All guides
Guide

How to build a docs support agent for your SaaS product, step by step

Your help center already has the answers. Here is how to turn it into an agent that answers customers from your docs, links the right article, and tells you which questions to write up next, with screenshots of every step.

How to build a docs support agent for your SaaS product, step by step

Most support inboxes at small SaaS companies hold the same twenty questions. How do I export? Why did my sync fail? Can I change the invoice number? The answers are already in your help center. Customers just don't read it.

A docs support agent reads it for them. It answers from your help articles, links the page it used, and tells you which questions your docs couldn't answer. This guide builds one on AgentFront from a blank account to a customer getting an answer. Every screenshot comes from a real run.

For the example we use Acme Invoice, a made-up invoicing app for freelancers, with seven help articles. Swap in your own product and docs as you go. Expect to spend about fifteen minutes.

What you need

  • Your help articles. A public docs site, a sitemap, or the files themselves (PDF, DOCX, Markdown, TXT or CSV) all work.
  • A rough idea of who asks the questions: your product's users, your trial users, or both.
  • A Stripe account, only if you plan to charge for the agent. You can add it later.

Part 1: Build the agent

1. Create your account

Go to app.agentfront.me/signup and enter your name, email and a password. There is no model key to set up; new accounts come with AI credits to build and test with.

The AgentFront sign-up form

2. Describe the agent

Give the agent a name. The name becomes its web address, so you'll see straight away whether it's free.

Then describe the job in plain words. Say which product it supports, what customers ask about, and what it should do when the docs don't cover a question. This description feeds every later step, so a specific paragraph here saves editing later.

Step 1: the agent's name, its address, and a description of the job

Our description:

A support agent for Acme Invoice, an invoicing app for freelancers and small agencies. It answers customers' how-do-I questions from our help docs (sending invoices, recurring billing, payment reminders, taxes, payouts, exporting to accounting tools), links the exact help article, and flags questions the docs can't answer so we know what to write next.

3. Answer the setup questions

AgentFront asks two or three short questions that fill gaps in your description. For a docs agent the important one is what happens when the docs fall short. We picked Professional and concise for tone and Collect details for follow-up for missing answers, so the agent admits the gap instead of guessing.

Two setup questions: tone, and how to handle missing answers

4. Add your docs

This is the step that matters most. The agent answers only from what you add here, and every answer cites it.

  • URL: paste your help center address, or a sitemap.xml to pull in up to 50 pages at once.
  • Files: upload exported articles. We uploaded seven Markdown files, one per help article.
  • MCP: connect a live source if your docs sit behind an API.

Wait until every source shows Indexed before you continue.

Step 2: seven help articles uploaded and indexed, with the URL option open

Two habits make the answers better:

  1. Put the article's public URL at the top of each file. We started every file with a line like Help article: https://help.acme-invoice.example/invoices/recurring. The agent then links customers to the real page instead of a file name.
  2. One article per file. Long exports that mix ten topics make it harder for the agent to find and cite the right one.

5. Skip connectors for now

Connectors let an agent act in other tools: look up an order, create a ticket, update a record. A docs agent only needs to read, so click Skip for now. If you later want it to hand hard questions to your help desk, add a connector or your own MCP server from the dashboard.

Step 3: connectors, skipped for a docs-only agent

6. Set pricing

AgentFront suggests plans from your description. Customers pay monthly for credits, and each message uses credits at its real model cost. A short docs answer costs well under one credit; ours ranged from 0.35 to 0.6.

Step 4: three suggested plans with credits per month

How you price depends on who the agent is for:

  • For your own product's users, keep one free plan with enough credits for normal use and treat the agent as part of your support.
  • As a product you sell, for example a docs agent you run for other companies, keep paid tiers and connect Stripe on the next screen. AgentFront takes no cut of what you earn.

One lesson from our run: don't give your agent's plans the same names as your product's plans. Acme Invoice has a Starter plan, and so did our agent. Later a customer on the agent's Starter plan asked a question, and the agent told them a feature wasn't "on your current Starter plan". It mixed up the two. Call the agent's plans something like "Free" and "Team" instead.

7. Pick a look

Choose a template and a color scheme for the storefront, the site where customers find the agent and chat with it. We used Modernist in Blue. You can change both later.

Step 5: storefront templates and color schemes

8. Review the system prompt

AgentFront writes the agent's instructions from everything you've entered: tone, sources, plans and limits. Read it through. It already tells the agent to rely only on your sources and to say when they don't cover a question.

We changed one line so every answer ends with a link to the help article:

Each source starts with its help article URL; end every answer with a link to the article you used.

Step 6: the generated system prompt, ready to edit

Click Create storefront.

9. You're live

In under a minute the agent has a landing page, sign-up, pricing page and chat at its own address.

The storefront is ready, with its live address

From the dashboard you can edit any part of the page by clicking it and describing the change. The landing page already offers sample questions based on your description, and customers can type their own.

The dashboard's storefront editor

Part 2: What your customers see

Here is the same agent from a customer's side. Share the storefront link in your app's help menu, your docs site and your support auto-reply.

Ask from the landing page

A customer can type a question straight into the landing page.

A customer types a question on the storefront landing page

Sign up in one step

They're asked to create an account so their conversations are saved, and the question they typed is carried through.

The customer sign-up form

Get an answer with sources

The answer arrives in a few seconds as numbered steps, with links to the two help articles it came from. Note the plan check: the docs say recurring billing needs Pro or Agency, so the agent says so first.

The agent answers with steps and links the help articles it used

When the docs don't know

Then we asked something our docs didn't cover: can invoices be sent from your own email domain? The agent didn't invent a feature. It said the docs don't cover it and offered the closest thing they do describe, downloading the PDF and sending it yourself.

A question the docs can't answer, handled honestly

This is the behavior you want from a support agent. A confident wrong answer about your product costs more than "I don't know".

Returning customers land on their own chat home, with sample questions and their history in the sidebar.

The customer's chat home

Part 3: Find and fill the gaps

The agent's second job is telling you what to write next.

Read the conversations

Every conversation shows up under Conversations in the dashboard, with the customer, the credits used and each step the agent took, including which knowledge searches it ran.

The creator's conversation list with a full transcript open

Check what went unanswered

About thirty minutes after a conversation goes quiet, AgentFront reviews it. It sorts conversations into topics, marks whether the customer got what they needed, and lists the questions behind them. Questions the agent couldn't resolve land under Unanswered.

Topics, top questions, and the email-domain question under Unanswered

For a SaaS team this list is the useful part. It is your docs backlog, ranked by how often people ask.

Write the article, then add it

Write the missing help article and publish it on your docs site. Then add it to the agent under Agent → Knowledge: upload the file, or add the page's URL.

Use Test retrieval to check the agent will find it. Type the question the way a customer would and see which sources come back first. A new source can take a minute to appear in search after it shows Indexed.

The new article added, and Test retrieval returning it first

Ask again

We asked the same question in different words. This time the agent gave the steps from the new article and linked it.

The same question now answered from the new article

Keep it useful

  • Add an article for every unanswered question that comes up twice. Check the list weekly.
  • Re-add articles when the product changes. Remove the old source and add the new one, or the agent will keep quoting the old steps.
  • Read a few transcripts each week, even good ones. You'll spot answers that are correct but confusing, which usually means the article is too.
  • Keep the scope narrow. A docs agent that only answers from your docs is easy to trust. Add connectors when you're ready for it to take actions, not before.

Your help center already has the answers. Create your agent and point it at them.

af
AgentFront team