What Not to Put in an AI Chatbot Knowledge Base: Risky Content

The content you put in an AI chatbot knowledge base decides what the chatbot can say, to whom and with what confidence, so what you leave out matters as much as what you add. An assistant that answers only from your content is safe only if the content is safe: if a draft price list, an internal margin table or a customer’s personal details sits in the knowledge base, the assistant may repeat it to anyone who asks the right question. This guide lists what not to put in an AI chatbot knowledge base, why each kind of content is risky and how to keep the knowledge useful, current and clean.

Talkmio is an AI live chat. Its assistant, Mio, reads your website and the documents you upload and answers visitors only from that content. That design avoids invented answers, and it also means one rule must guide everything you add: treat the knowledge base as if it were a public page.

Treat the AI Chatbot Knowledge Base as Public

The simplest test before adding anything is to ask whether you would be comfortable with every visitor reading it. A chatbot does not know who is reading. It answers a question from the content it has, and a clever or curious visitor can ask many questions in a row. Anything in the knowledge base is, in practice, available to whoever can start a chat.

That is why you should never rely on an instruction such as “do not share the internal price list” to protect a document. Instructions shape tone and limits, but they are not access control. If the content must stay private, keep it out of the knowledge base. The same principle is described for language-model applications in general in the OWASP guidance on large language model risks, which lists sensitive information disclosure among the main risks. Our guide to prompt injection and chatbot security explains how visitors try to talk assistants into oversharing.

What Not to Put in an AI Chatbot Knowledge Base

The table below groups the content that causes most trouble, with the reason and a safer alternative.

Do not add	Why it is risky	Safer alternative
Personal data of customers or staff	It can be repeated to the wrong person and creates data protection duties	Keep it in your CRM; add only general policies
Internal prices, margins and supplier terms	Visitors can ask for them and the assistant may answer	Publish only the prices customers may see
Drafts and unapproved text	The assistant treats a draft as a fact	Add content only after approval
Expired offers and old policies	Visitors act on the wrong price or deadline	Remove or replace with a dated update
Contradictory documents	The answer depends on which piece is found first	Keep one source of truth per topic
Contracts and documents under NDA	Breach of confidentiality	Summarise public terms on a page
Passwords, keys and access instructions	A security incident waiting to happen	Never store them in content at all
Staff handbooks and internal procedures	They contain exceptions and escalation tricks	Publish a customer-facing version

The common thread is that every one of these is useful inside the company and harmful outside it. A clean split between internal documents and customer-facing documents is the single best habit for a safe knowledge base.

Personal Data: Keep It Out Entirely

Customer lists, order exports, support transcripts with names and e-mail addresses, and HR documents do not belong in a chatbot knowledge base. Under data protection laws such as the GDPR, personal data needs a lawful reason, a defined purpose and limited retention, and a chatbot knowledge base is none of those. The overview of personal data shows how wide the definition is: a name, an e-mail address, an order number linked to a person, even an IP address.

The practical rule is to add policies, not people. A document that says how refunds work is fine. A spreadsheet that lists refunds by customer is not. If you want Mio to improve from real conversations, copy the question and the correct answer into a Q&A pair with the personal details removed. Talkmio stores data on servers in the EU, in Germany, and you can export or delete conversations at any time, but the safest data is the data you never added.

Internal Numbers, Margins and Negotiation Rules

Sales and operations documents often mix public and private facts: a price list with wholesale tiers, a discount matrix with approval limits, a supplier agreement with minimum quantities. If such a file is uploaded because it contains one useful table, the rest comes along. Split the file first. Create a customer-facing version that contains only what customers may see, and upload that.

The same applies to rules for staff, such as “give up to ten percent if the customer complains”. That sentence helps an employee and harms the business if a visitor can ask for it. Keep negotiation rules, escalation thresholds and fraud checks out of the content, and let Mio hand such conversations to a person. Mio already hands over when a visitor asks for a person, complains or asks about a specific order.

Drafts, Outdated Pages and Contradictions

The most common failure is not leakage but staleness. An old pricing page, a promotion that ended last spring or a policy that was replaced still lives on the website or in an uploaded PDF, and the assistant repeats it faithfully. Visitors then act on wrong information and blame the company, not the assistant. Review your sources on a schedule, and remove or replace anything that is out of date. Our article on keeping an AI chatbot knowledge base updated offers a simple routine.

Contradictions are the next problem. When two documents disagree, for example a returns policy of fourteen days on one page and thirty days on another, the answer depends on which piece is found first. Read our guide to conflicting sources in an AI chatbot to find and fix them. Use Try Mio to ask the same question in different words and see which knowledge pieces were used.

Content You Do Not Have the Right to Use

Copyright and confidentiality apply to chatbot content just as they do to a website. Do not upload a competitor’s brochure, a paid industry report, a supplier’s manual or an article you have no licence to reuse, and do not copy long passages of someone else’s work. Mio answers from your content, so any text you add can be shown to visitors, and a rights holder will not care that the words came through a chatbot.

Third-party documents under a non-disclosure agreement are a special case. Even summaries can breach the agreement. When in doubt, ask whoever owns the relationship, and add only a public statement that you are permitted to make. If a document is useful but not allowed, a better route is to write your own short description in your own words and publish that.

How Talkmio Reads Your Sources

Knowing how the tool reads your content makes the rules easier to follow. In Mio AI, under Knowledge base, you can add your website, and Mio reads every public page on the same domain while respecting robots.txt. You can add single pages, Q&A pairs, free text or files, including PDF, DOCX, TXT, MD, CSV and HTML up to 15 MB, and sources can be re-read on demand. The plan limits how many pages Mio can keep.

Two consequences follow. First, if a page is public on your domain, it can be read, so private information must not be published on a public page in the first place. If you must exclude a public page, remember that robots.txt rules also affect search engines, so test the change carefully. Second, selective sources give you control: you can add only the pages and files you choose instead of the whole site. Business instructions then set the tone and limits, such as never quoting delivery dates, and Try Mio lets you test before going live.

Examples of Content That Looks Harmless but Is Not

Some risks hide in ordinary-looking files. An FAQ export from a help desk may include customer names in the example questions. A price list may contain a column of cost prices that nobody noticed. A slide deck may carry speaker notes with candid remarks about competitors or customers, and a spreadsheet may keep hidden sheets with the data it was built from. A PDF may include comments, tracked changes or metadata with the names of staff. Open each file the way a stranger would, look at every sheet, note and comment, and export a clean copy before you upload it.

Website pages have their own versions of the problem: a staging page that was never removed, a thank-you page with order details in the address, a forgotten test page with lorem ipsum. Ask Mio about your own company with a few unusual questions, such as “what is your cheapest supplier?” or “who is your biggest customer?”, and see what happens. The right answer is that it does not know, and a hand-off to a person if the visitor insists.

A Short Review Routine

Run a ten-minute review before adding any new source and a longer one each quarter. For a new source ask five questions: who owns it, is it approved, is it current, does it contain personal or internal data and does it contradict anything else. For the quarterly review, list every source, remove what is outdated and ask Mio ten awkward questions to see what leaks. Our guide to testing an AI chatbot before launch contains a question list you can reuse.

Keep a one-page register of what is in the knowledge base, who approved it and when it was last checked. It sounds bureaucratic, but it takes minutes and makes every future change safer. Plans and limits are on the Talkmio pricing page, and the Free plan, with 20 knowledge base pages, is enough to practise the routine.

Frequently Asked Questions

Can I tell the chatbot not to reveal a document in its instructions?

Do not rely on that. Instructions shape tone and limits, but they are not access control. If content must stay private, keep it out of the knowledge base entirely, and publish a customer-facing version of anything that mixes public and private facts.

Is it safe to upload customer conversations to train the assistant?

Not as raw transcripts. They contain personal data. Turn useful exchanges into Q&A pairs with names, e-mail addresses and order details removed, and add only policies and general answers that you would be happy to show any visitor.

What happens if my website contains outdated pages?

Mio reads every public page on your domain, so it may repeat an outdated price or policy. Remove or update old pages, re-read the source in Talkmio and use Try Mio to check which knowledge pieces an answer used.

Can I keep internal documents and public documents together?

Keep them apart. Add only the files and pages you choose, and create customer-facing versions of any document that mixes internal and public information. Always treat the whole knowledge base as if every single visitor could read it, because in practice they can.

How often should I review the knowledge base?

Check each new source before adding it, and review the whole knowledge base every quarter or after any change in prices, policies or products. Also review it when Try Mio or real chats reveal a wrong or contradictory answer.

Does robots.txt control what Mio reads?

Yes, Mio respects robots.txt when it reads your website. Remember that robots.txt also affects search engines, so test any change carefully. A much safer approach is simply never to publish private information on public pages at all, whatever the instructions say.

The Bottom Line

An AI chatbot knowledge base is only as safe as the content you put in it. Keep personal data, internal numbers, drafts, outdated pages, contradictions and documents you have no right to use out of it, and add customer-facing content that is approved and current. Then test the result with awkward questions. Start free with Talkmio, add your public pages and try to make Mio say something it should not before your visitors do.