Build an Intercom Knowledge Base That Powers AI

Before Intercom turned Fin on for its own support team, the team reviewed and updated more than 700 help center articles. Intercom documented this themselves: the model was ready, the content wasn't. After the rewrite, resolution rates climbed in a way the model alone could not have produced.

Knowledge base prep is the step that moves the needle. Teams that invest two to four weeks of cleanup before launching Fin see resolution rates roughly 12 percentage points higher than teams that deploy with the existing help center untouched. Article quality is the usual bottleneck, and the work is mechanical enough to plan around.

This piece is the prep work that gets your Intercom knowledge base for AI ready before launch. It's mechanical, it's a few weeks, and it's the highest-leverage thing you can do.

TL;DR

• Resolution rate is downstream of knowledge base quality. Skipping prep caps the ceiling around the low 30s.
• Audit before rewriting. Rank articles by Fin involvement, ticket volume, and view count, not alphabetically.
• Rewrite the top 20% of articles for AI readers. Plain language, restated questions, explicit version info.
• Treat retrieval as the design constraint. One question per article. Headings that match real customer phrasing.
• Validate with red-team queries before launch. Maintain with Fin's Content Gap Suggestions after.

Why the knowledge base sets the automation ceiling

Fin is a retrieval and reranking pipeline with an LLM on top. It pulls passages from your knowledge sources, picks the relevant ones, then composes a grounded answer. If the right passage isn't in your knowledge base, or it's there but written in marketing voice, or it answers a different question than the one the customer asked, the AI either guesses or hands off.

Industry benchmark data puts median tier-1 deflection at 41.2%, with the top quartile at 58.7%. The gap between median and top quartile is rarely about which AI agent the team picked. It's about what the agent gets to read.

You can confirm this on your own data. Pick five questions Fin failed on last week. Find the help center article that should have answered each one. Four out of five times, the answer is technically there, written in a way the model couldn't match to the question or extract cleanly.

That's the work. TicketSwap, for example, automates 67% of conversations across 19 million users, and the leverage comes from sustained work on the help content the AI reads, structured deliberately for retrieval.

Step 1: Audit what you have

Run the audit data-first. Pull two lists from Intercom:

1. The top 100 help center articles by view count, last 90 days.
2. The top 100 conversation topics by ticket volume, same window.

The overlap is your priority queue. Articles that match high-volume topics get rewritten first. Articles that nobody reads, even when the topic is hot, get either rewritten urgently or removed and the content folded somewhere it will be found.

What to log in the audit:

• Article URL and title
• Views in the last 90 days
• Ticket volume on the topic (rough estimate is fine)
• Current Fin involvement rate if Fin is live
• Resolution rate per article if Fin is live and reports it
• One-line note on what's wrong (vague answer, missing version info, screenshot-dependent, written in marketing voice, etc.)

Intercom's product gives you Content Gap Suggestions inside Fin. The feature surfaces conversations Fin couldn't resolve and ranks where articles are missing or need updating, ordered by likely impact. If Fin is already running, that's your starting point. If Fin isn't live yet, synthesize the same signal from the manual audit.

By the end of step 1 you should have a ranked list of 80 to 120 articles to work through. About three weeks of dedicated work for one person, or one week for a small team.

Step 2: Rewrite the high-traffic articles for AI readers

Help center content written for humans browsing a search results page reads very differently from content written for an AI extracting an answer to a specific question.

Humans are good at stitching context across paragraphs and screenshots. The AI is reading one passage at a time. If the article shows different UIs in screenshots for self-serve and enterprise users, the human reader infers which case applies. The AI cannot see the screenshots and the inference rarely lands.

What "rewrite for AI" looks like in practice:

• State the question at the top of the article in plain language. "How do I update my billing address?" beats "Account preferences and billing management."
• One question per article. If the article currently covers three related questions, split it. Fin retrieves at the article level. A focused article gets pulled cleanly; a mixed one gets the wrong passages pulled or none at all.
• Replace screenshots with text where possible. If a screenshot is essential, describe what's in it in the surrounding text so the AI has the same information a sighted reader has.
• Cut marketing voice. "Our award-winning billing experience" is noise that confuses retrieval. State the policy.
• Use the customer's words. If customers say "refund" and your docs say "settlement adjustment," the retrieval scoring will struggle. Match the phrasing.

Intercom themselves recommend running content through Claude or ChatGPT to optimize phrasing, remove ambiguity, and align with common customer language. The point isn't that LLMs write better than humans. They strip out the cleverness humans add. For AI-readable content, plain is the right register.

A before-and-after

Before:

Welcome to our billing experience! Our streamlined dashboard makes managing your account easier than ever. To access billing preferences, navigate to the account menu and select the relevant option. From there, you'll be able to update payment methods, billing addresses, and tax information.

After:

How do I update my billing address?

To update your billing address:

1. Click your account avatar in the top right.
2. Select Billing.
3. In the Billing address section, click Edit.
4. Enter your new address and click Save.

Changes apply to invoices issued after the update. Past invoices keep the address that was active when they were issued.

The second version is two thirds the length and answers the question three times: in the heading, in the steps, and in the edge-case note. Retrieval handles it cleanly.

Step 3: Add context, versions, and edge cases

The most common Fin failure mode after launch is the confident wrong answer. The AI quotes a policy that used to be true, or applies a default scenario to a customer who's on a different plan.

Three additions fix most of these.

Version stamps. Date the article. "This article applies to plans purchased after January 2025." If your product has tiered behavior, say which tier each section refers to. The model cannot infer that a 2023 article describes the old refund window unless the article says so.

Explicit scope. State who the article applies to in the first paragraph. "This applies to admins, not regular users." "This is for the EU region; US customers should see [other article]." The AI respects these boundaries if they are stated. Implicit boundaries get smoothed over.

Edge cases. Customers ask about the edges. "What if I'm in a trial?" "What if my plan changed mid-month?" Anchor the edge-case answer in your content. Otherwise the AI improvises, and improvisation is where most reputation damage from AI support originates.

A useful test: after rewriting an article, ask "what's the closest wrong question someone could ask that this article looks like it might answer?" Then add a sentence redirecting that case.

Step 4: Restructure for retrieval

Retrieval pipelines work better on chunk-friendly content. Structure matters as much as wording.

The practical rules:

• Short paragraphs. Two to four sentences each. Retrieval chunks documents into passages roughly that size.
• Descriptive headings. H2s and H3s that name what the section answers. "Refunds for cancelled subscriptions" beats "Refunds." Headings are weighted heavily in retrieval scoring.
• Restated questions inside the article. If the article answers "Can I get a refund after 30 days?", say so explicitly in the body, even if the title already does. It anchors the passage during reranking.
• Don't bury the answer. The first paragraph after the heading should contain the direct answer, then the context. If the answer is in paragraph four, retrieval may pull paragraphs one through three instead.

Intercom's product team frames this as "treating the Knowledge Hub like you're optimizing for AI search". The substitution is useful. Anywhere you would have SEO-optimized for a search engine, do the equivalent for the AI.

Step 5: Validate before launch

Before turning the AI loose on customers, put it through a red-team pass.

Build a list of 30 to 50 real customer questions from the last 90 days of tickets. Include the easy ones, the edge cases, and the questions the team usually struggles with. Run them through Fin (or your dedicated AI agent) in a sandbox or with a test instance.

For each question, score:

• Did the AI answer correctly?
• Did it cite the right article?
• Did it confidently quote a wrong policy (the dangerous case)?
• Did it escalate when it should have?

Track the answers in a spreadsheet. Every wrong answer is either a knowledge base gap, a content quality issue, or a place where the AI should defer to a human. Fix each one before launch. Post-launch fixing is more expensive because there is a real customer on the other end.

A useful benchmark: aim for 80%+ correct answers on your red-team set before launch. Teams that hit 80% in the sandbox typically land 50–65% in production once real customer phrasing variation kicks in.

Maintaining the knowledge base after launch

AI knowledge bases are not set-and-forget. The maintenance loop:

Weekly. Pull the worst-performing 10 conversations from the past week. Read them. Each one is either a knowledge base fix, a workflow fix, or a "this should be human" tag.

Monthly. Review Fin's Content Gap Suggestions (or the equivalent in whatever AI agent you run). Add the missing articles. Update the flagged ones.

Quarterly. Revisit the audit list. Topics shift. Volume distributions change. The top 20 articles a year from now will not be the top 20 today.

Teams running on open.cx spend roughly three to five hours a week on this loop and see the resolution rate climb one to two percentage points a month for the first six months after launch. The gain is almost entirely from knowledge base work.

What changes when the knowledge base is right

A competent team with a clean knowledge base hits 50–65% resolution in the first 90 days. The top performers climb past 70% by month six. The mechanism is unglamorous content work, focused on the top 20% of articles and repeated weekly. Article quality and retrieval-friendly structure carry the program. Pick the AI vendor that fits your stack and ambition, and let the content do the rest.

If your AI agent has been plateauing, look at the top 50 articles before you look at the vendor. Almost every time, that's where the next 15 points of resolution rate are sitting.