Knowledge base privacy comes down to one rule: every article, attachment, search result and AI answer should reach only the people who are meant to see it, and you should be able to show that it did. The work is mostly decided before anyone writes a page. You choose what goes in, who can see it, how long it stays and what the search layer is allowed to surface. Kept is built around this order: participation is opt-in, access is set at the workspace level, and people can review and delete what they have shared.
This guide is for team leaders, HR professionals, project managers and business owners who are building a shared knowledge base, or fixing one that has grown messy. It explains the terms in plain words, then walks through a setup you can follow.
TL;DR
- A knowledge base leaks through more than the article page: search snippets, exports, backups, integrations and AI answers can all expose restricted content.
- Decide what never goes in (credentials, unreleased financials, personal identifiers) before anyone starts writing, and keep only what the work needs.
- Set access by workspace and role, and make sure AI answers respect the same boundaries as the articles they come from.
- Give people control: opt-in participation, a way to review what they shared, and a way to delete it.
- Plan retention up front, including backups, so deletion means deletion.
Table of Contents
- What knowledge base privacy actually covers
- Public, private and hybrid: choosing the right boundary
- What to keep out of the knowledge base
- Set access around roles, not around hope
- What AI search changes
- Give people control over what they share
- Retention and deletion: plan the end at the start
- A step-by-step setup for a private business knowledge base
- What this looks like for onboarding
- Perspective
- Where Kept fits
- FAQ
- Sources
What knowledge base privacy actually covers
A knowledge base is a shared place where a company keeps answers: how a process runs, why a decision was made, what to do when something goes wrong. Privacy in that place has two sides.
The first is confidentiality of company information: pricing logic, security settings, client details, plans that have not been announced. The second is privacy of people: names, employee IDs, personal opinions, performance conversations and anything else that points to an individual.
Both sides are exposed in more places than most teams expect. One guide to knowledge base security lists search snippets, notifications, cached pages, analytics, exports, backups, integrations and retrieval-based AI as paths by which a restricted article can still escape. If you only lock the article page, you have locked one door in a house with several.
That is why privacy belongs in the design of the knowledge base, not in a policy filed beside it.
Public, private and hybrid: choosing the right boundary
Most teams end up choosing among three setups.
| Setup | Who can read it | Typical content | Privacy question to ask |
|---|---|---|---|
| Public | Anyone, no login | Help articles for customers, product guides | Has anything internal slipped in through an attachment, screenshot or comment? |
| Private | Signed-in members of one workspace | Processes, decisions, onboarding material, judgment calls | Are roles and groups set so people see only what their job needs? |
| Hybrid | Some spaces open, others restricted | Customer help plus internal runbooks | Do the two spaces share a search index or an export that crosses the line? |
For the reader of this article, who is documenting how the company really works, the private workspace is the home. The reasons behind a process, the exceptions and the workarounds are exactly the content you do not want on the open web. If you also publish customer-facing help, keep it in a separate space with its own permissions, not a folder in the internal one.
A good test: imagine the most curious person with access to each space. What could they learn by browsing, searching and asking the AI assistant? If the answer is more than their role needs, tighten the boundary.
What to keep out of the knowledge base
The most effective privacy control is not storing something in the first place. The FINOS AI Governance Framework names the categories worth filtering out before content enters a knowledge base that feeds AI:
- Personal information: names, financial account numbers, employee IDs, Social Security numbers.
- Proprietary business information: trade secrets, unreleased financials, merger and acquisition details.
- Sensitive operational data: security configurations, credentials, audit findings.
- Confidential customer data.
- Regulated or compliance-sensitive information.
The same framework recommends masking, redaction, generalization and tokenization (replacing a real value with a stand-in) where the information is needed in some form. It also recommends the principle of least exposure: include only the data the business function requires, and de-identify it where possible.
In practice that gives you three habits:
- Describe the step, not the secret. Write "log in to the billing system with the shared finance account" rather than pasting the password. Keep credentials in a dedicated secrets manager.
- Replace identifiers with roles. "The account manager for the client" lasts longer and exposes less than a named person and a customer number.
- Write the reason, not the gossip. A good entry explains why a process works this way and when it applies. It does not need a verdict on a colleague.
This matters because personal data usually enters a knowledge base by accident, through pasted tickets, email exports and meeting notes. For help with spotting it, this overview of PII handling in AI covers how personal data should be dealt with at the point it enters a system.
If you are not sure which categories of people your documents touch, the NIST Privacy Framework gives a useful discipline: it asks organizations to inventory the categories of individuals whose data they process, such as employees, prospective employees and customers. A one-page list of those categories, and which spaces they might appear in, is a strong starting point.
Set access around roles, not around hope
Once you know what belongs, decide who sees it. A workable model has three layers.
Workspace level
Separate workspaces for separate audiences: a personal workspace for one professional's own notes, and a business workspace for a team or company. Kept offers both, with workspace-level access controls, so a person's private notes and the team's shared answers do not blend by default.
Role and group level
Within a business workspace, match access to the job. HR material on onboarding may be open to all new hires, while content about compensation or legal matters stays with a small group. Review the groups when people change roles, because stale permissions are one of the most common ways content ends up in front of the wrong person.
Edit rights versus read rights
Even open content needs a short list of people who can change it. Wide edit access invites accidental overwrites and accidental disclosure alike.
A simple routine: every quarter, list who can see each space, and remove anyone who no longer needs it. Put an owner's name next to every space so the question "who decides?" always has an answer.
What AI search changes
A knowledge base you can ask in plain words is far more useful than one you have to browse. It also adds a new surface to protect: the search and answer layer.
The European Data Protection Board's guidance on large language models, published in April 2025, recommends input filters or warnings to discourage people from entering personal data, automated detection of such data, and encryption in transit and at rest. Those are sensible baselines for any AI-enabled knowledge base, wherever you operate.
Two questions then decide whether your setup holds up:
Does the assistant respect the same permissions as the articles? If a person cannot open a page, the assistant should not quote from it. Access limits should apply when the system decides which content to consider for an answer, not only when the answer is displayed.
Can you see where an answer came from? Citation-backed answers let a reader check the source page and let an owner spot content that should not have been available. Kept answers with citations, so each response points back to the material behind it.
It also helps to treat the search index as its own data surface. When a document is copied, split into chunks and made searchable, each chunk can carry sensitive content. A narrower index reduces exposure; a broader one gives fuller answers. Decide on purpose, space by space, instead of indexing everything by default. For a deeper look at how conversational assistants can leak information and which safeguards address it, see this piece on privacy risks in chatbots.
If you want to understand how this kind of retrieval works under the hood, our guide to semantic search for a knowledge base explains it without the jargon.

Give people control over what they share
A knowledge base that captures how people really work depends on those people trusting it. If they fear that what they say will be used against them, they will say less, and what remains is the tidy official version.
Three design choices protect that trust:
- Opt-in participation. Each person chooses whether to take part. Nobody is enrolled by default.
- Review before it is final. People can read what has been recorded in their name and correct it.
- Deletion on request. If someone wants a contribution removed, there is a defined way to do it.
Kept works this way: taking part is each person's choice, the platform never scores or ranks anyone, and nothing a team says is used to train AI. The point of the system is to show the company where people do the same job in different ways, so leadership can decide which way becomes the company's own.
Tell your team this plainly before you start. A two-sentence note covering what is recorded, who can see it and how to opt out will do more for adoption than a long policy.
Retention and deletion: plan the end at the start
Privacy is not only about who can read content today. It is also about how long you keep it and what happens when you delete it.
A few points to settle early:
- Retention periods. Not everything deserves the same lifespan. Our article on three retention bands for conversational documentation shows one way to sort content by how long it stays useful.
- Backups. Deleted content can linger in backups. Write down who can access backups and how a restore is approved, so a deletion in the live workspace is not undone silently.
- Copies outside the system. A private workspace does not stop someone with legitimate access from copying content elsewhere. Minimizing what you store, and keeping secrets out, limits the damage if that happens.
- Departures. When someone leaves, you want to keep their working knowledge for the team without keeping personal material they would expect to be removed. Our knowledge retention plan walks through that balance.
Deletion is also a legal matter in many places. The EU's General Data Protection Regulation, Regulation 2016/679, gives individuals a right to erasure of their personal data, and similar rights appear in other countries' laws and in many US state privacy laws. If your company has employees or customers in the EU, or in a US state with such a law, ask your legal adviser which rules apply to the personal data in your knowledge base. A business that keeps personal data out of the knowledge base to begin with has far less to erase.
A step-by-step setup for a private business knowledge base
Here is a sequence you can run in a week.
- Name an owner. One person is accountable for the privacy of the knowledge base. In smaller companies this is often the operations lead or the HR manager.
- List what must stay out. Use the categories above, and add anything particular to your industry. Share the list with everyone who will contribute.
- Split the spaces. Create separate spaces for different audiences, and decide whether any content is ever public. Keep public and internal material apart.
- Set roles. Assign read and edit rights by role. Start narrow and widen on request.
- Connect systems deliberately. Each integration is a path for data to flow in. Connect only the systems the work needs, and note what each one brings. Kept connects to the systems you already use, so the numbers behind a process sit next to the reasons for it. See how Kept works with the systems you already use.
- Tell your team. Explain what is recorded, who sees it, and how to review or delete their contributions.
- Test the answers. Ask the assistant questions a new hire should and should not be able to answer, and check that the citations and the boundaries match.
- Review on a schedule. Quarterly is a good rhythm for access lists, retention and stale content.
For a fuller walkthrough of making a knowledge base the whole team actually uses, read how to build an internal knowledge base that gets used. For how the layers fit together, the five-layer knowledge base architecture guide is a good companion.
What this looks like for onboarding
Onboarding is where privacy and usefulness meet most visibly. A new hire needs real answers, such as "What do we do when a client asks for an exception?", without seeing payroll data or another team's confidential plans.
With a private workspace, role-based access and citation-backed answers, the new hire asks in their own words and gets an answer drawn only from what their role may see, with the source shown. The HR team stops answering the same questions repeatedly and does not have to hand out broad access to make that happen. Our guide on AI for HR onboarding covers this in more detail.
Perspective
We see the privacy question differently from the way most knowledge tools frame it. Other tools index the documents a company has already written, so privacy becomes a matter of locking files. But the knowledge that runs a company mostly sits in people's heads: why a step is done this way, when it applies, what the exception looks like.
That knowledge only gets shared when people trust the place it goes. So we build for trust first. Taking part is each person's choice. Nobody is scored or ranked. Nothing a team says is used to train AI. People can review what they have shared and delete it.
Access follows the workspace, answers come with their sources, and the company decides what becomes its own way of working. We think the best privacy design is the one that makes people willing to tell you how the work really gets done.
Anthony
Where Kept fits
If you want a private place where your team's processes, reasons and exceptions live together, Kept offers a plan for individuals, kept for you, and a plan for teams, kept for business. Both come with personal and business workspaces, opt-in participation, review and deletion, and answers that cite their sources. Pricing is on the pricing page, and you can read more about the team plan on the Kept for Business page.

To see how Kept handles data protection, visit the Security page and the Trust Center.
FAQ
Is a private knowledge base automatically secure?
No. Private means only signed-in people with the right permissions can open it, but content can still travel through search, exports, integrations, backups and copying. Keep sensitive data out, set roles carefully, and review access regularly so the boundary stays accurate.
What should never go into a company knowledge base?
Credentials, Social Security numbers and other personal identifiers, unreleased financials, security configurations and confidential customer data are the main categories. FINOS lists them as data to filter before content reaches an AI-enabled knowledge base. Write the process and the reason, and leave the secret itself in a secrets manager.
Does AI search make a knowledge base less private?
It adds a layer you need to design for, because the assistant can only be as private as the permissions behind it. Apply the same access rules to answers as to pages, show the source of each answer, and keep the indexed content narrow where the material is sensitive.
Can employees ask to have their contributions removed?
They should be able to. A well-run knowledge base lets people review what they have shared and delete it, and your retention plan should say what happens to backups. In Kept, participation is opt-in and review and deletion are built in.
Sources
- Knowledge base security and compliance guide
- FINOS AI Governance Framework, MI-2: Data Filtering from External Knowledge Bases
- NIST Privacy Framework 1.1 initial public draft
- EDPB: Large Language Models (LLMs), AI privacy risks and mitigations
- Regulation (EU) 2016/679, EUR-Lex
- PII Handling in AI: How to Keep Personal Data Safe
- Privacy in Chatbots: Risks, Leakage Paths and Safeguards
