Skip to content
S SiPesan Docs
ID EN

AI Agent

The AI Agent is an assistant that answers customers automatically using a language model. It follows the persona and business details you set, handles free-form questions, and hands the conversation to a human agent when it needs to. This guide covers how to switch it on, shape its personality, choose the engine behind it, and what happens when it gives way to a person.

Two Modes

Before anything else, decide how large a role the AI plays:

  • FAQ mode — your chatbot stays in charge. The AI only answers when the chatbot tree fails to match the customer's input. A good way to start: the standard flow stays under the chatbot, and everything else goes to the AI.
  • Standalone mode — no chatbot tree. The AI holds the conversation from the first greeting through to escalation.

When in doubt, start with FAQ. You can always move up to Standalone once you trust the AI's answers.

Switching AI On Per Channel

The AI is set per channel, not once for everything. Open a channel's edit form (WhatsApp, Telegram, Instagram, or WhatsApp Web) and pick one of three states:

  • Use a specific AI Agent — choose one of the AI Agents you've built. This channel takes on that agent's persona and settings.
  • Follow the business default — leave the agent field empty. The channel follows the AI settings under AI Settings.
  • Humans only — turn on the option to disable AI. Every chat on this channel is handled by agents and the chatbot alone, with no AI.

This way one business can have, say, a Sales channel whose AI leans into offers, and a Support channel whose AI plays it safe — or no AI at all.

Persona

The persona shapes how the AI speaks. Set it under AI Settings (for the business default) and on each AI Agent (for named agents):

  • Name — the name the AI uses to introduce itself, e.g. "Sipa".
  • Tone — choose casual & friendly, professional, or formal. This sets the register of its replies.
  • Extra instructions — free-form guidance, such as "Always mention the official warranty" or "Never promise a discount". Up to about 2,000 characters.

Create several named AI Agents when you need different personas per channel or per department. Manage them under AI Agent, then attach one to a channel using the steps above.

The Engine Behind the AI

The AI needs a language model to think with. There are three routes, set by the provider strategy under AI Settings:

  • Local — the qwen3:8b model runs on SIPESAN's server. Free, no API key needed, and no data leaves for a third party. This is the default.
  • Cloud — a model from the provider of your choice: Anthropic (Claude), OpenAI (GPT), Google (Gemini), or DeepSeek. You supply your own API key, stored encrypted.
  • Hybrid — a mix of both. Local-first tries the local model before reaching for the cloud; cloud-first does the reverse.

For cloud models, SIPESAN does not bill tokens. You use your own API key and pay the provider directly (Anthropic, OpenAI, and so on).

The Limits That Keep the AI Honest

The AI in SIPESAN is deliberately fenced in so it doesn't make things up. It may only draw on three sources: the business information you fill in, results from an AI Connector, and the conversation history. For facts that change — prices, stock — wire up a connector or a Knowledge Base rather than trusting the model's memory.

There's also a cap on turns. If a conversation goes in circles too long without progress (ten turns by default), the AI stops and hands it to an agent.

When a Customer Is Upset

SIPESAN reads customer sentiment on every AI turn. If the tone reads as angry, the conversation escalates to a human right away, without waiting for the turn limit. Labels like happy, neutral, frustrated, and angry are recorded in the AI Interaction Log, along with a confidence score, so you can review the conversations that went sideways.

Escalation & Follow-up

When the AI hands off a conversation (say the customer asked for an agent, ran out of turns, or turned angry), the session is "parked" and joins the Follow-up Inquiries queue. There, an agent sees the customer's name, the target department, the last message, and the time left in the 24-hour WhatsApp window. Click Open & Reply to jump straight into the chat; the session is still active, so there's nothing to reopen.

If you like, turn on per-agent SLA re-engagement: if no agent picks up a parked session within a set window (3 minutes by default), the AI can step back in so the customer isn't left hanging.

Connecting Data & Documents

An AI Agent rarely stands alone. Three features enrich its answers:

  • AI Flow — for flows with fixed steps (order tracking, bookings, lead qualification).
  • AI Connector — pulls live data from your systems (stock, order status).
  • Knowledge Base — answers from documents you upload (price lists, policies). Documents can be attached per AI Agent, so each agent has its own "reading list".

AI Billing (Optional)

By default, using the AI adds no SIPESAN cost beyond your MAU bill — you simply pay the cloud provider directly (or nothing at all with the local model). But if you want to cap and track AI usage with a quota, switch on credit-based billing under AI Credits. Credits are deducted after each successful reply, at different rates for local and cloud models, and the system warns you when the balance runs low.

Steps at a Glance

  1. Create an AI Agent under AI Agent, and set its persona (name, tone, instructions).
  2. Choose a mode (FAQ or Standalone) and a provider strategy under AI Settings.
  3. For cloud models, enter your provider API key.
  4. Attach the AI Agent to a channel from the channel's edit form — or disable AI on channels you want handled purely by people.
  5. Wire up a Connector or Knowledge Base if the AI needs to answer changing data.
  6. Watch performance in AI Analytics and work the parked sessions in Follow-up Inquiries.