AI assistants
An assistant talks with your customers or your team over web chat or WhatsApp and resolves end to end: it answers from your documents, looks up a catalog, books an appointment, runs a robot to fetch or record data and, when needed, asks a person for approval or hands the conversation to an agent. Everything is measured in cost and protected in personal data.
Assistants live under Centro › IA › Asistentes (Center › AI › Assistants; see The menu). Like the rest of the Center, they are organized in folders.
Creating an assistant
Section titled “Creating an assistant”- Under Centro › IA › Asistentes, click Nuevo workflow (New workflow).
- Pick a template. Each one comes with instructions, rules and tests ready for its industry: Asistente en blanco (Blank assistant), Agente inmobiliario (Real estate agent), Citas de consultorio o clínica (Clinic appointments), Soporte al cliente (Customer support), Calificación de clientes potenciales (Lead qualification) and Reservas de restaurante o local (Restaurant or venue bookings). To draw the conversation step by step instead, pick Flujo en blanco (Blank flow).
- Name it and, if you use folders, pick the Carpeta (Folder).
The Studio opens, where you configure, test and publish.
From each assistant’s ⋮ menu you can also Probar (Test), see its Conversaciones (Conversations), Duplicar (Duplicate: copies the configuration and the flow; channels are connected again on the copy), Mover a otra carpeta (Move to another folder) or Eliminar (Delete).
Two ways of working
Section titled “Two ways of working”| Mode | When it fits | How it is defined |
|---|---|---|
| Agente autónomo (Autonomous agent) | Open conversations where the model decides the next step. | The Instructions and the Capabilities you turn on. |
| Flujo guiado (Guided flow) | Journeys with fixed steps and full control of the order. | A canvas with nodes; the Instructions give context to the AI nodes. |
The guided-flow canvas has AI nodes for Mensaje (Message), Clasificar (Classify), Extraer (Extract) and Recolectar datos (Collect data), plus Agente (hands the conversation to an agent with tools until a goal is met), Condición (Condition), Proceso RPA (runs a published robot, with success and error outputs), Procesos en paralelo (Parallel processes), Bifurcar and Unir (Fork and Join), Esperar msg (Wait for message), Escalar (Escalate), Aprobación (Approval: creates a task with its options and its own target and deadline), Plantilla WhatsApp (sends an approved template, also after 24 hours) and Fin (End).
Capabilities
Section titled “Capabilities”Capabilities are turned on under Configuración and define what the assistant can do:
- Knowledge base: answers from your documents and cites the source. See Knowledge bases.
- Catalog lookup: tabular data from a spreadsheet, an API or a robot.
- Send files: from Drive, OneDrive or Zoan Cloud storage.
- Book appointments: in the Microsoft 365 or Google calendar.
- RPA processes: the robot runs on the machine and session you choose, and the assistant resumes the conversation when it finishes.
- API actions: operations of an API imported in Integrations. See API actions.
- Ask for approval: creates a task in the Tareas inbox and continues when someone decides.
- Escalate to an agent, Offer options (a list on WhatsApp, buttons on web chat) and End conversation.
Confirmation before acting
Section titled “Confirmation before acting”API actions that write are added with Pedir confirmación antes de ejecutar (Ask for confirmation before running) checked: the assistant shows what it’s about to do and waits for the person’s yes. Read operations aren’t. RPA processes don’t ask for confirmation by default: check Pedir confirmación (Ask for confirmation) on each process, or turn it on for all of them under Privacidad y seguridad with Pedir confirmación al usuario antes de cualquier acción con efecto (ejecutar un proceso) (Ask the user for confirmation before any action with an effect). It’s worth it for processes that record or change data; for those that only look things up, the conversation flows without extra steps.
The model
Section titled “The model”Also pick the model. If your plan includes AI, you don’t need your own account: spend comes out of the plan budget. If you prefer your own Anthropic, OpenAI or other provider account, connect it under Centro › Recursos › Integraciones. The assistant uses the integrations in its folder, those linked to it and those at the environment root (see Integrations in folders). See AI usage.
Channels
Section titled “Channels”- Web chat: with the assistant published, click Emitir clave (Issue key) and paste the Snippet de integración (integration snippet) into your site, before
</body>. - WhatsApp · Mosend: click Conectar con Mosend (Connect with Mosend), pick the Mosend account (from Integrations) and the number the assistant will serve. Zoan Cloud creates the webhook for you. Set up a template to resume conversations after 24 hours. See Mosend (WhatsApp).
Voice notes and audio
Section titled “Voice notes and audio”With Transcribir notas de voz y audios (Transcribe voice notes and audio), in the Conocimiento (Knowledge) tab of the configuration, WhatsApp voice notes and web chat audio are turned into text before the assistant answers: people can talk instead of typing. It comes on, with Idioma del audio (Audio language) set to Spanish; you can pick English, Portuguese, French or Detectar automáticamente (Detect automatically). The transcript shows in the conversation and each audio is transcribed only once.
It uses the environment’s OpenAI account if there is one in Integraciones; otherwise, the AI included in your plan, and the minutes come out of the AI budget (see AI usage). Each audio can be up to 20 minutes and 25 MB. The audio goes in full to the transcription service; the transcript enters the conversation like everything else the person writes. If you turn it off, the assistant only knows a voice note arrived.
Privacy
Section titled “Privacy”Under Privacidad y seguridad (Privacy and security), Enmascarar datos personales antes de enviarlos al modelo (Mask personal data before sending it to the model) is on: ID numbers, tax IDs, emails, phones, accounts or cards, addresses and names are replaced by placeholders before going to the model and restored in the answer. The organization can make it mandatory, and every call is recorded in AI governance.
Privacy notice
Section titled “Privacy notice”When your organization needs the person’s authorization to process their data before chatting —as required, for example, by Colombia’s Law 1581 of 2012—, turn on Pedir aceptación del tratamiento de datos antes de conversar (Ask for acceptance of data processing before chatting) under Privacidad y seguridad. It comes off and, once on, applies to both modes and every channel:
- The first message is the notice with two options: a list on WhatsApp and buttons on web chat. The person can also reply «1» or «2», or type one of the replies that count as acceptance or refusal.
- Until they accept, nothing they write reaches the model or is stored (only a mark that a message arrived remains), and any files and voice notes they send are discarded unread.
- On acceptance, the date and the notice they saw are kept in the conversation as evidence. If you change the notice text, it is asked again; also in every new conversation.
- If they don’t accept, they get the Mensaje si no acepta (Message if they don’t accept); if they write again, they get the notice again.
- If they later ask to withdraw their authorization, the assistant revokes it with the built-in tool «Revocar autorización de datos» (Revoke data authorization): what was said up to then stops going to the model and the notice is asked again.
You set the Texto del aviso (Notice text; empty uses a suggested one), the Enlace a la política de tratamiento de datos (Link to the data processing policy; optional, with https://), the Opción para aceptar and Opción para no aceptar (accept and decline options), other replies that count as acceptance or refusal and the Mensaje si no acepta, with a preview alongside. The notice is automatic and does not detect emergencies: if you handle health or safety matters, mention the emergency line in the text (for example, 123 in Colombia). Your organization defines the text and the policy; Zoan Cloud enforces the notice by system and keeps the evidence.
Test, publish and promote
Section titled “Test, publish and promote”- The bottom panel has Playground (tests what is saved), Validación (Validation) and Debug. The Pruebas (Tests) tab keeps cases you can re-run after each change.
- Publicar (Publish) puts the version on the channels.
- Promover (Promote) copies the configuration to another environment (for example, from Testing to Production); channels, keys, machine and robot session are set up in the target.
- Exportar and Importar save or load the assistant as JSON.
Conversations and monitoring
Section titled “Conversations and monitoring”Under Conversaciones you filter Todas, Mías, Por atender or Solo escaladas (All, Mine, To handle, Escalated only), and you can Tomar (Take) a web chat conversation, Soltar (Release) it, Devolver al bot (Return to the bot) or Cerrar (Close). Escalated WhatsApp conversations are handled in Mosend.
Centro › IA › Monitoreo shows how your assistants are doing: conversations, contacts, resolved, converted, handed to a person, abandoned, first response time and cost, with what Requiere acción (Needs action), the breakdown per assistant and per hour of the day, a channel filter and CSV export.
Next steps
Section titled “Next steps”- Knowledge bases
- Tasks — where the approvals the assistant asks for are decided.
- AI usage and Governance