Guides

How to Build a Self-Service Knowledge Base From Messy Docs

Portrait of Vera Sun

Vera Sun

Last update

Summary

  • A self-service knowledge base is the set of pages customers and an AI agent answer from without a person.

  • The most common reason an AI agent misses is a page nobody wrote.

  • Build it from last month’s questions, give each one a page that answers first, and delete the pages that contradict.

  • Test it with a question it can’t answer, then write each weekly miss once.

Wonderchat’s study of 5,848,078 conversations, recorded between March 2023 and August 2026, took a sample of 5,000 the AI agent couldn’t resolve. In 58.5% of that sample, the answer was never documented anywhere the agent could reach, and only 1.2% hit an out-of-date page.

So build a self-service knowledge base from the questions customers already ask, one answer-first page per question, and retire the pages that contradict. Then connect it to an AI customer service agent and ask it something the docs don’t cover.

Take, as an example, a B2B software company about to put an AI agent on a half-finished help center. Its answer to “How do I reset two-factor authentication for a teammate?” is a Slack thread and an out-of-date PDF.

Why do self-service knowledge bases stop answering?

Most stop answering because nobody wrote the answer down.

The study sorted that sample of 5,000 unresolved conversations by reason. In the sample, never documented came to 58.5%, outside the agent’s scope 12.6%, ambiguous 8.3%, and out of date 1.2%, with 19.3% not classified.

In a random sample of 5,000 unresolved conversations, most asked for something nobody had written down, and 19.3% weren’t classified (shares rounded).

Old pages cause a different failure, a confident wrong answer. The wrong answer Wonderchat’s support team sees most often comes from retrieval pulling an archived policy and answering from it correctly.

Luke Tan, CTO and co-founder of Wonderchat, lists what an AI agent commonly gets wrong. “Common things that are answered wrongly are old policies, old pricing, or outdated documentation.”

A messy knowledge base is a coverage problem first and a cleanup problem second. A perfectly formatted page still can’t answer a question nobody wrote it for.

The work has two jobs, writing what’s missing and removing what’s old.

How do you build a self-service knowledge base?

Start from what customers ask, sort what your team already wrote, then write and test what’s missing. Counting, gathering, and deleting come before any new writing, and they show exactly which pages to write next.

Before you start, have these ready:

  • An export of the last 30 to 90 days of tickets, chats, and emails.

  • Access to every place your team keeps answers.

  • A free Wonderchat account to test the result.

1. Start from last month’s questions

Export last month’s conversations, write each customer’s need in a few words, and count the repeats. Keep the top 20.

Then give each of the 20 one of four labels:

  • Has a page: a current page answers it in full.

  • Partly answered: a page touches it but leaves out a step or a plan limit.

  • No page: the answer sits in a reply, a Slack thread, or one teammate’s memory.

  • Needs a person: the question asks for an action or a judgment no page can give, such as a refund.

In the example, the two-factor question sits near the top and gets the label “No page.” Its only current answer is a Slack thread.

In an example tally, the most asked question with no page is the first page to write.

Count and label the top 20 first, then pick a tool or a category tree. A tree built after the count matches what customers actually ask.

The same count lets you test your help center against your tickets later.

Past tickets are where the questions come from, and once an AI agent runs, the top questions list in Wonderchat’s Analytics keeps the count for you. That list requires at least the Basic plan.

2. Gather what you already wrote into one place

List every place your team keeps answers, then decide what each one becomes:

  • Help center or docs site: connect it by its URL.

  • Google Drive: connect it as a source.

  • PDFs, decks, and spreadsheets: upload them as files.

  • Slack threads and saved replies: rewrite each answer as a page.

An AI agent answers only from what you connect, and a customer can’t open a Slack thread to check an answer. A help center behind a login is a common snag, and uploading its pages as files or connecting Google Drive gets around it.

Wonderchat reads your website and help center, files (PDF, Word, PowerPoint, Excel, CSV, and text), and Google Drive on every plan. Microsoft SharePoint requires at least the Scale plan, with limited uploads.

Each agent stores 100 webpages on the Free plan, 200 on Starter, 1,000 on Basic, and 2,000 on Scale, per Wonderchat’s plans. A site bigger than its plan connects the top 20 pages first, the ones that answer the counted questions.

In Wonderchat’s study, knowledge bases over 150,000 characters resolved about 74% of conversations and those under 10,000 about 61%, an association across whole support operations.

A big base also takes longer to settle. Knowledge bases over 500 pages take a median 15 days to go live, by Wonderchat’s product data.

3. Retire the pages that contradict each other

Keep one current page per answer. Delete or unpublish the old version, and merge near-duplicates into the page customers find first.

Check these pages first:

  • Pricing and plan limits.

  • Refund and cancellation policies.

  • Anything with a date or a version number.

An AI agent that can read both versions may answer from the archived one, and answer it correctly. Deleting the old page is the quickest fix on the whole job.

In the example, with both refund pages live, the agent can answer from the old one, and deleting it leaves one answer.

When an old page has to stay live, keep it out of the index. In Wonderchat, add its URL under Ignore URLs when indexing on the agent’s Data Source tab.

For a whole archive, keep an old folder out of the index with a pattern such as yoursite.com/archive/**, the form the box itself suggests.

Conflicting pages are easy to miss by eye, and Wonderchat’s review inbox keeps conflicts as a group of their own. Luke Tan gives two ways teams catch an old page that answers:

  • “with our review agent (Automatically flags knowledge conflicts)”

  • “by manually looking through their chat sessions”

4. Write one page per question, answer first

Each page answers one question for one kind of customer, with the answer in its first two lines. An AI agent retrieves passages one at a time, and each one has to answer without a “see above.”

The two-factor page

Before

After

Title

Security settings

How do I reset two-factor authentication for a teammate?

First two lines

An overview of every security option

A workspace admin resets it from the teammate’s profile. The teammate sets it up again at the next sign-in.

Steps

One paragraph in a Slack thread

Four numbered steps, with the button names in bold

Conditions

Not stated

Workspace admins only

Date

None, and an old PDF still shows the previous screens

The release the steps were checked against

Luke Tan points to three kinds of page that make an AI agent’s answers worse:

  • “Conflicting or vague pages”

  • “Poorly segmented pages”

  • “Overly long help center pages”

“Pages should be separated into smaller pages where possible” is his fix for the long ones.

5. Write the missing pages, most asked first

Every question labeled “No page” gets a page, in the order of how often customers ask it. A “Needs a person” question gets a short page that says what to send, and a teammate can act on the first reply.

Missing pages are where the gains are, with a caveat. The 84 businesses in Wonderchat’s study that added missing documentation averaged 75.4% resolved before and 80.8% three months later, though 57% improved and 23% got worse.

Across 84 businesses that added missing documentation, the average rose, with 57% improving and 23% getting worse.

“If a question is asked many times, which you can see using our most FAQ feature, then that usually indicates that a dedicated article is helpful” is how Luke Tan decides when a question gets its own page.

For edge cases, he points to a correction instead. “A correction (a small one-off FAQ) is usually good for extremely specific questions or edge cases.”

Write the missing page before you polish an existing one. A question with no page fails every time it’s asked.

In the example, the Slack thread becomes the two-factor page, and it goes first. Each page still missing keeps a repeat question on a teammate’s desk, and AI tools that cut customer service costs can only answer what a page covers.

6. Connect it to an AI agent and ask what it can’t answer

The test takes one sitting in Wonderchat:

  • Sign up on the free plan and create an agent.

  • Paste your help center URL and upload the files behind your most asked questions.

  • Ask the three questions customers ask most.

  • Ask one question your docs don’t cover, on purpose.

The free plan holds 100 webpages and 10,000 characters of files per agent, and Starter holds 200 webpages and 500,000 characters.

If Wonderchat’s helpdesk should run from the first day too, follow the help desk setup in the same account.

Read each reply for three things:

  • The source: which page the answer cites.

  • The gap: which question found nothing.

  • The miss: whether the reply says so and offers a person.

Every answer should show its source. Wanda, Wonderchat’s own AI agent, shows what the miss should look like when asked about a status page for outage alerts.

With no page on status alerts, Wanda says so and points to the mail icon for a person.

An honest miss passes the test, and the question it missed is the next page to write.

Setup rarely holds the test up. Of 478 Wonderchat agents that went live, 384 (80%) did so the same day.

7. Read the misses every week and write each one once

Give the knowledge base one owner and a weekly half hour. The owner reads the week’s misses and writes or links a page for each one.

In Wonderchat, every question the AI agent couldn’t answer lands in Review, sorted into user reported, knowledge gaps, quick fixes, conflicts, and model issues. From a knowledge gap, the owner has three ways to close it:

  • Paste a URL for a page you wrote, and Wonderchat adds the page to the knowledge base if it isn’t there yet.

  • Pick from Files for an answer in a file you already uploaded.

  • Write it myself, then Open editor, for a Create Correction form, a question and answer pair with an optional Source URL.

The app notes that a correction won’t fix the underlying page. Use corrections for edge cases, and write a page for any question customers keep asking.

The two-factor question misses once, waits in Review, gets a correction with its page as the source, and is answered next time.

Changed pages need a recrawl as well. Automatic webpage retraining runs monthly on Starter and Basic and weekly on Scale, and on the Free plan a teammate recrawls by hand.

Without an owner, a knowledge base goes stale one release at a time. The weekly misses name the next page to write.

Should the help desk and the knowledge base share one tool?

They should share one loop, where every handover a teammate answers and no page covers becomes a page. A help desk knowledge base holds the pages the help desk team and the AI agent both answer from.

Wonderchat is an AI customer service agent and AI helpdesk in one. Wonderchat’s helpdesk comes with every plan (tickets that carry the full chat, a shared inbox, and live chat), so the handover and the missing page sit in one account.

Teams can use Wonderchat’s helpdesk from the first day or keep the help desk they run. Ticket sync with Zendesk and Freshdesk requires at least the Basic plan, with up to 50 tickets until Scale, and the AI agent reads their help center by URL.

A customer service knowledge base has to cover more than support questions. Across all 5,848,078 conversations in Wonderchat’s study, support made up 53% and pre-sales 23%.

Pricing and plan pages belong in the knowledge base too.

How do you know your self-service knowledge base is working?

Three numbers show it, each with its base:

  • Answered without a person: the self-service rate, counted against every eligible session, abandoned ones included.

  • Misses that are a missing page: the share of review items that are knowledge gaps.

  • Repeat contacts: customers who come back with the same question.

Count a conversation as answered without a person only when there’s evidence the customer got the answer, such as a thank-you or no repeat contact within a set window. A chat that ended without a ticket can still mean the customer gave up.

Expect a wide spread between teams. Across 687 businesses with enough volume, the weakest tenth resolved 46% of conversations with the AI agent alone.

The strongest tenth resolved 88%, with a median of 69%.

The median business on the same platform resolved 69% of conversations with the AI agent alone.

All 687 businesses ran the same software, and the study doesn’t say why their rates differ. Documentation is one difference a team controls, and in 2026 Wonderchat’s AI agent alone resolved 75.3% of scored conversations.

When is a dedicated knowledge base tool the better buy?

Wonderchat answers from your pages and never drafts them. A team that needs software to write, version, and publish its help center buys a docs tool and connects the published pages by URL.

If your team needs

Use

An editor, versions, and a branded help center

A docs tool, with the AI agent reading its pages by URL

Software that drafts the pages

A docs tool with an AI writer

A help desk it already runs

That help desk, with the AI agent reading its help center

Document360 generates documentation from videos and files with its AI Writing Agent, has its own AI chatbot, and detects duplicate content. A team that wants software to draft and deduplicate the pages starts there.

Helpjuice uses its AI to write docs from Slack threads and tickets. For a team whose answers sit mostly in Slack, that saves rewriting each thread by hand.

Help Scout includes one Docs site on its free plan, capped at 10 articles, and up to five on Pro. Help Scout’s AI assistant, Envoy, costs $0.75 per resolution, billed monthly, on the paid plans.

A team already on Help Scout can keep its pages there, and teams with Zendesk keep Zendesk and point the AI agent at its help center.

CustomGPT also links every answer to its source. What Wonderchat adds is the helpdesk, and a miss and its handover land in one account.

A side-by-side read of AI options for support documentation covers the rest.

Should you start with Wonderchat?

Yes, if your answers sit mostly in replies and threads. Build pages from last month’s questions, and let the AI agent show which ones are still missing.

In the example, Wonderchat answers the two-factor question from the new page and cites it.

A question no page covers gets an honest miss and the offer of a person. The handover reaches the helpdesk with the full chat, and the question waits in Review for a teammate to write the page.

The free plan refills 50 AI replies every month, with no credit card, and includes the helpdesk, a shared inbox, and live chat.

Start for free, paste your help center URL, ask your hardest question and then one your docs can’t answer, and see the source and the honest miss, or book a demo.

Frequently asked questions

How can I create my own knowledge base?

Start from the questions customers asked last month and write one page per question, with the answer in the first two lines. Delete pages that contradict each other, then connect the pages to an AI agent and test it with a question it can’t answer.

Can you give me some examples of knowledge bases?

A public help center and a developer docs site are external knowledge bases, and an internal wiki of procedures is an internal one. Connect the public help center and the docs customers already read to an AI agent first.

Can you give me an example of self-service?

A customer asks the chat widget how to reset two-factor authentication for a teammate and gets the steps, with the help page linked, without writing to support. Resetting a password from a login screen is self-service too.

What is a helpdesk knowledge base?

A helpdesk knowledge base holds the pages a help desk team and its AI agent both answer from, the same pages customers read. With Wonderchat, teams can use its helpdesk or keep their own, and each miss waits in Review until the team writes the page.

How is a self-service knowledge base different from an FAQ page?

An FAQ page lists short answers to common questions on one page. A self-service knowledge base gives each question its own page with steps, conditions, and a date, and an AI agent can retrieve and cite each page on its own.

Stop answering the same questions. Let your docs do it.
Portrait of Vera Sun

Vera Sun

Vera Sun is the co-founder of Wonderchat, an all-in-one AI Support and Conversion agent platform built for companies with large knowledge bases. Her background is in product design and product management, which shapes how she thinks about the messy space between what a customer asks and what a product can answer. She writes about AI agents in production, customer support, and go-to-market for technical buyers, and can usually be found tinkering with new AI tools.