If you use tools like ChatGPT, Claude, or Gemini regularly, you have probably felt something strange. On the one hand, these systems seem to know almost everything about the world. They write convincing emails, summarize research, and answer detailed questions about obscure topics. On the other hand, when you try to use them for serious work inside your company, they forget who you are almost as soon as the window closes.
As individuals, we already rely on these tools as a kind of extended memory. They remember snippets of previous chats, they surface old code examples, and they help us pick up a thought we left weeks earlier. Even with simple chat history, they can reconstruct enough context to feel surprisingly personal. When you ask a follow up question four weeks later, they often still “know” that you were working on a proposal or debugging a specific piece of code. Inside the company, the experience is very different. Every time you explain something important about how your product works, you start from zero. You paste in background documents, provide example tickets, and walk the system through your edge cases. It takes several iterations before the model begins to respond in a way that reflects how your team actually does things. A week later, you ask a related question, and it acts as if that entire explanation never happened.
Now multiply that by hundreds or thousands of employees. Each person is separately trying to teach their agent how the company works. Sales explains pricing exceptions. Support explains escalation paths. Legal explains approved language. Product explains the rationale behind roadmap decisions. The same institutional knowledge is re explained over and over again, in slightly different words, to systems that have no shared memory. At some point, a reasonable question arises: if it is so easy to give an LLM a memory for individual use, why is it so difficult to build a shared memory for an enterprise?
Why Individual Memory Is Easy And Enterprise Memory Is Hard
There are five structural reasons why this gap exists. None of them are technical curiosities. Together, they explain why employees keep searching for answers AI should already know.
In Companies, There Is No Single Source Of Truth
For your personal use, there is an implicit source of truth: your own preferences, your own working style, your own files. The system only has to align with one person. Organizations are different. There is rarely one definitive answer that everyone agrees on. Many “facts” inside a company are in fact negotiated compromises. Pricing, for example, is not a static table. It is shaped by sales strategy, finance constraints, and what customers will accept. Policies evolve through edge cases. Product decisions reflect trade offs that different teams would describe differently. Research from APQC shows that knowledge workers spend several hours every week not just searching for information, but also reconciling conflicting versions of it, recreating work, and tracking down the one colleague who “really knows” how things are done now. This is a symptom of consensus driven truth. Any system that tries to represent “how the company works” has to aggregate many partial, sometimes conflicting perspectives and then expose them in a way that does not pretend there is more certainty than actually exists[1][2].
Even With RAG, Most Systems Are Static Snapshots
To make LLMs more useful in enterprises, many teams add retrieval augmented generation. They index a set of documents, connect that index to the model, and let the AI pull in relevant passages at query time. In theory, this grounds answers in company knowledge and reduces hallucinations. In practice, most of these indexes are built as one time or periodic jobs. They represent a snapshot of the organization at a given moment. Real companies move faster than that. Product packaging changes. Contract templates are updated. New markets are entered. Exceptions become the norm long before anyone updates the official documentation.Stanford’s research on legal LLMs highlights how even domain specific, RAG based tools still hallucinate in roughly one out of six benchmark queries, often because retrieval fails to bring in the right authorities or because the underlying corpus is incomplete. If that happens in a highly structured domain like law, it is even more likely in messy corporate environments where knowledge lives in a mix of formal documents and informal agreements. A static index cannot keep pace with this evolution[3].
Critical Knowledge Does Not Live In One Place
When individuals give an LLM memory, they usually point it at a small number of sources: personal notes, documents, email, or a project space. Enterprises do not have that luxury. Their knowledge is distributed by design. APQC’s survey of nearly one thousand knowledge workers found that the biggest productivity drains come from collaboration and information flow: managing internal communication, looking for or requesting needed information, and participating in meetings that exist solely to transfer knowledge from one person to another. That is because the answers people need are split across chat tools, wikis, ticketing systems, CRMs, data warehouses, and people’s heads[1]. Traditional AI search functions and chatbots can help employees find a document or message faster. They act like smarter indexes over this sprawl. But they do not actually do the work of integrating these fragments into a coherent, up to date understanding of how things work. As a result, the same questions are asked again and again, and different teams continue to make local decisions with partial information.
The Context Window Is Still Too Short
Large language models have made dramatic progress on context length. Some can technically ingest hundreds of pages in a single prompt. But this is still a narrow window compared to the life of an organization. Your company’s knowledge is not a single conversation. It is years of customer interactions, design decisions, policy exceptions, and incidents. Trying to load that into a single prompt is neither practical nor desirable. Even when technically possible, you pay a price in speed, cost, and often answer quality. This is why many enterprise deployments fall back to a pattern where each query is stateless: the model sees a slice of context retrieved from an index, generates an answer, and then forgets. There is no long term memory that accumulates what the organization has already explained. McKinsey’s work on AI adoption shows that while many companies experiment with genAI in isolated use cases, relatively few have invested in the deeper integration and governance needed to turn it into a persistent capability. Without that investment, the context window remains short in practice, no matter what the model specification says [4].
Public Models Are Trained On The Internet, Not On Your Business
Finally, public LLMs are pre trained on vast amounts of public text. That is why they are so impressive at general tasks. But almost none of that training data reflects the specifics of your company. When you ask a base model about your pricing, your escalation path, or your regulatory obligations, it has no idea. It will interpolate from generic patterns it has seen elsewhere. This is precisely why hallucinations are not an edge case; they are a direct consequence of how these models work. Stanford’s studies on legal AI systems and general purpose chatbots underscore this point. Even when connected to curated corpora, models still generate plausible but incorrect statements because their internal world model was never trained on the ground truth of your organization. For personal use, this is often acceptable. If the model misremembers a movie release date, you lose nothing. For enterprises, incorrect statements about pricing, compliance, or customer commitments create real risk. Without a way to systematically replace generic internet knowledge with verified organizational knowledge, corporate AI will continue to sound confident while speaking about the wrong company [3][5][6].
Why Hallucinations Make Your AI Useless For Real Work
All of these constraints would already make enterprise AI challenging. On top of them sits the problem of hallucinations. Stanford’s Human Centered AI institute has shown that even specialized legal research tools marketed as “hallucination free” still generated incorrect or misgrounded answers in more than 17 percent of benchmark queries, roughly one in six. Their earlier work found that general purpose chatbots hallucinated between 58 and 82 percent of the time on legal questions, despite sounding confident [3]. In a business context, employees cannot reliably detect which answers are wrong. Studies described by MIT and others on human AI collaboration point out that people tend to be overconfident in AI generated content and often fail to catch subtle errors [7]. The result is predictable:
Teams either distrust the system and re check everything, which cancels out any productivity gains.
Or they over trust it, and incorrect answers quietly leak into decisions, emails to customers, and internal processes.
Without a reliable way to ground answers in current organizational knowledge and expose uncertainty, AI remains a risky advisor at best. It can still help draft emails or summarize long threads, but it is not a system you would rely on to answer “What did we actually agree with this customer?” or “Which exceptions have we already approved in similar cases?”
What Employees Experience: Searching, Recreating, Asking Again
For most employees, the problem does not look like “AI strategy.” It looks like another morning lost to finding the one answer they need to move work forward. They jump between Slack, email, Confluence, Google Drive, Jira, and a growing list of tools, hoping that somewhere there is a message, a deck, or a ticket that explains what to do in this specific situation. APQC’s large scale surveys of knowledge workers show that roughly a quarter of their time is lost to productivity drains such as looking for information, requesting it from colleagues, and sitting in meetings that exist mainly to transfer knowledge that already exists somewhere else. Another study summarized in a Nasdaq release finds that more than half of knowledge workers say they cannot find the information they need at work when they need it, and many report recreating content because they simply give up on the search. Harvard Business Review, drawing on Bloomfire’s research, estimates that inefficient knowledge management can consume close to a quarter of a company’s annual revenue, with employees spending significant time both searching for knowledge and rebuilding work that should have been reusable [1][2].
Gartner’s digital worker research adds another layer to this picture. Almost half of digital workers say they struggle to find the information they need to perform their jobs effectively, despite using an increasing number of workplace applications. The tools are there, but they are fragmented. Important decisions and problem solving happen in side channels, comments, and one off calls. Even when people find a document, they worry it might be outdated or incomplete, so they ping the person who “really knows” anyway [8].
This is where most internal AI assistants live today. They sit on top of the same fragmented systems and act as a smarter search box. Instead of writing complex queries, employees can ask a question in natural language and get a list of snippets from Slack, Confluence, or shared drives. That is useful, but it still leaves all of the real work to the human. Someone still has to read through the snippets, reconcile conflicting answers, and decide which one reflects how the company actually operates today. Starmind’s research on productivity drains shows how this plays out over time. Employees repeatedly ask the same questions because previous answers are buried in channels no one can find, they are unsure who the real expert is, and many of the documents they do uncover are no longer accurate. The organization pays three times. Once when the original work is done. Again when someone spends time searching for it. And again when another person recreates it from scratch because the search failed [2]. From the employee’s perspective, it is a loop that never quite closes. They know that somewhere inside the company, someone has already solved this problem or made this decision. AI may help them reach the right folder faster, but it rarely gives them the answer itself. So they keep searching, recreating, and asking again.
A Better Question: What Should AI Actually Know?
Most enterprise AI conversations still start with the wrong question. They ask, “Where can we deploy generative AI?” or “Which use cases should we automate first?” The more useful question is simpler: “What does AI actually need to know about our business to be genuinely helpful to our employees?”. If you follow the thread from day to day work, the answer is not “everything.” Employees do not need an all knowing oracle. They need a system that can reliably answer a narrower set of high value questions. These are the questions that currently trigger long search sessions, repeated pings to colleagues, and high risk guesswork. APQC’s research on knowledge worker productivity makes this clear: the biggest drains come from finding process information, understanding how work is really done, and locating the right person to ask when formal documentation is incomplete. [1][2]
Framed this way, the bar for enterprise AI changes. It is no longer enough for a system to be impressive in a demo or fluent on generic topics. To be useful in real work, it should at least know three things. First, how your organization actually does things today, not just how the intranet says it should be done. Second, what has already been tried, decided, or promised so that teams do not repeat the same experiments or renegotiate the same issues from scratch. Third, where the knowledge is uncertain or contested, so that people see not just an answer, but the confidence and trade offs behind it. This aligns with what Harvard Business Review calls “evolvable scripts” concise, living instruction sets that embody how work is done and can be refined as teams learn. The point is not to freeze a perfect process in place, but to give people a shared starting point that reflects the current consensus and can adapt as reality changes. McKinsey’s work on AI high performers tells a similar story from a different angle. The organizations that capture real value from AI are the ones that embed it deeply into core business workflows and link it to clear decision rights, data sources, and governance rather than treating it as a standalone tool. [4]
Seen through this lens, the gap in most companies is obvious. Their AI does not actually know the things employees wish it knew. It has a vague notion of the domain, pulled from public training data, plus a static snapshot of documents that were easy to index. It does not understand the shape of consensus inside the organization, the history behind key decisions, or the patterns that link similar situations across teams. As a result, it can summarize what is already written, but it cannot reliably answer the question people really ask: “Given everything our company has learned about this topic, what should I do next in this specific case?”
What A Living Organizational Memory Looks Like
Inside most organizations today, consensus lives in people, not in systems. Teams use AI search and internal assistants to pull up documents, old tickets, and Slack threads. Those tools give them snapshots. The real work still happens afterward, when people read through those fragments, debate what is outdated, and quietly reconstruct what the company now believes to be true. Harvard’s work on workplace knowledge flows and knowledge sharing shows why this human layer matters so much. Productivity gains in their field experiments did not come from better storage, but from making it easier for people to access and exchange know how in context. Similarly, Harvard Business Review’s “new approach to knowledge sharing” argues that traditional artifacts like manuals or static trainings are too rigid to keep up with reality. They provide useful anchors, but teams must constantly reinterpret and update them as conditions change. AI search built on those artifacts inherits the same limitation. It helps employees find where an answer might live, but it rarely gives them the answer itself. [9]
A living organizational memory looks different. It is not just an index of past messages. It is a system that continuously digests what people do and decide, reconciles overlapping signals, and captures the current consensus in a form that AI can reason over. Instead of summarizing existing messages, it can respond with “Here is what we usually do in this situation, here is where we have made exceptions, and here are the people who have the deepest experience with this pattern.” It does not pretend that there is a single, eternal source of truth. It reflects the best collective understanding today, and it shows its work. Research on organizational memory and knowledge management underlines this point. Storing information is not enough. The value comes when organizations can retrieve, recombine, and update that information so it actively supports decision making. McKinsey’s analysis of AI high performers reaches a similar conclusion from a different angle. The companies that turn AI into real financial impact are those that treat knowledge architecture and governance as strategic assets, not as side projects. They invest in making sure that when AI gives an answer, it is grounded in current, trusted organizational knowledge rather than in a random mix of old documents and public training data. [4]
This is the gap Lumiérault is designed to close. Instead of asking every employee to re explain how the company works to their personal copilot, Lumiérault continuously captures signals from the tools you already use, organizes them into a graph of entities, decisions, and relationships, and uses that graph to keep a shared, evolving memory of your business. When someone asks a question, the system does not just search for similar sentences. It reasons over this graph, weighs sources by recency and reliability, and exposes both the answer and its confidence, so people can see where consensus is strong and where it is still forming. In practical terms, that means your AI can finally do more than highlight old threads. It can say, “For enterprise customers in this segment, we usually follow this playbook. Legal has approved these exceptions in similar deals. Here is the latest guidance from product, and here are the three cases that do not fit the pattern.” Employees spend less time stitching together partial information and more time making decisions. Knowledge that used to live in a handful of experts becomes accessible to everyone who needs it, without losing nuance.
If your internal AI still feels like an intelligent search bar, you are not alone. Most organizations are in that phase. But the teams that move beyond snapshots toward a true organizational memory will be the ones that capture the real value of AI. If you have not looked at Lumiérault yet and want to see what this could look like in your own environment, it is a good moment to start.