Skip to content

16 November 2026, 8 min read

Corporate wikis vs company knowledge bases: what changed when AI started reading them

Corporate wikis were built for people to browse. AI agents read differently, and that changes what a company knowledge base has to be.

Corporate wikis solved a real problem. In the 1990s and 2000s, company knowledge lived in heads and in email, and a page anyone could edit was a large improvement. Ward Cunningham's original wiki, installed in 1995, was designed around that idea: collaborative editing by the audience, directly in a browser. Most company wikis still work that way.

The trouble is that the reader changed. For twenty years the reader was a person who typed a keyword, scanned ten results and used judgement. Now a growing share of reads come from software that retrieves a few passages, treats them as true, and answers in a confident voice. A wiki that was good enough for the first reader can be a liability for the second.

This post looks at what a corporate wiki is good at, where it breaks when AI reads it, and what a company knowledge base needs that a wiki does not.

What a corporate wiki is good at

Be fair to the wiki. It is cheap to start, easy to write in and flexible about shape. A page can be a meeting note, a runbook, a policy or a half-formed idea. Nobody needs a schema before they can contribute.

Those strengths are real, and they explain why wikis win early. A team of ten with a wiki will usually find what it needs, because the team knows who wrote what and roughly when.

Wikis are also forgiving of messy structure. If two pages say different things, a person notices, asks a colleague and moves on. That social repair is invisible, and it is doing a lot of the work.

Where wikis break

The failures are well known to anyone who has maintained one for a few years.

Pages go stale without anyone noticing. A page has an edit date, but nothing says whether the content is still true. The org chart page from March lists someone who left in June.

The same fact lives in several places. The notice period appears in the handbook, in an onboarding page and in a slide that was pasted into a page. When it changes, one copy gets updated.

Structure is by convention. A "People" section is a list of pages that happen to follow a template. Nothing stops a page from omitting the manager, so nothing can reliably answer "who reports to Dana?".

Permissions are coarse. Many wikis control access by space or by page. That is fine for a handbook. It is awkward for a salary field on a person's page, which needs a different audience from the rest of the page.

None of this was fatal when a human reader compensated. It is a different matter when software reads the page.

What changes when AI reads it

An AI agent does not browse. It runs a search, takes the top results and reasons over them. Three things follow.

First, stale and contradictory pages become answers. A person who sees two conflicting pages stops and checks. A retrieval system may blend them. Retrieval-augmented generation, introduced in a 2020 paper from Lewis and colleagues, pairs a language model with a retriever over a document index. The model answers from whatever the retriever returns, so the answer is only as good as the pages behind it. Retrieval does not fix a wiki. It amplifies it.

Second, access becomes a bigger deal. A wiki search box shows a person only what they may open. A connector that indexes the wiki with an administrator's account may index everything, and then answer from it for anyone. OWASP lists sensitive information disclosure as a named risk for LLM applications and recommends limiting access to data on a least-privilege basis. A system prompt that says "do not reveal salaries" is not access control. The filtering has to happen before the model sees anything.

Third, writes become possible. Once an agent can update a page, the question of who is right when a person and an agent disagree stops being theoretical. A wiki has page history, but history is a log, not a policy.

A fact nobody owns is a fact nobody corrects, and an AI agent will repeat it with the same confidence as one somebody checked.

Corporate wiki vs company knowledge base

The names are used loosely, so here is the distinction we find useful. A corporate wiki is a set of pages. A company knowledge base is a set of records and pages, with structure and access rules that software can rely on.

Question Corporate wiki Company knowledge base
Unit of content Page Record (typed fields plus a page)
Structure Template by convention Collections with typed properties and relations
"Who reports to Dana?" Search and hope A query over a relation
Access control Space or page Per record and per field, enforced on every read path
AI reads as Whoever the connector is The person it acts for
Disagreement Last edit wins A proposal for the owner to review
Staleness Edit date An owner and a review date

The table is not an argument that wikis are bad. It describes a different job. A wiki is good for writing. A knowledge base is good for being asked questions by software.

A worked example: the salary field

Take a People collection. Each person has a name, role, manager, team, start date and salary. The first five are things most of the company may see. The last is not.

In a wiki, the person's page is one block of text. You can restrict the whole page, which hides the manager and team too, or leave it open and keep salary somewhere else, which means a second system. Either way, an AI tool connected to the wiki has no concept of "this one field".

In a knowledge base, salary is a property with its own audience. The record is visible; the field is visible only to the people it was shared with. When an AI tool asks "who is on the platform team and who do they report to?", it gets the answer. When it asks for salaries, the field is simply not there for that asker, and it is not there for the tool acting on their behalf either.

That is a small example, but it generalises. Customer contracts, board notes, performance reviews and security findings all have the same shape: one record, several audiences.

How to evaluate internal knowledge base software

If you are choosing internal knowledge base software today, ask questions that a demo cannot fake.

  1. Where is access enforced? Ask whether the rule runs on the search path, the API, the export and the AI connector, or only in the web interface.
  2. Whose access does an AI tool get? The right answer is the access of the person it acts for. A shared service account is the wrong answer even when it is convenient.
  3. Can a field have its own audience? Page-level permissions are not enough for anything containing a mix of public and sensitive data.
  4. What happens on conflict? If an agent or an import contradicts something a person wrote, ask whether the old value is overwritten or the change waits for review.
  5. Can an administrator read everything? Some roles need an override. The question is whether using it leaves a trace and tells the owner.
  6. How does content get in? Upload, database read access and a standard protocol such as MCP cover most needs. A long list of connectors is less useful than a few that work properly.

Most wikis do well on the first half of that list for people and badly for machines. Knowing which of your needs are for people and which are for machines is half the decision.

Do you have to choose?

Often not. Many teams keep a wiki for prose and add a structured layer for the records that matter. The risk is duplication, so it helps to decide where each fact lives. Prose, runbooks and meeting notes stay in the wiki. People, teams, projects, policies and their relations go into typed collections.

A reasonable migration path:

  • Start with one collection that causes the most pain, usually People or Policies.
  • Import existing pages and CSVs rather than retyping them.
  • Name an owner for every record.
  • Turn on AI access for that collection only, and watch what it returns before widening it.

We cover the access side in more detail in how to give AI agents your company's knowledge without leaking what they shouldn't see, and the question of who is right when an agent and a person disagree in when AI and people disagree.

How Contextely approaches this

Contextely is a company knowledge base built around collections such as People, Teams, Projects and Policies. Each has typed properties, relations, a rich page per record, and views such as table, board, calendar, timeline and org chart.

Everything is private by default, and sharing is a deliberate step to people, teams or everyone. AI tools connect over MCP, a CLI or a REST API, and each one sees exactly what the person it acts for can see. If an agent contradicts a value a person wrote, it files a proposal instead of overwriting it. Pricing starts free, with unlimited readers on every plan (see pricing).

The short version

Corporate wikis are still the right tool for writing things down. They were never designed for a reader that retrieves fragments, trusts them and answers for you. A company knowledge base keeps the pages and adds the three things that reader needs: structure it can query, access it cannot exceed and a way to disagree with a person without overwriting them.

Questions

What is the difference between a corporate wiki and a company knowledge base?

A corporate wiki is a set of editable pages that people browse and search. A company knowledge base keeps the same pages but adds structure (typed records, relations, owners) and access rules that software can enforce, so people and AI agents can both read it safely.

Can an AI tool read my company wiki?

Yes, but most wikis give it one of two bad options: read everything the connector's account can see, or read only what was public. Neither matches the rule you actually want, which is that the tool sees what the person asking can see.

Do we need to replace our wiki to get a company knowledge base?

Not necessarily. Prose that works in a wiki can be imported as pages. What usually needs a new home is the structured material: people, teams, projects, policies and the relations between them.

How do I choose internal knowledge base software?

Test four things: whether records have typed properties, whether permissions are enforced on every read path including the API, whether AI tools inherit the asker's access, and what happens when an AI tool disagrees with a person.

Start with what you already have.

Free for up to five editors. Unlimited readers on every plan.