Company knowledge base with AI: what it is, what goes in and how to build one
A company knowledge base is the place where anyone in the company who is entitled to it finds, in seconds, what the company already knows. With AI there is no need to rewrite everything into a wiki: you connect the documents where they are and ask questions, with the source in the answer. Here is what to put in, what to leave out and a 30-day plan.
- 19% of white-collar working time goes to searching for information (IDC and McKinsey)
- 56% of knowledge workers say that often the only way to get information is to ask someone or book a meeting (Atlassian, 2025)
- 70.8% of Italian companies using AI use it to extract knowledge from documents (Istat, 2025)
Key takeaways
- A company knowledge base is what the company knows and needs to work (procedures, manuals, price lists, contracts, answers given to customers), gathered so that whoever is entitled finds it in seconds and in the right version. With AI you query it by asking questions, and the answer cites the document it comes from.
- In goes the knowledge that answers real questions and has an owner. Out stay old versions, duplicates, passwords and personal data the purpose does not require, as the GDPR asks.
- Wikis and shared folders stall because writing and updating them is someone’s second job. In Atlassian’s State of Teams 2025 survey, 56% of knowledge workers say that often the only way to get information is to ask someone or book a meeting.
- AI changes the economics: documents stay where they are, answers cite the source and update when a file changes. It does not fix wrong documents, though: it repeats them, just as confidently.
- In 30 days you start with one team, 30 real questions whose answers you know, the sources connected and permissions decided first. You measure by running the same test again at the end, not by impression.
01 What a company knowledge base is
A company knowledge base is what the company knows and needs in order to work, gathered so that whoever is entitled to it finds it in seconds, always in the right version. It is not a particular piece of software: it is an agreement on three things. What goes in, who can see what, and who keeps it up to date.
It is often confused with neighboring tools. A shared folder stores files, but you search by name and by path: you have to know where to look. A wiki or an intranet is made of pages written on purpose, which work only as long as someone writes and updates them. A general-purpose AI assistant answers anything, but from its own memory, not from your documents.
A knowledge base with AI keeps the useful part of each: it leaves documents where they are, reads them, and lets you ask questions as you would a colleague who has read everything. The answer gives the document and the page it comes from. The technique is called RAG, and we explain it, with the evidence on how much it reduces errors, in this guide.
02 What to put in, and what to leave out
The most useful rule is to start from the questions, not from the documents: what goes in is whatever answers the questions people really ask, every week. Everything else can wait.
| Content | What it is for | Watch out for |
|---|---|---|
| Procedures and work instructions | The most repeated questions: “how do we do this?” | One valid version, with its date |
| Manuals and technical sheets | Answering customers and colleagues without interrupting the expert | Archive superseded revisions |
| Price lists and commercial terms | Quotes, discounts, payment terms | Only those in force: an expired price list produces wrong, confident answers |
| Contract templates and contracts | Clauses, deadlines, commitments made | Access limited to those who need it |
| FAQs and answers given to customers | Knowledge already tested in practice | Remove personal data that isn’t needed |
| Recorded calls and meetings, transcribed | Knowledge nobody ever wrote down | Tell people they are being recorded |
| Training and onboarding material | The first months of newcomers | Update it when the process changes |
And what stays out, or goes in only with restricted access:
- Old versions and duplicates. An AI system repeats what it finds: if two versions of a price list share a folder, sooner or later it will cite the wrong one. In a test presented at NeurIPS 2024, six leading models preferred a wrong document to the correct answer they already knew in more than 60% of cases.
- Personal data the purpose does not require. The GDPR asks that personal data be “adequate, relevant and limited to what is necessary” (article 5). Payslips, performance reviews and medical certificates have no place in a shared knowledge base.
- Passwords and credentials. Never, not even in an “internal” document.
- Documents without an owner. If nobody is responsible for a document, nobody will update it. Better to leave it out until someone takes it on.
Confidential know-how, on the other hand, can go in: with access limited to the people and the AI agents who need it, not to the whole company.
03 Why wikis and shared folders stall
Almost every company has already tried: a “Procedures” folder on SharePoint or Google Drive, a wiki, an intranet. The failure almost always takes the same shape, and it does not depend on the tool.
- Writing is someone’s second job. Every page is born from an hour taken away from real work, and whoever knows the most is also whoever has the least time. The pages stop at the first big deadline.
- Content ages in silence. No document warns you when it becomes wrong. Atlassian’s State of Teams 2025 report observes that people who find a piece of information usually have no way of knowing whether it is current, and too often end up acting on outdated or irrelevant information.
- You search by name, not by question. In a folder you need to know what the file is called and where it is. The real question (“what discount can we give a distributor above €50,000?”) is the name of no file.
- The answer lives in someone’s head. In the same survey, 56% of knowledge workers say that often the only way to get the information they need is to ask someone or book a meeting.
- Permissions are wrong, one way or the other. Too closed, and people give up; too open, and everyone sees everything. Microsoft itself points out, in its guide to preparing SharePoint for Copilot, that SharePoint’s default sharing settings are the most permissive ones.
The bill for all this is paid in time. The estimates are international and do not match, because they measure different things, but they point the same way.
| Source | Sample | What it measures | Figure |
|---|---|---|---|
| IDC and McKinsey Global Institute, 2012 | Managers, professionals and office staff | Share of working time spent searching for information or for the colleagues who have it | 19% |
| Microsoft, Work Trend Index 2023 | 31,000 people in 31 countries | People who say they spend too much time searching for information in their workday | 62% |
| Atlassian, State of Teams 2025 | 12,000 knowledge workers in six countries and 200 executives | Share of the workweek spent searching for information | 25% |
| Atlassian, State of Teams 2025 | As above | People who say that often the only way to get information is to ask someone or book a meeting | 56% |
Sources: for IDC and McKinsey, our July 2026 study; Microsoft, May 2023; Atlassian, 2025 report (United States, United Kingdom, Australia, India, Germany and France, so not Italy).
In our study The cost of knowledge that can’t be found we applied the lowest estimate, 19%, to Italian companies: €66.2 billion a year for the private sector alone, about three points of GDP. It is a deliberately conservative estimate, and it gives the order of magnitude of the problem.
04 What AI changes: the economics of a knowledge base
The traditional knowledge base carries a fixed cost that dooms it: someone has to rewrite into tidy pages what is already written somewhere else, and then keep the two aligned. A layer of AI changes the math in three places.
- No migration. Documents stay in the folders where they already live: Google Drive, SharePoint, OneDrive, Notion, the office file server. Nothing is moved and nothing is rewritten to get started.
- Questions instead of folders. You ask as you would a colleague, and the answer cites the document and the page it comes from. Whoever asked checks it in seconds, and the expert stops acting as everyone’s search engine.
- Updates without the work. When a file changes where it lives, what the AI reads changes too, with no retraining. The first experiments on RAG, in 2020, already showed that replacing the documents is enough to update what a system knows.
What is left to write is only what is written nowhere. And here AI helps in a less obvious way: the questions left unanswered show exactly which pages are missing. You write those, short, one question and one answer, instead of a whole wiki.
This is not an isolated bet. According to Istat, among Italian companies that use artificial intelligence the most widespread technology is precisely the extraction of knowledge and information from text documents: 70.8% use it, more than generative AI (59.1%).
Two limits deserve to be stated plainly. AI does not fix wrong documents: it repeats them, as confidently as it repeats the right ones. And the search can miss the right passage. That is why a citation to the page is not a detail: it is how people notice the errors.
In ekory, as an example, you connect folders on Google Drive, SharePoint, OneDrive and Notion, and websites: files stay where they are and, when they change at the source, ekory updates itself. Ekos, the AI agents that work on that knowledge, answer citing the document and the page, and each one reads only the documents you give it. The sources you can connect are on the Integrations page, the rest on Product.
05 How to build one: a 30-day plan
The surest way to fail is to start with everything. The plan below starts from one team and real questions, and reaches a measured result in a month.
- Days 1 to 5: a scope and 30 questions. Pick a team whose questions repeat (technical support, sales office, accounting) and a person accountable for it. Collect 30 real questions from emails, tickets and chats, and from the people who get asked most. For each one, write down the right answer and the document where it lives.
- Days 6 to 10: sources and permissions. List where those answers live and connect those folders, not the whole drive. Remove old versions of critical documents: price lists, terms, procedures. Decide who sees what by group, before connecting, not after.
- Days 11 to 15: the first test. Ask the 30 questions and sort every answer: right with the right source, right without a source, wrong, “not found”. The errors almost always tell you what to fix: a missing document, an old version, an ambiguous page.
- Days 16 to 25: the team at work. Open it to the team. Every week, look at the questions left unanswered and write the missing pages, short ones. Decide who can correct the answers, and show everyone how to check a source.
- Days 26 to 30: measure and decide. Run the 30-question test again and compare it with the first. If it holds up, pick the second team; if not, you know exactly where to step in.
The step most often skipped is the first test. Without it you will never know whether the system improved or you simply got used to its mistakes.
06 How to measure a knowledge base
A few numbers are enough, collected the same way every time. None of them needs special tools.
| Indicator | How to measure it | When |
|---|---|---|
| Answer quality | The 30 test questions: how many are right, with the right source | At the start, after 30 days, then every quarter |
| Coverage | Questions of the week left unanswered | Every week |
| Search time | A typical week, by sample: how often and how long people search for information | Before you start and after three months |
| Questions to the experts | How many questions reach the two or three people who “know everything” | One week before and one after |
| Freshness | Critical documents with an owner and a review date | Every month |
| Use | People who asked at least one question during the week | Every week |
| Onboarding | Days before a newcomer handles a typical task on their own | With every new hire |
Sampling a typical week to measure search time is the method proposed in our study on the cost of knowledge that can’t be found.
Be wary of savings measured elsewhere: they belong to other companies. Your own number comes from comparing the week before you start with a week three months later.
07 Permissions and governance: who sees what, who answers for what
A knowledge base with AI makes findable what used to be merely hidden. That is its strength, and the reason permissions have to be decided before the sources are connected. Five rules are enough to start.
- Everyone sees what they need. Groups by department or project, and broad access only where it is needed. The GDPR asks this for personal data: by default it must not be made accessible to an indefinite number of people (article 25).
- Permissions apply before the search. A document a person cannot open must not end up in their answer, not even by rephrasing the question.
- Every area has an owner. One person for price lists, one for procedures, one for contracts: they keep the critical documents current and decide what goes in.
- Who corrects is chosen. Everyone can ask; only a few designated people can correct the AI’s answers, so an error typed into a chat does not become knowledge.
- Contract and training. With the provider you need a data processing agreement that says where the data is and who processes it. And the people using AI need training: the AI Act asks anyone who uses it for work to take measures that support their staff’s AI literacy, meaning the skills to use it well (article 4, as amended in July 2026).
Before choosing a tool, check whether it decides permissions itself or takes them from the source folders, and who can change them. The questions to ask a provider are the ones your data protection officer would ask: where the data is, who sees it, and whether it is used to train the models.
Frequently asked questions
What is a company knowledge base?
It is what the company knows and needs to work, such as procedures, manuals, price lists, contracts and answers given to customers, gathered so that whoever is entitled finds it in seconds and in the right version. With AI you query it by asking questions, and the answer cites the document it comes from.
What is the difference between a knowledge base, a wiki and an intranet?
Wikis and intranets are pages written on purpose, which hold up only as long as someone updates them. A knowledge base with AI starts from the documents that already exist and makes them answerable without rewriting them. New pages get written only for the questions that remain unanswered.
Do we have to move our documents into a new system?
With an AI layer, no: you connect the folders where the documents already live, such as Google Drive, SharePoint, OneDrive or Notion, and when a file changes there, what the AI reads changes too. It is worth removing old versions of the most consulted documents first.
How long does it take to build a company knowledge base?
About 30 days for a first team: one week to collect 30 real questions and choose an owner, one to connect the sources and decide permissions, the rest for the first test, daily use and measurement. Extending it to the whole company is a sequence of teams, not one big project.
Who should be in charge of it?
You need a project owner and one person for each important area, such as price lists, procedures and contracts, who keeps the critical documents current and decides what goes in. You don’t need a dedicated department: you need someone who answers the question “how long does it take a newcomer to find what they need?”.
Is a knowledge base with AI compatible with the GDPR?
It can be, if it is built to be: personal data limited to what is necessary, access decided by group and not open to everyone by default (article 25 of the GDPR), a data processing agreement with the provider and data stored where you know. The right questions are the ones your data protection officer asks.
Sources
- ekory, The cost of knowledge that can’t be found, July 2026, with the 19% estimate from IDC and McKinsey Global Institute, The Social Economy, 2012.
- Microsoft, Work Trend Index 2023: Will AI Fix Work?, 9 May 2023.
- Atlassian, The State of Teams 2025, 2025.
- Istat, Imprese e ICT, Anno 2025, 15 December 2025.
- Lewis P. and others, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, NeurIPS, 2020.
- Wu K., Wu E., Zou J., ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence, NeurIPS Datasets and Benchmarks, 2024.
- Microsoft Learn, Get ready for Microsoft Copilot with SharePoint Advanced Management, updated August 2026.
- Regulation (EU) 2016/679 (GDPR), articles 5 and 25.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), article 1, replacing article 4 of Regulation (EU) 2024/1689, 8 July 2026.