need to contain everything the organisation knows. It can search the organisation's own information when someone asks a question.

But that changes the problem rather than removing it.

Suppose someone asks an internal assistant why a particular architecture was chosen. The approved design may explain what was implemented, but perhaps the real reasoning is spread across an options paper, an incident review, two meetings and an email from an engineer who left three years ago. The organisation may still possess all of those artefacts, but that does not mean the retrieval system understands how they relate to one another.

More recent work on enterprise retrieval reflects this difficulty. Microsoft Research's AgenticRAG preprint, for example, explores moving beyond a single retrieval step by allowing a reasoning model to search repeatedly, navigate within documents and refine what evidence it needs.[2] The interesting point is not that one retrieval architecture has solved enterprise knowledge. It is almost the opposite: finding the right evidence in a large organisational corpus is difficult enough that retrieval itself is becoming a more sophisticated reasoning problem.

And even an excellent retrieval system reaches a hard boundary.

It cannot retrieve what was never recorded.

Organisations had this problem long before AI

Institutional memory is not a new concern created by large language models. Management researchers were examining it decades before today's AI systems existed.

Walsh and Ungson's 1991 work on organisational memory described organisations as retaining information through multiple repositories rather than through one central memory.[3] A few years later, Sharp and Lewis used the term corporate memory to describe not only formal information but also the informal knowledge accumulated through employees' experience and position within an organisation.[4]

That distinction still feels remarkably current.

Anyone who has spent enough time inside a large organisation knows that there is the documented process, and then there is the way the process actually works. There is the service documentation, and then there is the engineer who remembers the strange dependency nobody included in the diagram. There is the decision log, and then there is somebody who remembers that the obvious option was tried before and knows why it failed.

We call some of those people experienced. What they often possess is something more specific: a map of relationships, history, exceptions and consequences that the organisation itself has never completely written down.

When they leave, the organisation does not necessarily lose all of the information they held. It loses the connections between pieces of information.

That may be the more important loss.

Search is not the same as memory

This is where I think some of the language around enterprise AI becomes misleading. We talk about giving machines access to organisational knowledge as though access and understanding were roughly the same thing.

They are not.

Imagine a system retrieving five documents about the same process. One is current policy. One was superseded eighteen months ago. One describes an exception. One came from a failed pilot. One is a working document that was never formally approved.

All five are real organisational records. All five may be semantically relevant. But they do not have the same status.

A person who has worked in the organisation for years may recognise those distinctions immediately. They know which team is authoritative, which system replaced another, which document is historical and which decision was temporary. A machine needs those relationships represented somehow: through metadata, provenance, chronology, ownership, retrieval logic or other forms of context.

Otherwise, the organisation risks confusing what can be found with what should be believed.

There is a subtler consequence here too. Once AI becomes an interface to organisational knowledge, the surviving archive starts to shape what the organisation can easily remember.

Formal decisions are more likely to survive than informal disagreement. Approved documents are easier to retrieve than the discussion that produced them. Successful projects often leave cleaner records than abandoned ones. Current structures are easier to represent than the organisational history that explains why they exist.

The result does not have to be false to be misleading. It can simply be incomplete in a consistent direction.

An organisation may gradually become better at retrieving its conclusions than remembering the arguments, uncertainty and failed alternatives that produced them.

That matters because the discarded alternatives are often where the learning is.

What gets preserved acquires power

There is a tendency to think of missing organisational knowledge as an inconvenience. Someone cannot find a document, so they ask around. A new employee needs longer to understand why something works the way it does. A project repeats an investigation that somebody completed three years earlier.

With AI, the consequences become more interesting because machines can turn retrieved material into answers at scale.

If an outdated document is difficult for a human to find, perhaps only a few people ever encounter it. If that same document is indexed, retrieved and repeatedly summarised by an assistant, its influence can become much larger.

The archive is no longer passive.

This makes questions that once sounded administrative surprisingly important. Who owns this document? Is it still valid? What superseded it? Was this an approved decision or simply somebody's recommendation? What evidence supported it? Which exception applies? When should the information stop being treated as authoritative?

These are not glamorous AI questions. They are nevertheless part of the AI system.

The quality of the knowledge environment becomes part of the quality of the machine's output.

Permission also shapes memory

Enterprise knowledge has another property that public information does not: not everybody is supposed to see everything.

Security investigations, employee information, legal advice, commercial negotiations, customer data and sensitive architecture may all legitimately sit behind different access boundaries. Connecting AI to enterprise information therefore creates a second problem alongside retrieval: the system needs to find useful evidence without weakening the controls surrounding it.

Recent Microsoft research has demonstrated how weaknesses in access-control enforcement around fine-tuning and RAG systems can create opportunities for sensitive information to leak to unauthorised users. The authors argue for deterministic, fine-grained access control across enterprise AI workflows.[5]

There is an organisational consequence hiding inside that security problem.

Two people can ask the same machine the same question and, correctly, receive answers based on different evidence because they are authorised to see different things.

Institutional memory is therefore not only fragmented by where information lives. It is also fragmented by who is allowed to retrieve it.

That is probably unavoidable. But organisations should understand what it means. There may never be one neutral AI representation of "what the organisation knows". There may instead be overlapping views shaped by role, access, provenance and organisational boundaries.

The machine learns our habits as well as our knowledge

This leads to the part I find most important.

What exactly are organisations teaching their machines?

Not intentionally. Operationally.

If an organisation records decisions but rarely records why they were made, the machine inherits conclusions without reasoning. If documents are created but rarely retired, it inherits contradictions. If incident reviews capture remediation but not learning, it inherits fixes without context. If exceptions accumulate quietly, it inherits policy without reality.

And if the knowledge people actually rely on lives primarily in private conversations, individual inboxes and the memories of experienced staff, the machine inherits whatever happened to escape into formal systems.

That makes enterprise AI a mirror of sorts. Not because it perfectly reflects the organisation, but because its failures can reveal what the organisation itself has failed to make legible.

A poor answer to "Why did we choose this?" may expose weak decision records. Conflicting answers may expose unresolved document ownership. An answer based on an obsolete process may expose poor information lifecycle management. An inability to explain an exception may reveal knowledge that was never institutionalised in the first place.

These look like AI problems when they appear through an AI interface. Often they began much earlier.

This is more than data quality

The immediate response may be to treat this as a data-quality exercise: clean the documents, improve metadata, remove duplicates, fix permissions and build a better index.

All of that matters. But I do not think it is enough.

An organisation is not simply a collection of information. It is also a history of decisions.

Why was something chosen? What alternatives were considered? What assumptions were true at the time? What failed before this approach succeeded? Which decision replaced which? What evidence supported the choice? Who owns it now? Under what conditions should it be reconsidered?

A document tells you something happened. Institutional memory should help you understand why.

That distinction becomes more important as machines move from finding information to interpreting it, recommending actions and eventually acting on the organisation's behalf.

If we want machines to reason from organisational knowledge, then the reasoning that produced important organisational decisions needs to become more legible too.

So what should organisations do differently?

I do not think the answer is to document everything. That would probably create a larger archive without necessarily creating better memory.

The more useful shift is to treat important organisational knowledge as infrastructure rather than administrative residue.

Decisions worth preserving should carry some account of their rationale, assumptions and alternatives. Important artefacts need ownership and lifecycle, so that systems can distinguish current authority from historical record. Provenance needs to travel with information: where it came from, who approved it, when it became effective and what replaced it.

Organisations also need to become more comfortable preserving uncertainty. Sometimes there is no single settled answer. Teams disagree. Evidence changes. A workaround remains under investigation. A useful AI system should be able to surface those conditions instead of compressing them into false certainty.

And organisations deploying internal AI should test more than whether the system can answer obvious factual questions. Ask questions that institutional memory finds difficult: Why did we choose this? What did we try before? Which assumption changed? What superseded this decision? Where is the exception? Who owns it now?

Then compare the machine's answer with what experienced people in the organisation actually know.

The gap may be more useful than the answer. Because that gap tells you something about the organisation itself.

It shows what has been preserved, what has become disconnected and what was never made institutional in the first place.

What we leave behind

There is understandable excitement about giving AI access to organisational knowledge. Done well, the value could be substantial. People could spend less time rediscovering information, navigating fragmented repositories or depending on a small number of colleagues simply because those colleagues remember where everything is buried.

But access to more information should not be confused with institutional memory.

Memory has structure. It has chronology. It distinguishes what happened from why it happened. It preserves relationships between decisions and consequences. It also forgets, revises and recognises when an old answer should no longer govern a new situation.

AI does not create those properties merely by being connected to a document store.

It inherits whatever version of them the organisation has already built.

Perhaps that is the more useful way to think about enterprise AI. Before asking how much organisational knowledge we can place within reach of a machine, we should ask what kind of organisation that knowledge describes.

Does it preserve only decisions, or the reasoning behind them? Does it know what is current and what is history? Can it recognise uncertainty? Does it retain the lessons of failure as carefully as the records of success?

Because the machine will learn from what remains.

And what remains is not necessarily everything the organisation once knew.

Sources and further reading

[1] Lewis, P. et al. (2020) “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” Available at: https://doi.org/10.48550/arxiv.2005.11401.
The foundational RAG paper, introducing generation grounded in retrieved non-parametric memory.

[2] Suresh, S. et al. (2026) “AgenticRAG: Agentic Retrieval for Enterprise Knowledge Bases,” arXiv.org [Preprint].
A Microsoft Research preprint examining iterative retrieval, document navigation and analysis over enterprise knowledge bases. It should be treated as a preprint rather than peer-reviewed literature.

[3] Walsh, J.P. and Ungson, G.R. (1991) “ORGANIZATIONAL MEMORY,” The Academy of Management review, 16(1), pp. 57–91. Available at: https://doi.org/10.5465/amr.1991.4278992.
A foundational treatment of how organisations acquire, retain and retrieve information, and how organisational memory is distributed across different repositories.

[4] Sharp, C. and Lewis, N. (1993) “Information Systems and Corporate Memory: design for staff turn-over,” AJIS. Australasian journal of information systems, 1(1). Available at: https://doi.org/10.3127/ajis.v1i1.434.
Particularly relevant to the informal or “soft” knowledge accumulated through employees’ experience and the risk of losing it through organisational change and staff turnover.

[5] Bhatt, S.S. et al. (2025) “Enterprise AI Must Enforce Participant-Aware Access Control,” arXiv.org [Preprint]. arXiv:2509.14608.
A Microsoft Research preprint examining information leakage and fine-grained access control in enterprise LLM, fine-tuning and RAG environments. It is cited here specifically for its access-control argument, not as evidence about organisational memory itself.