Set up an AI helpdesk by giving one AI Agent the answers it needs. Place it in one customer channel and define how a person takes over. You can build and test that first useful route in one session.
The short answer: add the pages and files your team already uses. Publish the agent in one channel, configure human handoff and replay real customer questions before launch.
The five steps below cover the complete path: knowledge, one customer channel, human handoff, testing and the first review after launch.
1. Give the AI Agent the answers your team already uses
The AI cannot answer a business-specific question from the model alone. Before customers can use it, add the same information your team would check before replying.
Open the AI Agent you want customers to talk to and add these sources to its knowledge:
- Public support and policy pages
- PDF, Word and spreadsheet files
- Product or service information
- Exact answers for details that must be phrased precisely
You do not need to write new documentation first. Start with current webpages, documents and exact answers. Leave out old versions and internal drafts that could contradict what customers should receive.

For a billing-email question, add the page that explains billing details and the current invoice policy. Then ask:
When does a changed billing email appear on an invoice?
Open the sources shown below the answer. The timing should come from the current policy. If an old page appears, remove or update that source instead of trying to hide it with a longer prompt.
With the sources connected, the next question is where customers will meet the AI. Start in one channel so you can see the whole experience clearly.
2. Put the AI where customers already ask questions
Choose one channel for the first launch. Website chat is easiest to observe because you can ask, answer and hand over in the same browser.
Publish the agent on a test page using the website snippet or the installation option for your platform. Then open that page in a private browser window and ask the billing-email question.
Check the customer experience, not only the preview:
- Is the widget easy to find on desktop and mobile?
- Does the answer fit on a small screen?
- Can the customer see what to do next?
- Is the human route visible when the answer is not enough?
Once this works, connect other supported channels such as WhatsApp, email, SMS or Telegram. The knowledge can stay the same, but the reply and handoff must still make sense in each channel.
The AI can now answer a real customer. Before anyone else uses it, make sure the conversation also has a safe ending when the answer is not enough.
3. Set the human route before you go live
Set a human handoff for the agent before customers see it. Turn on live chat when a person should be able to join the website conversation. Choose the team and availability so the promise matches who can actually answer.
In the same handoff settings, choose what happens when nobody can help live:
- Ask for email for a later reply
- Open Form to collect the fields your team needs
- Don't offer follow-up when you only want to show contact details or keep the conversation with AI
Write a short expectation that your team can keep. Do not promise an instant reply outside working hours.

Now ask the full question: "I changed our billing email yesterday, but the next invoice still shows the old address. What should I do?" The AI explains when invoice details normally refresh. When the answer is not enough, the request moves to the chosen team.
Open Inbox. The conversation should include the customer's question, the answer already given and the customer profile built from the details you collected. Click Take over while the customer is still online, or use the offered email follow-up after they leave.

Send the reply from the same conversation. The customer should not have to repeat the billing issue.
The route is complete, but one billing question proves very little. Ten real questions will reveal whether the same setup survives different wording and missing details.
4. Test ten real questions, not ten polished demos
Copy ten recent questions from your inbox. Include the awkward ones:
- Three common questions with clear answers
- Two questions phrased differently from your documentation
- One question with a missing detail
- One question that is outside the agent's scope
- One complaint
- One direct request for a person
- One question that needs private account data
In the test chat, replay the conversation yourself. Save important variations as scripted chatbot tests. You can then run them again after changing a source, instruction or action.
For every test, mark the outcome as:
| Outcome | What to do |
|---|---|
| Correct and complete | Keep it |
| Correct but hard to read | Improve the source or response instruction |
| Wrong source | Remove, replace or retag the source |
| Missing knowledge | Add the smallest useful answer |
| Needs customer data | Plan an AI Action later |
| Needs a person | Check the handoff, owner and reply path |
Do not block launch because one request needs an action. The first version can answer known questions and hand the rest to your team. That already removes repeat replies without hiding support.
The useful questions now work and the difficult ones reach a person. That is enough to launch one channel and learn from real conversations.
5. Go live, then review the first conversations
Publish the agent to the first channel. For the first week, review its conversations, outcomes and handoffs. Then open the questions your AI Agent could not answer. Similar questions are grouped so you can fix a repeated gap once.

The grouped questions show where customers are asking for the same missing answer in different words. Open the matching conversations before adding a source. This keeps the change tied to what customers actually needed.
Use the evidence to make one change at a time:
- Add a missing answer when several people ask the same question
- Fix the product or page when the same confusion causes contact
- Adjust handoff when requests reach the wrong team
- Add a read-only lookup when many answers depend on the same customer data
The billing case is useful even though a person finished it. The AI answered the documented part, the boundary was clear and Finance received the existing conversation. That is better than a confident guess or a dead end.
Add an AI Action only when a repeated task needs it
After launch, you may see many questions that need the same lookup. Connect a ready-made AI Action to your AI Agent or create a custom connection to your system. Start with read-only access.
For the billing example, a lookup could return the saved billing email and the state of the next invoice. Test a normal result, missing customer, existing invoice and failed connection. If a later action changes billing data, put it behind human approval. A failed result may return invoice_already_created, which is a reason to stop and ask a person, not to promise a change.
The first setup does not need this action to go live. Add it because real conversations prove the work repeats, not because a setup checklist says every AI helpdesk needs an integration.
The finished first version
The agent now answers from material your team trusts. Customers can use a real channel. A person can take over with the earlier conversation intact. AI Insights shows what to improve next.
The customer receives a useful answer and reaches Finance when the answer is not enough. The team does not rebuild the case from scratch. That is a complete first version, not a miniature demo.
Next step: If unresolved work must enter another ticket system, use the chatbot-helpdesk integration guide to map the case and reply path. If your team works from Boei Inbox, you do not need that extra integration.
Frequently asked questions
Do I need to write new support documentation first?
No. Connect the material your team already trusts. Review real conversations and improve the sources that actually cause a wrong or incomplete answer.
Do I need technical skills?
You can connect knowledge, website chat and human handoff without building an API. A custom AI Action needs a working endpoint and someone who understands its fields and access.
Should I connect every channel at once?
No. Prove the answer and handoff in one channel, then add another. This makes problems easier to find.
Should AI answer every request?
No. Let it answer supported questions. Give customers a clear human route for sensitive, uncertain or private work.
When should I add AI Actions?
Add an action when many questions need the same live data or repeatable task. Start read-only and require approval before sensitive changes.