Can an AI agent keep your help center up to date?

Can an AI agent do knowledge base maintenance?

Yes, as a drafter and auditor with a person approving every change. An AI agent for knowledge base upkeep can find the questions customers ask that your help center does not answer, draft the missing articles, propose edits when your product or policy changes, and flag duplicate or contradictory articles. It does not decide what is true, and in the two products documented here its changes wait for a person to accept them.

The evidence comes from the vendors. Intercom's launch post says Operator takes a brief, such as release notes or a policy change, finds every article that needs updating and drafts the edits, and VentureBeat reported roughly 200 customers in its beta in May 2026. Intercom's documentation says Fin Operator "never publishes, modifies, or deletes content directly"; every change is a proposal someone approves. Zendesk saves articles drafted from ticket data as drafts for admins to review. Intercom reports that its own knowledge team ships 10–15 product releases a week, which involves creating and updating around 90 articles. The team uses Operator to audit the knowledge base and propose updates, which it says it can review and publish for each release in less than an hour; a significant or widespread release may take 2–3 hours. The same article lists about 30 hours a week for writing and updating product-release content. That is the vendor describing itself. No independent accuracy study was found.

Treat this as a supervised hire. The agent does the searching and drafting; you keep the approval.

This page covers keeping help content current. For answering customers see tier-1 support, for routing tickets see ticket triage, and for the role as a whole see customer service representative.

What the agent does and what stays with you

Agent

2of 9

• Find customer questions your help center does not answer
• Flag articles that look outdated or underperforming

Agent with review

5of 9

• Draft new articles from resolved tickets and conversations
• Update every affected article after a product or policy change
• Find and resolve duplicate or contradictory content
• Restructure articles so AI agents can use them
• Translate articles where demand exists

Human

2of 9

• Decide what is true and approve what gets published
• Own the content standards, article owners and style guide

The split follows one rule: finding problems is cheap to check, so the agent does it; changing what customers read needs a person's approval. Gap lists and freshness flags change nothing on their own. Intercom's recommendations come from comparing failed Fin answers with human replies to similar questions, and from checking for duplicates and contradictions across sources. Zendesk's admin copilot compares recent tickets with your articles and gives weekly insights on coverage, freshness and AI readability when needed. Intercom shows the source conversations for each recommendation; Zendesk's recommendations can include related tickets or other supporting insights, and its documentation does not say every recommendation has them.

Drafts and edits are different. An article reaches every customer who reads it and every AI agent that answers from it. One Intercom employee wrote on the company's community forum in May 2025 that they review the suggestions once a week and approve "about 70%". Read that the other way round: about 3 in 10 suggestions were not approved, in the vendor's own workspace.

A worked example: one policy change

Assume, for this example, a software company that extends its refund window from 14 to 30 days, with a help center of 200 articles.

The operations lead gives the agent the new policy text and asks it to update every article that mentions refunds. (Intercom's Operator FAQ gives a similar example prompt.) The agent searches the help center and finds six articles: the refund policy itself, two billing FAQs, a cancellation guide, a plan comparison page and an internal macro article. It proposes an edit to each as a diff: "14 days" becomes "30 days" in five places. In the plan comparison it also adds a sentence saying refunds apply to annual plans, citing a support reply from last month.

The knowledge owner reviews the six proposals on Thursday. Five are correct. The added sentence about annual plans is not in the policy text; it came from one agent's reply. She rejects that edit, sends the question to finance, and approves the other five. The billing owner signs off because refunds are on the sensitive list. Before publishing she searches the help center for "14 days" and finds no remaining matches.

What you need to provide

The agent reads customer conversations, so it touches personal data: names, emails, order details and whatever customers write. It needs:

• Edit access to the knowledge base through a role that can draft but not publish or delete. In Intercom, accepting content recommendations requires a specific publishing permission; give that only to your human approvers.
• Read access to solved tickets and AI agent conversations, including the human replies. Zendesk's knowledge builder uses public comments only and excludes internal notes; Intercom's recommendations skip conversations without a teammate response.
• Release notes and policy changes as they happen. Intercom's guide recommends asking product managers to share details of every product in build.
• A style guide and an owner for each sensitive topic.
• Enough volume. Zendesk says a small number of qualifying tickets might not produce a proposed article, and Intercom says low-volume customers may get few or no recommendations.

Our recommendation, not a requirement of any tool: name one person who approves everything, plus a topic owner for billing, refunds, security and legal terms.

On cost, the published prices are bundled into support platforms rather than charged per article. Intercom's Pro add-on, which includes content recommendations and Operator, costs $0.20 per customer conversation with a 500-conversation monthly minimum, so at least $100 a month ($0.20 × 500). Operator usage above the plan's monthly credit allowance is charged at $0.01 per credit. Zendesk includes admin copilot at no additional cost on Suite Professional and above; Suite Professional is listed at $115 per agent per month, paid yearly. Neither price covers the reviewer's time, which in Intercom's own summary is about 1 hour a week for AI content suggestions. Because the units are per conversation and per seat, not per article, this page does not compare them with a wage.

Job brief template

Use this as the brief for a vendor, an operator or your own team setting up the agent. Replace the bracketed values with your own.

Keep the 2-business-day update rule and the 20-proposal weekly sample. They measure the job: is the help center current, and is what the agent proposes correct.

Risks and human checkpoints

A confident, wrong article can do more harm than a missing one, because your AI support agent may repeat it to customers who ask. The defense is the evidence behind each proposal. Intercom attaches the source conversations to each recommendation, and Zendesk's recommendations can include related tickets or other supporting insights; read them before approving, and reject anything whose only support is a single reply.

Another failure is quiet: duplicates. Zendesk says you can generate a help center from ticket data when you have no help center yet, or one that is incomplete or out of date, and says a well-maintained help center "will likely not be helpful" because the tool creates new and possibly duplicate sections and categories. For ongoing upkeep of a maintained help center, our recommendation is to use recommendation-style tools instead of bulk generation.

Check your tool's limits too. Intercom documents that Operator analyzes at most 250 conversations per session, cannot add new translations to existing articles, and saves new articles as drafts disabled for Fin until someone publishes them.

How to run a trial

Run the agent for 4 weeks on your existing help center before you rely on it.

1. Pick a slice: one product area or one collection, with a named approver.
2. Week 1, baseline: let the agent produce its gap and contradiction list without changing anything. Check 20 items against the linked conversations. Count how many are real.
3. Feed it one real release note or policy change. Time how long it takes to propose edits to every affected article, then search the help center for the old wording to see if it missed any.
4. Each week, review 20 proposals against the acceptance criteria. Log factual errors, unsupported claims, personal data and duplicates.
5. Track your reviewer's time and the acceptance rate. A falling acceptance rate may indicate that proposals are becoming less useful; inspect the linked evidence and inputs.
6. Watch repeat contacts on topics whose articles changed. If they drop, the edits may have helped; check for other changes before attributing the decline.

1. 1

   Pick a slice

   One product area or one collection, with a named approver

2. 2

   Week 1, baseline

   The agent produces its gap and contradiction list without changing anything; check 20 items against the linked conversations

3. 3

   Feed it one real release note or policy change

   Time how long it takes to propose edits, then search for old wording it missed

4. 4

   Each week, review 20 proposals

   Against the acceptance criteria; log factual errors, unsupported claims, personal data and duplicates

5. 5

   Track reviewer time and acceptance rate

   A falling acceptance rate may mean proposals are becoming less useful

6. 6

   Watch repeat contacts on changed topics

   If they drop, the edits may have helped; check for other changes first

If the criteria hold for two weeks in a row, widen the slice. For structuring the test, see how to run a paid trial with an AI agent; for the brief, how to write a job description for an AI agent; and for the weekly review, how to supervise an AI agent. A current help center also matters for phone support. More support jobs are on the customer support overview.

Sources

1. 1
   Use AI-powered content recommendations to improve Fin

   IntercomPrimaryAccessed Nov 14, 2026

2. 2
   Fin Operator FAQs

   IntercomPrimaryAccessed Nov 14, 2026

3. 3
   Creating help center content using ticket data and generative AI

   ZendeskPrimaryAccessed Nov 14, 2026

4. 4
   Optimizing knowledge with admin copilot insights and recommendations

   ZendeskPrimaryAccessed Nov 14, 2026

5. 5
   Announcing admin copilot for Suite Professional and above accounts

   ZendeskPrimaryAccessed Nov 14, 2026

6. 6
   Mastering knowledge management for great AI support

   IntercomVendorAccessed Nov 14, 2026

7. 7
   How to upgrade Fin's knowledge base over coffee

   Intercom CommunityVendorAccessed Nov 14, 2026

8. 8
   Meet Operator: An Agent for your customer operations

   IntercomVendorAccessed Nov 14, 2026

9. 9
   Intercom, now called Fin, launches an AI agent whose only job is managing another AI agent

   VentureBeatPressAccessed Nov 14, 2026

10. 10
    Pro add-on

    IntercomVendorAccessed Nov 14, 2026

11. 11
    Zendesk Pricing Plans

    ZendeskVendorAccessed Nov 14, 2026