AI Agents

Create your first AI agent

What it's for An AI agent automatically answers your customers for you, in the tone you choose and based on the information you give it. It is the heart of Madyis Hub. Here is how to create your very first agent in a few minutes. Steps to create your agent 1. In the sidebar, click AI Agents. 2. Click New agent. 3. Enter the agent's name (for example "Lea" or "Support Assistant"). This is the name you will see in your dashboard. 4. Add a description. Describe the agent's role in one sentence (for example "answers questions about orders and deliveries"). This description matters: if you use several agents, it is used to route each request to the right agent. 5. Enter the personality (the system prompt). This is where you describe how the agent should speak: its tone, how it addresses the customer, what it represents. Example: "You are the warm assistant of the Flowers & Co shop. You answer politely and always stay positive." 6. Choose the agent's main language. 7. Save. Giving knowledge to your agent An agent with no knowledge knows nothing about your business. Once the agent is created, open it and go to the Knowledge tab to teach it your information (your website, your documents, your text). See the dedicated knowledge base article for details. Activating the agent on an inbox For the agent to actually answer your customers, you must link it to an inbox (for example your website widget, WhatsApp or Instagram). 1. On the agent page, open the Channels tab. 2. Tick the inbox or inboxes the agent should answer on. 3. Save. 💡 Tip — Before activating the agent for real, test it in the Playground to make sure it answers the way you want. See the article "Test your agent (Playground)".

The knowledge base: feeding your AI

What it's for The knowledge base is everything your AI agent knows about your business: your products, your opening hours, your return policy, your usual answers. The more complete and up to date it is, the more accurate the AI's answers are. You manage it in the Knowledge tab of your agent. The 4 ways to add knowledge 1. Import a website Provide your website address. Madyis Hub analyses it automatically and extracts the content. Ideal to quickly pull in your "About", "FAQ" or "Shipping" pages. 1. The Knowledge tab → the Import a website card. 2. Paste the page address. 3. Start the analysis and wait a moment. 2. Import documents Already have a FAQ PDF, a Word product sheet or a spreadsheet? Import it directly. - Accepted formats: PDF, Word, TXT, CSV. - Maximum size: 20 MB per file. 1. The Knowledge tab → the Import documents card. 2. Choose one or more files. 3. Confirm: Madyis Hub extracts the text automatically. 3. Add text For a specific piece of information (opening hours, an instruction, a current promotion), simply type it in by hand. 1. The Knowledge tab → the Add text card. 2. Write your information. 3. Save. 4. Import from Notion If you keep your documentation in Notion, you can sync it into the knowledge base. ℹ️ Coming soon — The Notion import is built but still requires the Notion connection to be finalised on the platform side. To prepare it, first connect Notion in Settings → Integrations. Rest assured: the AI follows your information For every answer, the AI reads your entire knowledge base (up to about 60,000 characters). It does not rely on a single snippet: it genuinely builds on what you have written. In practice: - It does not make things up: it answers from your information. - When you add new knowledge, it is taken into account immediately. - If your base becomes very large, the AI automatically selects the passages most relevant to the question asked. 💡 Tip — Information that must always be respected (a tone, a strict rule, your opening hours) is also best placed in the agent's personality or guidelines, since those are sent to the AI on every answer.

Scenarios

What it's for A scenario is a named instruction that your agent applies depending on the customer's situation. Instead of mixing everything into the personality, you define clear procedures that the AI follows at the right moment. Example: a "Refund request" scenario whose instruction is "Ask for the order number, then check the return policy before answering". When a customer mentions a refund, the AI applies exactly this procedure. Create a scenario 1. Open your agent in AI Agents. 2. Go to the Scenarios tab. 3. Click to add a scenario. 4. Give it a clear title (for example "Refund request"). 5. Write the instruction: what the AI should do in this situation, step by step. 6. Save. Enable or disable a scenario Each scenario can be enabled or disabled without being deleted. Handy, for example, to enable an "End-of-year promo" scenario only during that period. 1. In the Scenarios tab, find the scenario you want. 2. Use the toggle to enable or disable it. 3. A disabled scenario is no longer applied by the AI but is kept for later. 💡 Tip — Write short, concrete instructions. A "ask X, then do Y" instruction is followed better than a long paragraph.

Guardrails and response guidelines

What it's for Guardrails and response guidelines are strict rules that your AI agent always respects. This is what makes the AI reliable and safe for your business: you stay in control of what it can and cannot say. You define them in the agent's Settings tab. Guardrails: the prohibitions A guardrail tells the AI what it must never do. It is a safety barrier. Examples of guardrails: - "Never promise a refund." - "Never give medical advice." - "Never make up a price: if you don't know it, say you will check." Response guidelines: style and obligations A response guideline tells the AI how to answer: its style, its phrasing, its obligations. Examples of guidelines: - "Always greet the customer by their first name." - "Always end with: Have a great day!" - "Answer in two sentences maximum." How to set them up 1. Open your agent in AI Agents. 2. Go to the Settings tab. 3. In the Guardrails field, write your prohibitions, one per line. 4. In the Response guidelines field, write your style obligations, one per line. 5. Save. 💡 Tip — Guardrails and guidelines are sent to the AI on every answer. So this is the ideal place for a rule that must always apply, whatever the customer's question.

Availability: make the AI answer according to your hours

What it's for You don't have to let the AI answer all the time. You can decide when it takes over, according to your hours. The AI knows the time and whether your inbox is open or closed. This setting is in the agent's Settings tab, under Availability (scheduling). Prerequisite: set up your business hours For scheduling to work, the agent must know your hours. You define them on the inbox. 1. Go to Settings → Inboxes. 2. Select the relevant inbox. 3. Open Business hours and fill in your open time slots. 4. Save. The 3 availability modes Always active (24/7) The AI answers at all times, day and night, weekends included. Only outside business hours The AI covers evenings and weekends, when your team is not around. During your business hours, it lets your human agents answer. Ideal to never leave a customer without a reply while keeping humans in charge during the day. Only during business hours The AI answers only when you are open. Outside those hours, it steps back. How to choose the mode 1. Open your agent in AI Agents. 2. Go to the Settings tab. 3. Under Availability (scheduling), choose the mode you want. 4. Save. 💡 Tip — Combine availability with a guardrail. For example, in "outside business hours" mode, add a guardrail: "Outside hours, only say that we are closed and give our opening hours." Your customers are reassured without the AI committing on your behalf.

A team of AI agents (multi-agent)

What it's for Rather than a single agent that has to know everything, you can build a real team of AI agents. A "lead" agent (the supervisor) receives each request, hands it to the right specialist (support, billing, appointment booking, etc.), then reviews the answer before it goes out to the customer. Your customers feel like they are talking to an expert team, not a generalist robot. How it works - The supervisor reads the customer's request and picks the best-suited specialist, based on each agent's description. - The specialist drafts an answer from its knowledge base. - The supervisor reviews and refines the answer before sending. Setting up your team 1. Create several agents (see "Create your first AI agent"), one per specialty. Give each a precise description of its field: this is what lets the supervisor route requests correctly. 2. For each agent, open the Settings tab and set its role: specialist or supervisor. 3. Designate one agent as the supervisor; the others stay specialists. 4. Link the supervisor and all specialists to the same inbox: in each agent, the Channels tab, tick the same inbox. Write good descriptions Routing relies entirely on each specialist's description. Be clear: "Handles refunds, returns and delivery issues" is far more effective than "Support". 💡 Tip — Start simple: a supervisor plus two or three specialists is often enough. You can add more later. Test the whole team in the Playground before going live.

Supervised mode: approve answers before sending

What it's for You may want to keep an eye on what the AI says before it answers your customers. That is exactly the role of supervised mode: the AI prepares the answer, but you decide whether to send it. The mode is set in the agent's Settings tab. The two modes Supervised mode The AI does not answer the customer directly. It creates a draft that you find on the AI Validation page. You can approve it as is, edit it before sending, or reject it. Ideal to start with confidence or for sensitive topics. Autonomous mode The AI sends its answer directly to the customer, with no approval step. Ideal once you trust your agent and want to reply instantly, 24/7. Choosing the mode 1. Open your agent in AI Agents. 2. Go to the Settings tab. 3. Choose supervised or autonomous. 4. Save. Approving drafts 1. In the sidebar, open AI Validation. 2. You see the queue of pending answers. 3. For each draft: edit the text if needed, then approve to send it to the customer, or reject to discard it. 💡 Tip — Many merchants start in supervised mode to check the quality of answers, then switch to autonomous mode once reassured. You can change the mode at any time.

Test your agent (Playground)

What it's for The Playground is a private testing space. You chat with your agent just like a customer would, but with no risk: no real customer sees these messages. This is the step to take before activating the agent for real, to make sure it answers well. How to test your agent 1. Open your agent in AI Agents. 2. Go to the Playground tab. 3. Type a question the way a customer would (for example "What are your opening hours?" or "How do I return a product?"). 4. Read the AI's answer and judge whether it is correct, in the right tone, and faithful to your information. What to check - Does the AI know the right information? If it hesitates, complete its knowledge base. - Is the tone right for you? If not, adjust the personality or response guidelines. - Does it respect your guardrails? Deliberately ask a tricky question to verify. - Do your scenarios trigger at the right moment? The winning loop Test, adjust, test again. Edit the knowledge, personality or guidelines, then come back to the Playground to check the effect. Once satisfied, activate the agent on an inbox (the Channels tab). 💡 Tip — Above all, test the hard or rare questions: the simple cases almost always pass; it is the edge cases that reveal what needs improving.

Human handover and alerts

What it's for The AI is efficient, but some situations deserve a human: an unhappy customer, a sensitive request, a complaint. Madyis Hub lets you define when the AI should step aside in favour of a team member, and who to notify. Everything is set in the agent's Rules & Alerts tab. Human handover via keywords You define keywords that, as soon as they appear in a customer's message, pause the AI on that conversation. A human can then take over without the AI continuing to answer. Example keywords: "lawyer", "refund", "cancel", "talk to a human". 1. Open your agent in AI Agents. 2. Go to the Rules & Alerts tab. 3. Add a handover rule and enter your keywords. 4. Save. Alerts: notify the right agent An alert does not necessarily pause the AI: it notifies a specific agent so they can take a look. Handy to make sure an important topic doesn't slip past anyone. 1. Still in the Rules & Alerts tab, add an alert. 2. Set the trigger condition. 3. Choose the agent to notify. 4. Save. The chosen agent will receive a notification (bell, email and real-time notification). Pausing the AI on a single conversation You can also take over manually on any conversation, with no keyword. 1. Open the relevant conversation. 2. In the header, use the AI on / AI paused toggle (or the "Take over" option in the actions menu). 3. The AI stops answering on that conversation; turn it back on whenever you want. 💡 Tip — Combine both: an "urgent" keyword that pauses the AI, plus an alert to your manager. You never miss an important request.

Import your help center into your AI agent

Already have a help center? Reuse it to feed your AI agent in one click — it'll answer from your published articles. Import 1. Open your AI agent → Knowledge tab. 2. Choose Import from the Help Center. 3. Select the portal and the language. 4. Run the import: each published article becomes a knowledge document. Good to know - Only published articles are imported. - When you edit your articles, re-run the import: it refreshes the knowledge (without creating duplicates). - That's the strength of the system: the same content serves your customers in self-service AND your AI — you write it once.

The AI action log: understand replies (and non-replies)

The action log shows, conversation by conversation, what your AI agent did — and why. It's the tool to understand a reply… or the lack of one. Where to find it Open your AI agent → Log tab. What it contains For each generated reply: the model used, the response time, and the knowledge documents used. The log also records sent replies, skipped replies (monthly quota reached, outside the agent's hours…) and any failures. Debugging "why didn't the AI reply" - Quota reached? → the AI skips replying until next month. - Agent's schedule? → check its Availability (24/7, after-hours, or business hours). - A human already replied? → the AI steps back so it doesn't talk over you. - Agent not connected to the inbox? → check the agent's Channels tab. Good to know: the log also shows whether the AI sent the whole knowledge base (small base) or only the relevant passages (large base) — useful to understand its answers.