Your AI support agent answers from your help centre. How do you get the articles ready?
A method for preparing help centre articles before an AI support agent answers from them: coverage, conflicts, structure, scope, retrieval tests and owners.

The short answer
Get the help centre ready for an AI support agent before the build. Map articles to the questions customers ask, remove conflicting and stale versions, rewrite each article so one section answers one question, mark what the agent may quote, test retrieval on real questions and name an owner for every article.
Key takeaways
- An AI support agent answers from whatever its knowledge source says, so a wrong, stale or conflicting help article becomes a wrong answer given with confidence.
- The readiness audit has six checks: coverage of real customer questions, conflicts, staleness, structure, scope of what the agent may quote, and a named owner per article.
- Coverage is measured against the questions customers actually send, taken from recent tickets and chats, not against the article list the team already has.
- Retrieval should be tested on its own before the agent is tested, because an agent that finds the wrong article will write a fluent answer from it.
- In Gartner's survey of 3,566 B2B and B2C customers in February and March 2026, 87% said a company using GenAI for service must offer a way to reach a human agent, so the help centre also needs a clear route to a person.
- After launch, every question the agent could not answer is a content task with an owner, and the help centre becomes the agent's change log as much as the customer's reference.
In this article
Why the help centre decides what a support agent says
The help centre decides what an AI support agent says because the agent retrieves an article and writes its answer from it. If the article is right, the answer usually is. If the article is wrong, out of date or contradicted by another one, the agent repeats the error in fluent, confident prose.
Most support agents work this way. A customer asks a question, the system searches the knowledge source for the closest passages, and the model writes a reply from them. The model does not know your refund policy changed in March unless an article says so. It cannot tell which of two articles is current unless something marks it.
A disclosure: Sigzen AI builds AI agents for clients at published fixed prices, and a content audit is part of every support agent build we scope.7 We have an interest in agents going live. This method is the audit we run first, written so your team can run it without us.
How do you check whether a help centre is ready?
Check a help centre for readiness with six tests: coverage, conflicts, staleness, structure, scope and ownership. Each test produces a list of articles to write, merge, update, rewrite, restrict or assign, and the lists together are the work plan before the build.
Run the tests in that order. Coverage comes first because it tells you which articles matter. There is no point rewriting an article no customer needs while the top question has no article at all. Conflicts and staleness come next because they cause the most damaging answers: confident and wrong.
Customers already expect AI in service, and they compare yours with the alternatives. In Gartner's survey of 3,566 B2B and B2C customers, run in February and March 2026, people were about three times as likely to use a third-party GenAI tool, such as ChatGPT, as a company's own chatbot to sort out a service problem.1 Those outside tools read your public help centre too, so the audit improves both.
AI in customer service, 2026
Customers use AI for service, and they still want a person within reach
Map articles to the questions customers actually ask
Map coverage against the questions customers actually send, not against the article list you already have. Pull recent tickets and chat transcripts, group them by intent, and check each intent against the help centre. The gaps are the questions with many contacts and no article.
Take a recent quarter of contacts, or at least a few hundred conversations if volume is low. Tag each one with its intent in plain words: "change delivery address", "return a damaged item", "update card details". Count the contacts per intent and sort the list from most to least frequent.
Then mark each intent with one of four labels. Covered: an article answers it fully. Partial: an article mentions it but leaves out a step or a condition. Missing: no article exists. Action: the customer wants something done, such as a refund, which an article alone cannot complete.
| Label | What it means | What to do before launch |
|---|---|---|
| Covered | An article answers the intent fully | Check it for conflicts and dates |
| Partial | A step, condition or exception is missing | Add the missing part to the article |
| Missing | No article answers it | Write one, or keep the intent with people |
| Action | The customer wants something done | Scope a tool or a hand-over, not an article |
Illustrative. Our labels for a coverage map; the intents come from your own tickets and chats.
Our view: launch the agent only on intents marked Covered after the audit, and route the rest to people until their articles exist. A smaller agent that answers well earns more trust than a broad one that guesses.
Why do conflicting articles produce wrong answers?
Conflicting articles produce wrong answers because the agent may retrieve either version, and it has no reliable way to know which one is current. A person on the support team knows the old returns article is wrong; the agent sees two plausible sources and picks one.
Conflicts build up quietly. A policy changes, a new article is written, and the old one stays published because a customer email links to it. A product is renamed and both names live on. A regional exception is written into one article and missing from the next.
Find them by searching the help centre for every policy term in your top intents: refund, return window, delivery time, cancellation, warranty. Read every article that mentions each term side by side. Any two that state different numbers, conditions or steps are a conflict, and one of them must be merged, corrected or removed.
Others have covered this ground, and the credit belongs to them. Zendesk's guide to optimising help centre content for generative AI explains how articles are split into chunks and matched to messages, and it advises removing duplicate and conflicting information.5 What this method adds is the order of work and the tests that prove each fix.
Stale articles: give every article a date and an expiry
Treat every article as stale until someone confirms it, and give each one a review date and an expiry. An article with no date is the hardest kind of error to catch, because it reads as current to the agent and to the customer.
Add three fields to every article if your help centre allows it: last checked, checked by, and next review. Where the platform has no custom fields, put a short line at the foot of the article. The point is that a reviewer can see at a glance which articles nobody has looked at for a year.
Prioritise by risk, not by age alone. An article on prices, fees, refunds, legal rights or safety does the most harm when wrong, so it gets the shortest review cycle. An article on how to change a password can wait longer. Intercom's guide to preparing a knowledge base for AI bots also advises checking last-updated dates and fixing high-traffic articles first.6
Our view: any article that states a price, a time limit or a legal condition should name the date it took effect, in the text. The agent will quote the date with the rule, and a customer reading an old answer can see that it is old.
How should an article be written so an agent can use it?
Write each article so that one section answers one question on its own, with the answer in the first sentence and every condition spelled out. Retrieval systems usually pull passages, not whole articles, so a section that depends on the paragraph above it loses its meaning.
Five habits cover most of the rewrite. Put the question in the heading, in the customer's words. Answer it in the first sentence of the section. Name the subject again instead of writing "it" or "this". State conditions as conditions: "if the order has shipped", not "in some cases". And write out what a screenshot or video shows, because the agent cannot see either.
Tables need care. A short table with clear headers reads well, but a wide table with merged cells can be split badly when it is indexed. Where a table holds the rule itself, such as fees by plan, add one sentence per row in plain text beneath it so the rule survives being split.
The same answer, written for a person and for an agent
Written for a person
- Heading: Returns
- "It depends on the item, see below"
- Conditions in a screenshot of the form
- "In most cases we can help"
Written for an agent and a person
- Heading: How do I return a damaged item?
- First sentence states the return window
- Each condition is a sentence with "if"
- Effective date and route to a person
Public articles, internal notes and what the agent may quote
Decide, article by article, what the agent may quote to customers, what it may read but not quote, and what it must never see. Support teams keep internal macros, workaround notes and escalation guides next to public articles, and an agent that reads all of them may quote any of them.
Sort every source into three groups. Public: articles any customer may read, which the agent may quote. Internal guidance: notes that help the agent decide, such as "check the order status before offering a refund", which it may follow but not repeat. Restricted: anything with personal data, security steps, pricing exceptions or negotiation limits, which stays out of the agent's index entirely.
The internal group is where most leaks happen. A macro that says "offer a discount if the customer threatens to cancel" is useful to a person and harmful when an agent repeats it to every customer who asks. Our view: an internal note the agent needs should be rewritten as a rule in its instructions, and the original note kept out of its index.
Write down the sorting in a short register: source, group, owner and reason. When the agent says something it should not have, the register tells you within minutes whether the content was in scope or the configuration was wrong. Our method for rewriting a procedure so an agent can follow it covers turning internal guidance into rules.
Test retrieval before you test the agent
Test retrieval on its own before you test the agent's answers, because an agent that finds the wrong article will still write a fluent answer from it. A retrieval test asks one thing: for each real question, did the system find the right passage?
Build the test from the coverage map. For each Covered intent, take three to five real customer questions in their original wording, typos included. Next to each, write the article and section that should answer it. Run the questions through the platform's search or retrieval preview, if it has one, and record the top passages returned.
Score each question as found, found low or missed. Found means the right section is the top result. Found low means it appears but below another passage. Missed means it is not returned. A missed question points to a wording problem: the article uses the company's terms and the customer uses different ones.
Fix by changing the article, not the test. Add the customer's words to the heading or the first sentence, split a long article in two, or merge two thin ones. Then rerun. Our guide to testing a support agent before switching it on takes over from here, with the full answer test, the pass marks and the rollout.

Does the help centre need a route to a person?
The help centre needs a clear, public route to a person, and the agent needs to know it. Customers expect one, and every question the agent cannot answer has to go somewhere better than a dead end.
In Gartner's survey of 3,566 B2B and B2C customers in February and March 2026, 87% said it is essential that a company using GenAI for service offers a way to reach a human agent.2 Write one article that states how to reach a person, through which channels and when, and make it the agent's fallback answer for anything out of scope.
The same article belongs in the coverage map as an intent of its own, because "how do I talk to someone" is usually near the top of the list. Our view: an agent that hides the route to a person to raise its resolution rate damages trust faster than it saves cost. Our escalation design for support agents covers when the agent should hand over without being asked.
Who keeps the articles right after launch?
A named owner keeps each group of articles right after launch, and every question the agent could not answer becomes a task for that owner. Without owners, the help centre decays at its old pace while the agent repeats it to far more customers.
Assign owners by topic, not by article: billing, delivery, returns, accounts. The owner approves every change in that topic and answers for its accuracy. Give the support team a simple way to flag a wrong or missing article from inside a conversation, and route each flag to the topic owner.
Review the agent's gaps weekly. Pull the conversations it handed over or could not answer, group them by intent and decide for each group: a new article, an edit, a new rule in the instructions, or a confirmed hand-over. Large firms are building the same habit around agents. In KPMG's survey of 314 US leaders at firms with $1 billion or more in revenue, run from 24 July to 25 August 2026, 74% included cost reviews in their AI approval processes.4
A four-week help centre readiness plan
Pull a quarter of tickets and chats; tag intents; build the coverage map; name topic owners.
Find and resolve conflicts on policy terms; add review dates to every article.
Rewrite top-intent articles section by section; sort sources into public, internal and restricted.
Run the retrieval test; fix missed questions in the articles; agree the launch intents.
What the readiness work costs, and how to know it paid off
The readiness work costs the time of the people who know the products and the policies, and it pays off when the agent answers the launch intents correctly without a person stepping in. Measure both from the start.
The spending pressure is real. In a Gartner survey of 1,303 senior leaders run from January to April 2026, service and support teams put a median 12% of their 2025 budget into AI, the most of the ten functions Gartner assessed, and 24% of service and support leaders showed a positive financial return across their AI use cases.3 Content readiness is one of the cheaper ways to avoid joining the other side of that figure.
Track four numbers: the share of retrieval tests found at the top, the share of agent answers a reviewer marks wrong, the hand-over rate per intent, and the number of open content tasks. A falling wrong-answer rate with a stable hand-over rate means the articles are improving. A falling hand-over rate with a rising wrong-answer rate means something was loosened that should not have been.
What to do this month
Run the six tests on the help centre this month, before any agent configuration, and agree the launch intents from the results. None of it needs a new tool.
- Export a quarter of tickets and chats, tag each with an intent and count them.
- Label every intent Covered, Partial, Missing or Action against the current articles.
- Search the policy terms in your top intents and resolve every conflict you find.
- Add last checked, checked by and next review to every article, shortest cycle for prices and rights.
- Rewrite top-intent articles so each section answers one question in its first sentence.
- Sort every source into public, internal and restricted, and keep the register.
- Run the retrieval test on real questions and fix misses in the articles, then name topic owners.
The order matters. A team that prepares the content first launches an agent that repeats the business's current policy, and a team that skips it launches an agent that repeats whatever was published last. Our view: the audit is worth doing even if the agent is delayed, because customers and outside assistants read the same articles today.
Sources
- Gartner, survey of 3,566 B2B and B2C customers, Feb–Mar 2026: third-party GenAI vs company chatbots; geography not stated (Jul 2026)
- Gartner, survey of 3,566 B2B and B2C customers, Feb–Mar 2026: access to a human agent; geography not stated (Aug 2026)
- Gartner, survey of 1,303 senior leaders, Jan–Apr 2026: service and support AI budget share and financial return (Jul 2026)
- KPMG, AI Quarterly Pulse Q3 2026: 314 US leaders at $1bn+ firms, 24 Jul–25 Aug 2026 (Sep 2026)
- Zendesk, Optimizing your knowledge content for generative AI (help article, accessed Nov 2026)
- Intercom, Get ready for AI bots by optimizing your knowledge base (Apr 2023)
- Sigzen AI, published prices (Nov 2026)


