AI Bot

AI Bot (ai-bot) is the substrate you build conversational interfaces on — customer service bots, internal helpdesk assistants, in-product copilots, WhatsApp concierges.

It is infrastructure, not a product. You declare the bot; the block provides the runtime, the channel plumbing, the conversation state and the escalation machinery. End users talk to your bot. There are no pre-built bot templates and no bot-builder UI here — those are things you build on top.

The idea: cheapest viable resolution

Most conversational products send every message to an LLM. That is simple, slow, and expensive in a way that scales linearly with success.

AI Bot instead runs each message through an escalating cascade. A message stops at the first stage that can handle it confidently:

inbound message
    │
    ├─▶ 1. rules / regex        microseconds, no model cost
    ├─▶ 2. dynamic model        fast, generated from your rules
    ├─▶ 3. fast classifier      milliseconds, tiny cost
    ├─▶ 4. small language model low cost, handles ambiguity
    └─▶ 5. LLM                  terminal — always resolves

“What are your opening hours” is a rule match that never touches a model. A paraphrase the regex misses is caught by the dynamic model. Only genuinely novel or complex messages reach stage 5. This is the central cost and latency property of the block — the cascade is why a high-traffic bot does not cost like a high-traffic LLM integration.

Declare a bot

A bot is a versioned, signed, tenant-scoped YAML artifact. Structure is declarative; where declaration is not enough, you drop into scripts at defined hook points.

kind: bot
metadata:
  name: acme-support
  tenant: acme
  product: support-portal
  version: 7
spec:
  channels:
    - kind: web
      enabled: true
    - kind: slack
      enabled: true
      workspace_ref: vault:acme/slack-token

  persona:
    prompt_ref: gateway:acme-support-persona-v7
    style: professional
    locale_default: en-US
    locales: [en-US, es-MX, pt-BR]

  nlu:
    cascade:
      - stage: rules
        confidence_threshold: 0.95
        rules_ref: rules/intents.yaml
      - stage: classifier
        confidence_threshold: 0.85
        model_ref: ai-ml:acme-support-classifier-v7
      - stage: llm
        confidence_threshold: 0.0      # terminal
        model_ref: gateway:acme-support-llm-default

  orchestrate:
    handlers:
      - intent: business_hours
        path: canned
      - intent: order_status
        path: skill
        skill_ref: ai-computer:acme-order-lookup-v2

Channels

One bot definition serves every channel it declares. Inbound messages are normalised to a single canonical shape before the cascade sees them, so your rules, intents and handlers are written once rather than per-channel.

Channel	Notes
Web API	The foundational channel; also what custom front-ends use
Slack	Workspace credentials from the vault
Microsoft Teams	App registration from the vault
WhatsApp	Account plus approved message templates
Custom	Bring your own transport against the same canonical message

What it composes

AI Bot orchestrates other blocks rather than reimplementing them:

Need	Handled by
Model calls	AI Gateway
Conversation memory	AI Memory
Knowledge retrieval	AI Knowledge
Multi-step autonomous work	AI Flow
Quality measurement	AI Evals

Every model call — the SLM stage, the LLM stage, the persona prompt — goes through the AI Gateway, which means bot traffic lands in the same budgets, caches and cost accounting as everything else.

Core Concepts — the cascade, confidence, conversation state and orchestration
Configuration — the bot definition in full
Operations — deploying, model generation, analytics and compliance

AI Bot

The idea: cheapest viable resolution

Declare a bot

Channels

What it composes

Next