An AI agent in Helpdesky is a teammate that drafts or sends replies from your help center articles and data sources, using your own AI provider key. This guide walks you through setting up your first AI agent: adding the key, creating the agent, choosing its identity and mode, and testing it in the Playground before it meets a customer. Once it is running, continue with Train an AI Agent with Skills and a Markdown File and AI Agent Guardrails, Pacing and Hand-off Rules.
AI agents live under Inbox > AI Agents in the dashboard, so messaging has to be enabled on the helpdesk. Teammates need the AI Agents permission in their role to open the page. See Inviting Teammates and Managing Roles.
Add your AI key
Agents run on the AI provider key saved in Settings → AI Settings. Until a key is added, the AI Agents page shows a No AI key configured card with an Open AI Settings button, and agents can be set up but nothing runs.
On the AI Settings card:
- Pick a Provider. Helpdesky supports ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Grok (xAI), DeepSeek, Qwen (Alibaba), Mistral AI and Perplexity.
- Paste the key into the API key field. You create it in your provider's console, for example the OpenAI API keys page or the Anthropic console.
- Choose a Model. Helpdesky lists the latest models available for your key, or falls back to sensible defaults.
- Click Test connection, then Save AI settings.
Every suggestion, draft, Playground run and reply the agent produces is one request to that provider, billed to your account there. Helpdesky adds nothing on top. The same key also powers Ask AI in the widget and replying with AI in the inbox, so you only set it once.
Create the agent
Open Inbox > AI Agents and click New agent. The New AI agent dialog asks for two things:
Personality can be changed later; it only sets the starting values.
- Name: the agent's name, for example Bella. Customers may see this name, so pick one that fits your team.
- Personality: one of three starting points. Human-like reads like a colleague, with natural pauses that grow with reply length, no AI label, no bot intro, and signs with its first name. Transparent (the default) is clearly an assistant: AI label on, introduces itself in its first reply, offers a human when unsure, replies promptly. Instant is the fastest possible: no pause on any channel, no typing bubble, AI label on.
Personality sets the AI label, the first reply intro and how fast the agent replies on live chat and on tickets & email. Click Create agent and the editor opens.
Set the identity
The Identity section controls how the agent presents itself.
Identity and Mode sit at the top of the editor. The Playground on the right uses whatever you have typed, even before saving.
- The avatar button uploads a picture that appears next to the agent's messages in the inbox.
- Name is used in the inbox and, when the sign-off is set to first name, at the end of replies.
- Title, for example Support assistant, is shown next to the name where the agent introduces itself.
- AI label is either Shown or Hidden and decides whether replies are marked as written by an AI. Hiding the label may be restricted where your customers live, and compliance with disclosure laws is your responsibility. The agent still admits it is an AI when asked, whatever you choose here.
- Reply in the customer's language uses the language detected on the customer's message instead of the help center language.
Choose a mode
The Mode section decides how much the agent does on its own. All three modes use the same instructions, skills and guardrails; they only differ in who presses send.
| Mode | What it does | What your team sees in the inbox |
|---|---|---|
| Copilot | Suggests a reply in the inbox. A teammate reviews, edits and sends it. | A Suggest reply button above the reply box. The suggestion appears as a Suggested reply card with Use, Edit and Regenerate. |
| Review | Drafts a reply and waits for one click approval before it goes out. | The draft appears in the thread marked Waiting for your approval, with Send as Bella, Edit and Discard. Nothing reaches the customer until someone approves it, and pacing does not apply. |
| Autopilot | Replies on its own within the pacing and guardrails you set. | A panel reading Bella (AI) is handling this conversation, with Pause and Hand back. Replying yourself pauses the agent. |
Copilot in the inbox. The card reminds the team that a teammate reviews and sends every reply.
Two mode specific options sit below the cards:
- In Copilot, Draft automatically when a conversation is opened prepares the suggestion as soon as someone opens the thread. It is off by default to limit token spend. Drafts are cached per customer message, so reopening a conversation does not bill again.
- In Autopilot, Staff email notifications decides when your team gets the new message email while the agent handles a conversation: Only on hand-off or Every customer message.
Start with Copilot. Your team keeps full control, every edit they make before sending is recorded as a correction, and after a week or two you have real evidence of how the agent would have done on its own. Switching the mode later is a single click and a save.
Write the first instructions
The Instructions section is the agent's voice, persona and standing rules. Click one of the Tone presets (Friendly, Professional, Concise or Warm) to insert starter text at the top of the box, then edit it freely.
A few sentences are enough to start. Knowledge comes from your articles, not from this box.
Describe who the agent is, how it writes and what it should always or never do. For example: "You are Bella, part of the Acme support team. Keep answers short, friendly and specific. Link the relevant help article when one exists. Never guess at order details: ask for the order number instead."
Do not paste your policies or product facts here. The Knowledge section below shows what the agent answers from: your published articles and enabled data sources. Keep facts in articles, where every agent and the widget's Ask AI can use them, and keep the instructions about behaviour.
Sign-off controls how replies end: No sign-off, First name or Custom text.
Test in the Playground
The right hand column of the editor is the Playground. Chat as a customer and the agent answers with your current settings, including unsaved changes, so you can iterate without touching a real conversation.
Each reply comes with a readout: the decision, the confidence and which sources and skills were used.
- Pick a Channel to see the pacing that would apply on live chat versus tickets and email. Copilot suggestions are always instant.
- Type as the customer and send. Under the reply you see Decision (Reply, Hand off or No reply), the confidence as a percentage, the response time, and the articles and skills the agent relied on.
- Try on a recent conversation runs the agent against the latest customer message of a real thread. Nothing is sent or saved on the conversation. This needs the Messages permission.
- Reset clears the chat.
Playground runs show up on the Activity tab alongside inbox suggestions and dry runs, so you can revisit what the agent decided and why.
Save and switch it on
When anything differs from the saved version, a You have unsaved changes bar appears with Save and Discard changes. Save, then go back to Inbox > AI Agents.
The list shows each agent's mode, channels and the last 30 days of results.
New agents are Active as soon as they are created. The Active or paused switch on each row turns an agent off without deleting it. Each row also shows the agent's mode, the channels it works on, its headline numbers for the Last 30 days and an Analytics button.
If you run more than one agent, order matters: when several agents match a conversation, the first one wins. Use Move up and Move down to set the priority. Which conversations an agent matches depends on its channels and availability, which are covered in the guardrails and pacing guide.
Keep an eye on results
Click Analytics on an agent to see what it achieved over 7 days, 30 days or 90 days: Replies sent, Conversations handled, Resolved without a human, Hand-off rate, Thumbs up and more, each with a daily trend.
Analytics for one agent. Switch the period at the top right.
Encourage your team to use the thumbs on suggestions and agent replies, and to edit drafts rather than rewrite them from scratch. Those edits and ratings become corrections, and Helpdesky turns them into suggested changes to the agent's skills and instructions that you approve on the Suggestions tab. How that loop works, and how to give the agent skills and example replies, is the subject of the training guide.