Customer Support

How to build a customer knowledge base with AI

Customer Support3 AI tools7 steps6 friction points

Want an agent to handle this on your live data? See the Starch version →

A customer knowledge base is the collection of answers, policies, procedures, and product details your team actually uses to respond to customer questions — plus the organized, searchable version you give customers so they can help themselves. Building one means gathering that information from wherever it currently lives (your head, Slack threads, a folder of Google Docs, old support tickets), structuring it, and maintaining it as things change. Most operators have been meaning to do this for months.

AI makes this feel tractable because the hard parts — turning rough notes into polished articles, organizing a pile of unstructured Q&A into categories, rewriting technical documentation for a general audience — are exactly what language models are good at. You don't need a technical writer or a documentation team. You paste in messy source material and get structured, readable output in a few minutes. That's a real and legitimate use of these tools.

ChatGPT, Claude, and Gemini can all contribute meaningfully here today. You can use them to draft articles from raw notes, generate FAQ structures from support email threads, rewrite dense internal docs into customer-facing language, and suggest category taxonomies for organizing your content. The work is real; so is the ceiling.

Skip the manual work → See the Starch version

Customer Support3 AI tools7 steps6 friction points

AI walkthrough

How to do it with AI today

A practical walkthrough using ChatGPT, Claude, and other off-the-shelf LLMs — what they're good at, what you'll have to do by hand.

Tools that work for this

ChatGPTClaudeGemini

Step-by-step

1 Gather your raw source material: pull your most common support emails, copy your existing policy docs, paste in Slack threads where you or your team have answered customer questions. This is your input set — the messier the better at this stage.

2 Open Claude or ChatGPT and paste in a batch of support emails or Q&A threads. Ask it to extract the distinct customer questions and your answers, then format them as draft FAQ entries with a question, a short answer, and a one-paragraph explanation.

3 Take your existing policy docs (return policy, shipping policy, account setup instructions) and paste them one at a time into the LLM. Ask it to rewrite each one in plain language, in second person, at a seventh-grade reading level. Edit the output to match your actual policy before publishing.

4 Ask the LLM to look at all the FAQ entries you've generated and suggest a category structure — five to eight top-level categories with descriptions of what belongs in each. Treat this as a starting point; adjust based on how your customers actually ask questions.

5 For each category, use ChatGPT or Claude to draft a longer-form help article from your notes. Paste in the rough source material and prompt for a structured article with a title, intro paragraph, numbered steps or bullet points, and a closing summary.

6 Copy your drafts into whatever knowledge base tool you're using — Notion, Confluence, a help center platform like Intercom or Zendesk. The LLM produces text; you handle placement, linking, and publishing manually.

7 Set a recurring reminder to revisit the knowledge base when your policies change. Return to the LLM with updated notes and re-run the relevant prompts to refresh affected articles.

Prompts you can copy

Here are 20 customer support emails from the past month. Extract every distinct question customers asked, group duplicates, and write a draft FAQ entry for each — question, one-sentence answer, and a short paragraph with detail.

Rewrite the following return policy in plain English, second person, for customers who aren't familiar with our product. Keep every rule intact but remove jargon. Policy: [paste text]

I'm building a knowledge base for a small e-commerce business. Here are 30 FAQ entries across different topics. Suggest 6-8 top-level categories, explain what belongs in each, and sort these entries into the right categories.

Turn the following rough notes into a help center article. Include a clear title, a one-paragraph intro explaining what the article covers, step-by-step instructions, and a short FAQ at the end. Notes: [paste notes]

Here is a knowledge base article we wrote 6 months ago. Here are the changes we made to our policy since then. Rewrite the article to reflect the changes, keeping the same structure and tone. Old article: [paste] Policy changes: [paste]

Reality check

Where this gets hard

The walkthrough above works — until your numbers change, the LLM hallucinates, or you have to re-paste everything next month.

No connection to your actual support tickets or email — you manually copy-paste threads each time, so you're always working from a snapshot, never the full picture.

Nothing persists between sessions. The structure, tone guidelines, and category taxonomy you carefully developed last month aren't available next time you open a new chat.

Output consistency drifts. The article format and voice you got from one prompt chain won't automatically match what you get when you draft new articles three weeks later with a slightly different prompt.

The LLM can't tell you which articles are stale. When your return policy changes, nothing flags the six knowledge base articles that reference the old policy — you have to remember to find them yourself.

There's no searchable layer. The LLM produces text documents; making them findable for customers or your support team requires a separate tool, manual tagging, and ongoing maintenance the LLM can't help with.

Large volumes of source material hit context limits. If you want to analyze a full year of support tickets to find knowledge gaps, you're chunking manually and losing cross-document coherence.

Tired of the friction?

Starch runs the whole workflow on live data — no copy-paste, no hallucinated numbers, no re-prompting next month.

See the Starch version →

Starch alternative

The same workflow on Starch

Starch is an agentic operating system — an agent builds and runs the software your work depends on, connected to your live business data. For this workflow, that means a persistent knowledge management system that stays current as your product and policies change, not a document you re-draft manually every few months.

Connect Notion from Starch's integration catalog and sync your existing docs on a schedule — Starch reads your current pages and structures so the knowledge base starts from what you already have, not a blank slate.

Connect Gmail or Outlook from Starch's integration catalog; Starch pulls your support email threads so the agent can identify recurring questions, surface gaps in your current documentation, and draft new articles without you manually copying emails.

Use the Knowledge Management starter app to get a team wiki with AI-powered search out of the box — everything your customers and team need to find answers in one searchable place, without scattered Google Docs.

Describe what you want in plain English and the agent builds it: 'Build me a knowledge base dashboard that shows which articles were searched most this week, which searches returned no results, and flags any article that hasn't been updated in 90 days.'

Customer Support Agent — coming soon — will use your knowledge base as its source of truth to resolve common tickets instantly across chat and email, escalating only the questions your documentation can't answer, with full context attached.

Automations run on a schedule without you re-running prompts: every Monday, Starch can scan your connected support inbox for new question patterns and Slack you a list of topics your knowledge base doesn't yet cover.

Get closed-beta access →

Toolkit