Control which knowledge base articles your AI uses

In each of the tools covered here, removing an article from the collection that grounds your AI is easy. Deciding what to remove is harder, because someone has to notice that the article is out of date or no longer relevant.

Four controls decide what content an AI uses: which kinds of source are available to it, such as a help center, a wiki, email or the CRM; what it can access inside each source; whether a page’s age affects whether it gets used; and how content is maintained, meaning whether someone is responsible for it and gets told when it goes stale, or has to check on their own. Zendesk’s and Intercom’s AI agents cover the first two. Only Guru covers the third. It can keep an overdue page out of its AI’s answers. Guru and Document360 both send review reminders, and in Document360 an overdue article stays live for readers while the reminder goes to staff.

Key takeaways

• Zendesk’s AI agents read the help center live: “When an AI agent searches a connected help center, it searches the current content of the help center at the time the search is performed.” Correct an article and the next answer is already different.
• The same agent is limited to what that customer could see anyway: “If the customer is unauthenticated, the AI agent can use only public articles to generate its response.”
• Intercom separates who can read an article from whether Fin can use it, so you can enable or disable content for Fin item by item or in bulk, and Fin audiences built from People and Company attributes decide which customers get which content.
• Guru can keep an overdue card out of AI answers: a card becomes unverified when its verification interval runs out, and a Knowledge Agent set to “Verified only” searches verified content alone.
• Document360 marks an overdue article Stale and keeps publishing it: “A stale article continues to appear and function normally on the knowledge base site.”

Scoping, and the difference between two settings that look alike

An AI answer tool starts by pointing at a source. Zendesk’s advanced AI agents connect to one or more of them, and Zendesk recommends keeping the number within a reasonable limit. The agent reads the help center at the moment it answers. “When an AI agent searches a connected help center, it searches the current content of the help center at the time the search is performed.” Edit the article, and the next answer based on it changes.

One layer down is the per-article switch. Intercom lets you enable or disable each piece of content for Fin from the Details panel while you edit it, or in bulk from the Content page in Knowledge, so an article can stay in the help center but out of the AI’s replies.

Help center visibility and AI use look like one setting, and a tool that collapses them into one will frustrate you. Who is allowed to read this article is a permissions question. Whether the AI may quote it when answering is an editorial one. A deprecated migration guide should stay readable and out of AI answers: customers on the old version still need to find it, but nobody wants it showing up as the answer to a new customer’s question. Intercom adds Fin audiences on top, built from People and Company attributes, so Fin can use one page for enterprise accounts and leave it out for self-serve.

Zendesk takes the other approach and ties the AI to what the customer could already see. “If the customer is authenticated, the AI agent can use relevant restricted articles to generate its response. If the customer is unauthenticated, the AI agent can use only public articles to generate its response.” The generative replies FAQ adds a requirement: “You need a public knowledge source, written in any language, for the AI agent to generate answers from.”

Zendesk’s default keeps the AI from showing a customer anything they couldn’t read themselves, but it’s blunt. If you unpublish an article to keep it out of AI replies, it also disappears for every customer searching your help center and for the support agent who was linking to it in tickets.

Freshness rules mostly flag pages rather than block them

Document360 handles freshness through review reminders. Set a review date, and when a published article passes it, the article “enters a Stale state”, showing a red dot in the category tree and a Stale badge in the editor. Readers see no change: “A stale article continues to appear and function normally on the knowledge base site. The Stale state is an internal indicator visible only in the knowledge base portal. Readers on the site are not notified that an article is stale.”

Guru goes a step further. Each card has a verifier and a verification interval: does not expire, 30, 60 or 90 days, six months or a year, with does not expire as the default for new collections. The interval triggers reminders to the verifier, and when it runs out the card becomes unverified. The difference is on the AI’s side: a Guru Knowledge Agent can be set to search verified content only, which Guru describes as “highest quality, narrowest coverage”, so an unverified card drops out of its answers until someone verifies it again. Knowledge Agents also review content once a day and can verify or unverify it on their own, using signals such as thumbs-down feedback and what the content says. Auto-unverify is off by default.

Intercom handles time differently. You can schedule snippets, internal articles and external documents to turn on or off for Fin at a set date, time and timezone, optionally with an end date so the change reverses itself. A schedule also overrides later changes made by hand: “Manual changes don’t cancel a pending schedule. If you manually toggle availability after setting a schedule, the scheduled transition will still fire at its scheduled time and override your manual change.” Somebody correcting an item by hand can have their fix undone by a schedule set earlier.

There’s a case against letting age block anything, because age is a poor proxy for wrongness. A refund policy written years ago can still be exactly right, while a pricing page edited recently can already be out of date. Guru’s own advice leans the same way: “For most content, we recommend using ‘Does not expire’ and letting Knowledge Agents handle ongoing quality checks automatically."

"Block outdated pages” is three different requests

This page is wrong and I know it. Unpublish, archive, delete, or switch it off for the AI. Every tool covered here does some version of this, and it’s the only one of the three that takes effect the moment you press the button.

This page is old and might be wrong. Review dates and verification intervals answer this with a task for a person. Guru’s verified-only setting is the one place among these tools where lapsed time keeps a page out of AI answers. Intercom’s scheduled availability is a date-driven switch, but you set it per item in advance rather than deriving it from age.

This page contradicts a newer one. This is the request that would help most, because a contradiction is evidence of wrongness in a way a timestamp isn’t. Guru comes closest. An auto-unverify rule can flag a card as a near-duplicate of another, and Guru publishes a skill template for its Knowledge Agents that lists cards that contradict each other, quoting both sides “so a human can settle it in a minute instead of re-researching it.” Guru says the template is “not intended to be an out-of-the-box solution,” and deciding which card is right remains someone’s responsibility.

An old page that might be wrong is where these tools help least. A review date tells someone to look, but not what changed, so they reread the whole page and check it against current policy.

What connecting your CRM and support inbox gets you

The other thing to check in these tools is whether they reach your CRM and your support inbox. That connection runs in two directions, and vendors document one of them far better than the other.

The reading direction is well covered. Zendesk’s and Intercom’s AI agents both work inside their own help desks, so the answer goes out in the conversation the customer started, and Intercom’s audiences shape it from People and Company data.

The writing direction decides whether your docs stay true, but it’s thinner. Zendesk’s is the clearest example. A support agent working a ticket can flag an article from the knowledge section of the context panel, and Zendesk creates a ticket with a link to the flagged article, the source ticket, the agent’s username and their inline comments. On Enterprise plans Zendesk assigns that ticket to the article’s owner. On other plans it stays unassigned.

Every step depends on someone making time. A support agent has to notice the problem and file the flag, and the owner has to edit the article, so updates slip when support gets busy.

The page went stale because keeping it current was nobody’s job

Blocking an outdated page is a patch on a problem one level down. The page went stale because the information that would have corrected it never reached it.

Picture how that happens. Somebody announces a price change in a thread. A partner sends revised terms as an attachment. A customer replies that the setup steps in an article no longer match the screen. Support explains the new policy to customers by email before anyone opens the docs. None of it is secret, but nobody transcribes it into the place the AI reads.

Panopto’s 2018 study with YouGov, a survey of 1,001 US adults with at least five years of work experience at organizations of 200 employees or more, found that 42% of institutional knowledge is unique to the individual, and that 60% of respondents found it difficult, very difficult or nearly impossible to obtain information vital to their job from colleagues.

Keeping a help center current requires somebody to notice that a conversation contradicts a documented fact, then fix the article. In a company with no one whose job it is to do this work, a reminder that an article is due for review goes to nobody in particular. We wrote up the wider version of this argument in the missing middle of company knowledge, and the published research is on our unrecorded knowledge page.

Where we fit

this+that reads every message your team gets across Gmail, Outlook, Slack, Microsoft Teams, Google Chat, and WhatsApp Business, turns the requests in them into tasks in DoBox, and routes each one to a workflow or to a person. The knowledge layer is the brain: pages in personal, team and public spaces, with version history and conflict detection when two people or two agents edit the same page.

A page’s space decides who can see it. Personal pages are visible only to you. Team pages are shared with your team. Public pages are the ones you choose to make public as static content.

Writing back to the brain addresses the maintenance problem. Workflows read the brain to ground what they do, drafting included, and they write back to it as they run. The prebuilt Company Brain workflow builds pages from your public Slack channels, email and codebase, capturing decisions and architecture, and anything from a private source waits for your approval in DoBox first. The Support Reply Composer agent writes support replies from your brain’s documented answers, cites the pages it used, and tells you when the documentation is missing. The Runbook Keeper agent answers “how do we do X?” from the brain and writes the missing runbook whenever the answer had to come from a person. Every write lands on a page with version history, so a bad edit is recoverable.

The Company Brain workflow is an early version of what we’re building toward: one workflow designed to maintain most of the brain, rather than upkeep that happens inside workflows built to automate something else. Contacts already keep themselves current from your messages.

If your team has already settled on Notion or Confluence as the home for docs, our agents can read and write there through MCP instead. HubSpot, Salesforce and any other MCP server connect the same way.

The first 30 days are free if you want to point it at one mailbox and see what it writes.

How we sourced this

Every vendor behavior above comes from that vendor’s own documentation, linked in the sentence that makes the claim, read on September 23, 2026. These pages change, so check the date on the doc before relying on it.

Frequently asked questions

Which knowledge base chat tools let me control which articles the AI uses? Intercom lets you enable or disable content for Fin item by item or in bulk, and target content at audiences built from People and Company attributes. Zendesk scopes by connected knowledge source and by what the customer is entitled to see, so an unauthenticated visitor gets answers from public articles only. Guru can limit a Knowledge Agent to verified content.

How do I stop an AI from answering out of an outdated page? Two fixes take effect right away. Correct the article, in a tool like Zendesk whose AI reads the current version, or remove it from what the AI reads: unpublish it, archive it, or switch it off for the AI specifically where your tool allows that. Review dates and verification intervals mostly create a task for whoever owns the page. Guru is the exception, where an agent set to verified content only stops using a card once its interval lapses.

Can a knowledge base update its own articles automatically? Nothing in the vendor documentation cited here describes an article rewriting itself from customer conversations. What ships is detection and routing, such as a flag from a support agent, a reminder on a date, a ticket for the owner, an unverified badge. In this+that, workflows and agents write back to the brain as they run, and the Company Brain workflow holds anything from a private source for a person to approve.

Do AI answers respect help center permissions? Zendesk’s advanced AI agents do, and say so: restricted articles are available to an authenticated customer and public articles only to an unauthenticated one. Intercom treats help center visibility and Fin usage as separate settings, so verify the behavior in whichever tool you use rather than assuming one implies the other.

What should I check before connecting a knowledge base chat tool to my CRM and support inbox? Check which direction the integration runs. Zendesk and Intercom both read customer context into an answer. Getting a correction back out of a ticket and into the article is what keeps your documentation true, and in the tools covered here that runs through people rather than automation.