Knowledge Base for AI Agents: How to Write Help Articles Your AI Can Answer From
How AI support agents read your help center, seven writing rules for accurate AI answers, where to start, and how to keep the knowledge base current.
By Suhep Team · · 8 min read
An AI support agent does not know your product. It knows your knowledge base. When it gives a customer a wrong refund window or a setup step that no longer exists, the cause is usually an article that was vague, out of date or written for a reader who already knew the answer.
That matters more every month. On 22 October 2026, CX Network reported that support and CX job adverts increasingly ask people to manage AI agents, work that used to sit with IT. For a small team, most of that management is unglamorous: writing and maintaining the articles the AI answers from. This guide covers how an AI agent reads your help content, seven writing rules that make its answers more accurate, and a simple routine for keeping the knowledge base honest after launch.
How an AI agent reads a knowledge base
A person opens an article, scrolls, and fills the gaps from context. An AI agent works differently. Most support AI tools search your articles for the passage that best matches the customer's question and build an answer from that passage. Intercom's guidance for its Fin AI agent describes headings as "semantic chunk boundaries": each section should cover one thing, because a section that mixes several subjects produces an unfocused answer.
Three consequences follow:
- A section is often read on its own. Anything that only makes sense after the paragraph above it can be lost.
- Headings may not survive. Intercom notes that Fin may not always capture headings from HTML, so each section should restate its topic in its opening sentence.
- Contradictions are dangerous. If two articles answer the same question differently, the AI may pick either. Suhep's help article on training your AI lists duplicate articles that answer the same question as a common mistake, because they confuse the AI.
Rule 1: One question per article or section
Give every common customer question one clear home. "How do I change my billing email?" and "How do I download an invoice?" deserve separate sections, even if both live in a Billing article. When two articles overlap, merge them and redirect the old one rather than keeping both just in case.
A good test: could you paste this section alone into a chat window as a complete answer? If not, it is doing too much or too little.
Rule 2: Make every section stand on its own
Pronouns and shorthand that work in a full article break when a passage is read in isolation. Restate the subject instead of writing "it" or "this".
- Weak: "It can be turned off in settings."
- Better: "You can turn off email notifications for new tickets in Settings > Notifications."
Intercom calls a related idea query-answer symmetry: write as if you were giving a radio interview and did not want to be quoted out of context. A bare "No." under a question heading is a poor answer for an AI; a full sentence that repeats the question is a good one.
Rule 3: Write exact numbers, not "a few days"
Vague time frames turn into confident guesses. Intercom's guidance is blunt: numbers, limits and durations should be stated precisely, and phrases like "a few minutes" or "shortly" do not pass. If customers are likely to ask a question that involves a calculation, write the worked example into the article so the AI does not have to do the maths.
For example, instead of "Refunds are processed quickly", a refund article might say: "Once a refund is approved, we send it within 5 business days. Your bank may take up to 7 more business days to show it. If you have not seen it after 12 business days, reply to your ticket and ask for a person." Use your own real figures; the point is that every number is explicit and the total is already added up.
Rule 4: Say who the article applies to
Many wrong AI answers are right for somebody else: a different plan, a different role, a different country. State the audience at the top of the article or section:
- which plans include the feature
- which role or permission is needed (for example, "workspace admins only")
- any region, currency or platform limits
- which contact route applies to which situation, labelled clearly
If the answer differs by audience, write separate sections with clear headings rather than one paragraph full of "unless" clauses.
Rule 5: Use the customer's words
Customers write "I can't log in", not "authentication failure". They write "cancel" when your product says "downgrade to free". Search recent tickets for the phrases customers actually use and put those words in the article title and opening sentence. Suhep's guide to AI confidence scores recommends the same thing: use customer language rather than internal jargon, and keep each article focused on one topic.
When you must use a product term, explain it the first time it appears, and spell out acronyms.
Rule 6: Finish every set of instructions
Step-by-step instructions should run all the way to the end: what the customer sees after the last step, how they know it worked, and what to do if it did not. An article that stops at "Click Save" leaves the AI nothing to say when the customer replies "I clicked Save and nothing happened".
Use numbered lists for steps and bullet points for options. Dense paragraphs are harder for people to scan and harder for an AI to pull a single step out of.
Rule 7: Write down what the AI must not answer
Some questions should never be settled by an AI from an article alone: refund exceptions, legal or compliance questions, account ownership changes, anything involving security, and upset customers who have already been passed around. Keep those topics out of public answers, or write short articles that say plainly what happens next ("A member of our team reviews every refund request outside the standard window"), and make sure your tool routes them to a person. Our AI-to-human handoff checklist covers how to set that boundary.
Where to start: find the gaps first
You do not need to rewrite the whole knowledge base before switching an AI agent on. Start where the volume is:
- Export the last three months of tickets and group them by question. A short list of repeat questions usually covers a large share of volume.
- Check each top question against the knowledge base. Is there one clear, current article? If not, that is your first writing list.
- Mine your saved replies. Macros and canned responses often hold the most tested answers your team has, and Intercom suggests turning suitable ones into articles or snippets. Run them through a macro audit first so you do not publish outdated wording.
- Read the AI's low-confidence and escalated conversations every week. They show you the questions the knowledge base cannot answer yet.
A before-and-after example
Here is a typical article written for people who already know the product, followed by the same content rewritten so an AI agent can answer from any part of it:
- Before: "Exports. Go to Reports and hit the button. Large ones are emailed. Admins only."
- After: "How to export your ticket report as a CSV file. Only workspace admins can export reports. 1. Open Reports. 2. Choose the date range. 3. Click Export CSV. Reports with fewer than 10,000 rows download straight away. Larger reports are emailed to the admin who requested them, usually within 15 minutes. If the email has not arrived after an hour, check your spam folder, then contact support with the date range you chose."
The limits in this example are illustrative; use the ones your product really has. Notice what changed: a title phrased as the task, the audience stated, numbered steps, exact limits, and an ending that covers failure.
Keeping the knowledge base accurate after launch
An AI agent repeats outdated content with the same confidence as current content, so maintenance is part of the job:
- Give every article an owner. Billing articles belong to whoever owns billing; product articles to whoever ships the feature.
- Tie updates to releases. When a feature, price or policy changes, the matching article changes in the same week.
- Sort by last updated. Intercom suggests this as a quick way to find the content most likely to be wrong.
- Watch the signals. Reopened tickets, "not helpful" votes and customers who ask for a person straight after an AI answer all point at weak articles.
If you track resolution targets, remember that an AI answer that gets reopened was not a resolution. Our guide to helpdesk ticket statuses explains why the solved status should only be used when the problem is really fixed.
How this works in Suhep
In Suhep, the AI reads incoming tickets and widget questions, searches your knowledge base and suggests answers, and every AI response carries a confidence score from 0 to 100%. According to Suhep's help center, answers scoring 80% or above are eligible for automatic resolution (80% is the default threshold and can be changed in AI Settings), while anything below 40% is routed straight to a person. Ticket detail pages show which knowledge-base article was matched, and the dashboard's KB Gap Alerts point to topics customers ask about that have no good article yet.
The Test AI panel in AI Settings lets you type a sample customer question, see which article the AI matches and check the confidence score before customers do. To get started, read setting up your knowledge base and AI-assisted resolution in the Suhep help center.
Sources
- Intercom Help: Optimizing content for Fin
- CX Network: AI agent management, the new CX role you can't ignore (22 October 2026)
Frequently asked questions
What is a knowledge base for an AI agent?
It is the set of help articles, FAQs and snippets that an AI support agent searches to answer customer questions. The AI builds its reply from the passage that best matches the question, so the accuracy of its answers depends on how clear, current and complete those articles are.
How should I structure help articles so an AI can answer from them?
Give each common question one home, split articles into single-topic sections with clear headings, restate the subject at the start of each section, use numbered steps and exact numbers, say who the article applies to, and finish every set of instructions with what to do if it does not work.
How many articles do I need before turning on an AI support agent?
There is no fixed number. Start with clear, current articles for the repeat questions that make up most of your ticket volume and route everything else to a person. Then add articles each week based on the questions the AI could not answer confidently.