AI knowledge base for customer support: how it works and how to build one
By Roman Zakharenko · Facts checked
An AI knowledge base for customer support is a help center written so that both customers and AI agents can answer questions from it. The AI looks up the relevant articles and builds its reply from them, so its answers are only as accurate, complete and current as your articles. You can build it in the tool you already have: Zendesk's help center, Freshdesk's knowledge base or Confluence spaces linked to Jira Service Management.
At a glance
- Retrieval-augmented generation finds relevant passages in your documents and adds them to the prompt, so the model answers from your content.
- In a Zendesk help center, articles live in sections inside categories; they can't sit directly in a category.
- Freshdesk organizes its knowledge base into categories, folders and articles.
- Confluence spaces linked to a Jira Service Management knowledge base feed its help center, knowledge base search and virtual service agent.
- Jira Service Management counts a request as deflected when a customer picks a suggested article while submitting and marks it helpful.
How AI answers draw on your articles
Most AI support agents use some form of retrieval-augmented generation (RAG). AWS describes RAG as a technique that augments a large language model with external data, such as a company's internal documents. In outline:
- Your documents are converted into embeddings — numeric representations of the text — and stored in a vector database. This is a one-time setup step.
- A customer asks a question in natural language.
- The system runs a similarity search on the vector database and adds the passages it finds to the prompt as context.
- The language model writes its answer using that context.
Two things follow. The AI can't answer from what isn't written down, and it works with pieces of articles rather than your whole help center. Freshdesk's guidance makes the same point: Freddy AI retrieves and uses articles independently, so if an article leaves out definitions or context, the AI may make assumptions or give partial answers.
The major helpdesks build this in. Zendesk's AI agents can use generative replies on messaging and email, connected to multiple knowledge sources, with search rules that control which sources are searched under specific conditions. In Jira Service Management, articles from a linked Confluence space show up in the help center, in knowledge base search and in the virtual service agent.
Where the knowledge base lives in your helpdesk
| Helpdesk | Where articles live | How they're organized | What draws on them |
|---|---|---|---|
| Zendesk | Help center | Categories contain sections, and sections hold articles; sections can be nested, but articles can't sit directly in a category | Customers browsing and searching, and AI agents with generative replies |
| Freshdesk | Knowledge base (solution articles) | Categories contain folders, and folders hold articles; folders can be nested | Customers on the portal, and Freddy AI |
| Jira Service Management | Confluence spaces linked to the service project's knowledge base | Pages in the linked Confluence space | The help center, knowledge base search and the virtual service agent |
One practical detail: in Freshdesk, a category or folder appears on the customer portal only after at least one article in it is published. For AI answers, what each article says matters more than where it sits; the hierarchy mainly helps people who browse.
Write articles an AI can answer from
Freshdesk publishes guidance on writing knowledge content for Freddy AI, and it applies to any retrieval-based AI:
- Keep each article to a single topic, and say when the information applies and when it doesn't.
- Link to related articles instead of embedding secondary workflows.
- Split setup, usage and troubleshooting into separate articles when they serve different goals.
- Make every article self-contained: define the terms and context it relies on.
- Use one top-level heading for the article's purpose, and lower-level headings to separate tasks, options or concepts.
Add a few habits of your own. Title articles with the words customers actually use, not internal product names. Put the answer in the first lines, then the steps. Keep exact values — limits, deadlines, prices — in one place and link to it, so a change doesn't leave two different numbers for the AI to choose between. And start where the tickets are: a few weeks of tickets grouped by topic shows which articles to write or repair first.
Keep the content current
An outdated article turns into a confident, wrong AI answer, so give every article an owner and a review rhythm. Zendesk supports this with article verification: knowledge admins set verification rules that flag articles for review based on filters and intervals — for example, every three months for articles with a specific label, in a certain category, or not updated in a while — and article owners get email reminders to verify them. Check whether your plan includes it before you count on it.
Scheduled reviews catch slow drift. These events should trigger an update straight away:
- A product, pricing or policy change: update the articles on or before release day.
- An agent answers a ticket the knowledge base doesn't cover: create or flag an article.
- The AI hands a conversation to a person because it couldn't answer: review that question.
- Two articles give different answers: merge them and unpublish the duplicate.
Measure self-service
Measure whether articles prevent tickets, not just whether people read them. Keep the definitions the same every month so the trend means something.
| Measure | What it tells you | How to get it |
|---|---|---|
| Deflected requests | Customers who found an answer while submitting a request | Jira Service Management counts a request as deflected when the customer selects a suggested article during submission and marks it helpful |
| Requests resolved with an article | How often agents close requests using knowledge base content | Jira Service Management's Requests resolved report tracks requests closed with or without a knowledge base article |
| Self-service ratio | Help center use per ticket | Divide help center sessions by tickets created in the same period, and watch the trend |
| AI handoff rate | Share of AI conversations passed to a person | Your AI agent's analytics, or a tag added on every handover |
| Tickets on documented topics | Whether existing articles actually prevent tickets | Sample tickets each month and check whether an article already covered the question |
How TascaLab puts your knowledge base to work
TascaLab builds chatbots on top of the tools you already use — Zendesk, Freshdesk, Jira, Slack or internal systems — through APIs, webhooks and Make or n8n, so your knowledge base stays where it is. The Starter Chat Bot ($400 one-time) covers 1–2 channels and the Standard Chat Bot ($575) covers 3–5, across your website, phone, WhatsApp, Telegram, Facebook Messenger and Instagram. Upkeep workflows, such as review reminders for stale articles, fit Custom Automation (from $1,200).
Most setups take 1–3 weeks. Every project includes 30 days of post-launch support and basic documentation, with no hidden fees, and you can start with a free 30-minute Health Check call.
Frequently asked questions
What is retrieval-augmented generation in customer support?
It's a way for an AI to answer from your own content. Relevant passages are retrieved from your documents, typically through a similarity search on a vector database, and added to the prompt, so the language model writes its answer from that context.
Can Jira Service Management use Confluence as its knowledge base?
Yes. You link a Confluence space to the service project's knowledge base, and its articles appear in the help center, in knowledge base search and in the virtual service agent.
How do I keep Zendesk knowledge base articles up to date?
Use article verification if your plan includes it. Verification rules flag articles for review based on filters and intervals, and article owners get email reminders to verify them.
How should articles be structured for AI answers?
One topic per article, self-contained, with clear headings. Freshdesk notes that Freddy AI retrieves articles independently, so missing definitions or context can lead to assumptions or partial answers.
Want this set up for your team?
Book a free Health Check: a 30-minute call and a written report with 2–3 concrete recommendations, a fixed price and estimated running costs. No pitch.
Sources
- AWS Prescriptive Guidance — Understanding Retrieval Augmented Generation
- Zendesk Help Center — Organizing knowledge base content in categories and sections
- Zendesk Help Center — Generative replies in AI agents FAQs
- Zendesk Help Center — About article verification and how it works
- Freshdesk Support — Create categories, folders and articles in your knowledge base
- Freshdesk Support — Write Freddy AI–optimised knowledge content
- Atlassian Support — Link a Confluence space to your knowledge base
- Atlassian Support — Find out how your knowledge base articles are performing