Skip to main content

​
AI Chatbots (Kaanha Bots)

Kaanha AI includes a powerful chatbot system with two types: AI Agents (LLM-powered) and Flow Bots (visual decision trees). Both can run simultaneously and hand off to human agents. Both live under the Kaanha Bots section in the sidebar.

​
AI Agents

AI Agents use large language models to have natural conversations with your customers. They are configured with a business task so they stay focused on your use case. There is a single agent type today — Built-in LLM, powered by OpenAI, Anthropic, Google, Groq, or a custom OpenAI-compatible endpoint. (Two earlier agent types — a Python webhook bot and a Notion-only RAG bot — have been retired; a Python-style integration is now added as a tool an agent can call rather than a separate agent type, and Notion content is served the same way any other Knowledge Base is: link a Notion-backed KB to a regular agent.)

​
Creating an AI Agent

  1. Go to Kaanha Bots → AI Agents → New AI Agent.
  2. Choose a Mode:
    • Active conversational — replies to customer messages directly.
    • Passive watcher — silently observes conversations and takes side-effect actions (tagging, notes, Slack alerts) but never replies to the customer.
  3. Fill in the required fields:
  1. Configure the rest from the agent’s own tabbed editor — General, LLM Config, Knowledge Bases, Triggers, Behavior, Tools, Channels, Team, Test.

​
How an Agent Gets Matched to a Conversation (Engagement Scope)

Each agent has an engagement mode, set in its Triggers tab: An agent can also be scoped to specific channels (WhatsApp Cloud API, WhatsApp Connect, SMS, Slack) — leave this empty to match on every channel.

​
Business Task Requirement

Every active AI Agent must have a businessTask field. This is required by Meta’s WhatsApp Business Policy. The platform appends a safety prompt enforcing task boundaries to every agent’s system prompt.

​
Behavior Settings

The Behavior tab controls how an agent transitions in and out of a conversation:

​
Testing an AI Agent

Use the Test tab to simulate a conversation without touching real customer data. The test panel also returns a delegation trace showing which sub-agents (if any) were invoked, making it easy to debug multi-agent setups.

​
Knowledge Bases

Knowledge Bases let you ground AI Agents in your own content. An agent can be linked to multiple knowledge bases; retrieval is hybrid (vector + keyword) across all linked bases.

​
Source Types

​
CRM-Backed Knowledge Bases

Set a KB’s Entity field to contact, deal, task, product, company, or supplier (or a custom entity slug). The KB is automatically re-indexed whenever that entity type is mutated — imports, webhook events, AI tool updates all trigger a re-sync. This keeps the KB current without any manual sync step.

​
Document Q&A — Ask AI with a File Attached

The Ask AI panel (available on Contacts, Tasks, Suppliers, Products, Orders, and Companies) can now answer questions grounded in a file you attach on the spot, not just your CRM data.
  1. Open the Ask AI drawer (the paperclip button next to the input) and attach a PDF, Word document, CSV, or plain text file (up to 16MB).
  2. Ask your question. The relevant text is extracted from the file and used alongside your normal CRM/Knowledge Base context to ground the answer.
  3. The attachment stays available for follow-up questions in the same session — you don’t need to re-attach it for every message.
If the file can’t be read (unsupported type, corrupted, or a temporary storage hiccup), Ask AI tells you clearly rather than silently ignoring the attachment.

​
Contact Memory

AI Agents automatically extract facts from closed conversations and remember them for the next interaction with that same contact. Memory is per-contact and surfaced in the system prompt at the start of each turn. Memory entries include: name, preferences, past issues, purchase history, language — whatever the LLM extracts as meaningful. Entries have a confidence score and an optional expiry date.

​
Product Catalog Auto-Injection

Enable Auto-inject Catalog on an agent to automatically prepend your top-N active products into the system prompt. The agent can then answer pricing, availability, and product questions without any tool calls. Configure the number of products (default: 20) in the agent settings. The catalog block is capped at 3,000 characters and placed before the agent’s own system prompt.

​
Built-in Tools

AI Agents can invoke built-in CRM and product tools mid-conversation. Enable or disable specific tools from the agent’s Tools tab — by default an agent can call every registered tool; turn on “choose specific tools” to restrict it to only the ones you pick. Tools are grouped by risk category (Read / Write / Communication / Admin / Search) so you can see at a glance which ones can only look things up versus which ones can send messages or change data. A representative sample:

​
Connect External Tools via MCP

If you have your own tool server (or a vendor’s) that speaks the Model Context Protocol (MCP), you can connect it under Settings → AI → MCP Servers. Once connected, any agent can list and call that server’s tools mid-conversation — the same way it calls a built-in tool. This is how you extend an agent’s reach beyond what’s built into Kaanha AI, without needing a new agent type or a code change. To add one:
  1. Go to Settings → AI → MCP Servers → Add connection.
  2. Give it a name, the server’s URL, and an auth token if it requires one.
  3. Save, then use Test connection to confirm it’s reachable.
  4. Any agent’s tool-calling loop can now discover and call that server’s tools.
MCP tool calls are treated as a write-risk action and require confirmation in the same places other write-risk tools do, since an external server’s tools are opaque to Kaanha AI.

​
Multi-Agent Orchestration

Feature flag required. Multi-agent orchestration is gated behind the multiAgent feature flag. Contact your admin to enable it for your organization.

​
Overview

A Lead Agent can delegate work to one or more Helper agents mid-conversation. The caller (contact) sees one seamless conversation; the delegation is invisible to them.
  • Lead Agent — handles the primary conversation, decides when to delegate
  • Helper — a specialist agent called as a tool by the Lead
  • Solo — a standalone agent with no delegation

​
Setting Up an Orchestration

  1. Go to Kaanha Bots → Orchestration.
  2. Click + New Orchestration to create a Lead Agent.
  3. On the orchestration board, drag-and-drop Helper agents onto the canvas.
  4. Connect them to the Lead Agent.
  5. Each Helper needs a short description (used by the LLM to decide when to call it).
  6. Save.
The Lead Agent sees each Helper as a callable tool. When the LLM decides to delegate, it calls the Helper’s full LLM + tool loop and returns the result. A Helper agent can also reach outside the conversation entirely — cross-channel tools let it push a WhatsApp follow-up to the contact mid-call or post-call.

​
Depth Limits

​
Human Handoff Through Chains

If any Helper agent triggers a HUMAN handoff (keyword or tool call), the signal propagates up through the entire delegation chain and routes the conversation to a human queue immediately.

​
Testing Orchestrations

The Test button on a Lead Agent card runs a simulated conversation and returns a delegation trace — which agents were called and in what order.

​
Flow Bots

Flow Bots are visual, rule-based conversation trees. They’re deterministic — the same input always produces the same output. Like AI Agents, they use the same three-mode engagement scope (keyword match / all conversations / fallback) and can be scoped to specific channels.

​
Creating a Flow Bot

  1. Go to Kaanha Bots → Flow Bots → Create Flow.
  2. Enter a name and trigger keywords.
  3. Click Visual Editor → to open the drag-and-drop canvas.

​
Node Types

​
Connecting a Bot Directly to a Slack Channel or WhatsApp Group

Every bot has its own Channels tab in the editor. Use it to route a Slack channel or a WhatsApp group straight to that specific bot — every message posted there goes to it directly, skipping the usual keyword/catch-all matching entirely. This is the fastest way to give an internal team (Ops, Purchasing, Stock, whatever you name your bots) their own dedicated Slack/WhatsApp inbox handled by AI.
  1. Open the bot you want to connect and go to its Channels tab.
  2. Tick the Slack channel(s) or WhatsApp group(s) you want routed to it.
  3. If a channel is already routed to a different bot, you’ll be asked to confirm before reassigning it — the previous bot stops receiving messages from that channel once you do.
There’s also a channel-centric view of the whole map at Settings → CRM → Ops Automation → Ops Agent Routing, if you’d rather see every channel’s assignment on one screen and edit several at once.

​
Backing Up or Cloning a Bot — Export Manifest

Every bot’s full configuration — prompt, provider, tools, behavior settings, everything you can set in its editor — can be exported as a single JSON file.
  1. Open the bot’s editor.
  2. Click Export manifest near the top of the page.
  3. A JSON file downloads with everything needed to recreate this bot’s exact configuration.
This is useful for backing up a carefully-tuned bot before making risky changes, or for handing a known-good configuration to your team as a reference. (Re-importing a manifest to create a new bot from it isn’t available yet — export is a read-only snapshot today.)

​
Priority Chain

When an incoming message arrives, Kaanha AI checks in this order:
An operator can pause automation on a specific conversation without disconnecting the bot entirely — replies are silenced but the bot stays configured and resumes exactly where it left off once unpaused.

​
Human Handoff

AI Agents automatically hand off to humans when:
  • The contact types HUMAN, AGENT, HELP, or any of your configured handoff keywords
  • The agent’s own handoff signal appears in its response
  • Max turns are reached
  • A content safety violation is detected
After handoff, the conversation status changes to OPEN and appears in the agent queue — unless Silent AI transitions is on, in which case this happens invisibly to the customer.

​
Content Safety

All AI responses pass through a content safety filter (22 patterns across 7 categories, with Unicode normalization to catch obfuscated attempts) before being sent:
  • Harmful instructions
  • Personal data requests
  • Off-topic content
  • Prompt injection attempts
Blocked responses trigger automatic handoff to a human agent.

​
AI Disclosure

The first AI response in a new WhatsApp Cloud API conversation includes a disclosure that the contact is speaking with an AI, per Meta’s WhatsApp Business Policy. You can turn this off per-agent in its Behavior tab — WhatsApp Connect, SMS, and Slack conversations never show it regardless of the toggle, since Meta’s requirement is specific to Cloud API.

​
Chatbot Analytics

Go to Kaanha Bots → Analytics to see:
  • Total conversations handled per agent
  • Handoff rate
  • Average turns before handoff
  • Most common trigger keywords
  • Token usage per agent
  • Voice Agents — the phone counterpart, sharing the same prompt/tools philosophy
  • Integrations — connecting Slack, Notion, and other external systems
  • Security & Compliance — content safety, AI disclosure, and audit logging