Skip to content
Something
Start

How to build

How to build an AI customer support agent

The problem with most support bots is not that they cannot write — it is that they will happily invent a refund policy. A useful support agent has to be grounded in documents you control, has to know the edge of what it knows, and has to hand over cleanly when it reaches that edge. All three are setup decisions, not model decisions.

Start building free~20 min to first version

What you’re building

A support agent answers customer questions from your own help content rather than from general knowledge. It cites what it used, escalates to a human when confidence is low or the topic is sensitive, and leaves a record of every conversation so you can see what it got wrong.

Who this is a good fit for

  • Teams whose first-line queue is mostly the same twenty questions
  • Anyone with decent documentation that customers never find
  • Support leads who want deflection without the risk of a bot inventing policy

How to build it

  1. Step 1

    Describe the agent and its boundaries

    The boundaries matter more than the persona. Say what it must never attempt — refunds, account changes, anything legal or medical — and what it should do instead. An agent told explicitly to escalate is dramatically safer than one told to be helpful.

  2. Step 2

    Give it your documentation to stand on

    Upload your help articles, or connect the Notion or Confluence space where they live. Connected sources stay current, which matters: a support agent answering from a doc you updated three months ago is a slower version of the same problem you started with.

  3. Step 3

    It builds the retrieval, the guardrails and the handover

    It indexes your documents, builds the answer flow with citations attached, implements the escalation rule you described, and creates the conversation log. The guardrail is structural — the agent retrieves before it answers, so an unanswerable question produces an escalation rather than a confident guess.

  4. Step 4

    Test it on your hardest questions first

    Do not test the easy ones. Ask it the question your newest hire gets wrong, the one where the answer is "it depends", and one it has no business answering at all — the last is the important one. What you are checking is whether it escalates when it should, because that is the behaviour that decides whether you can trust it in front of customers.

  5. Step 5

    Put it live and read the logs weekly

    Publish it to your help centre or embed it in-product. Then actually read the conversation log for the first few weeks — the questions it escalated are a list of gaps in your documentation, and fixing those improves the agent without touching the agent.

What it does once it’s running

  • Answers grounded in your own documents, with the source cited
  • Topics it refuses to attempt, defined by you, enforced structurally
  • Clean handover to a human with the conversation summarised
  • A ticket created in your help desk rather than a dead end
  • Full conversation log, so you can see what it got wrong and why

Tools it connects to

ZendeskIntercomFreshdeskNotionConfluenceSlackGmail

What it won’t do

  • Only as good as the documentation behind it — an agent cannot answer what you never wrote down
  • Actions that change money or account state should stay with a human, and the agent should be told so explicitly
  • Highly regulated advice (medical, legal, financial) needs human review regardless of how well it performs

Start from a template instead

Each of these opens with the prompt already written. Edit it before you build.

Support → Product Loop Agent

Turn support signals into bugs, votes, and user follow-ups.

Documents Agent

Search Drive and Notion, summarize notes, and answer with citations.

Common questions

How do I stop it from making things up?
By grounding it: it retrieves from your documents before answering and cites what it used, so an answer with no source behind it does not get produced. You also name the topics it must refuse outright. Both are set up when you describe the agent, not patched afterwards.
Can it hand over to a real person?
Yes, and that is the behaviour worth testing first. It creates a ticket in your help desk with the conversation summarised, so the human picking it up is not starting from nothing.
What does it do with questions it cannot answer?
It says it cannot answer and offers to escalate, rather than guessing. Those escalations are also the most useful output in the first few weeks — they are a precise list of what your documentation is missing.
Where does it live?
It gets its own URL you can link from your help centre, and it can post into Slack or your help desk. You can also point your own domain at it.

Related use case

AI agent for company knowledge & Q&A

Answer questions grounded in your own documents, with citations, instead of guessing.

Build it in the next ten minutes

Start with a sentence. Leave with a working app on its own URL.

Start building freeSign in