Responsible AI Policy
Botmanor orchestrates AI agents on the models you choose. Here is how we handle the AI part of that — what's live, what's on the roadmap, and the principles that don't change either way.
Botmanor is pre-launch. Last updated: July 2026.
Our principles
Your models, your keys
Botmanor does not run its own foundation model. You connect the LLM providers you choose — OpenAI, Anthropic, Google, Cohere, Mistral, Groq, Together, HuggingFace, Azure OpenAI, AWS Bedrock, and custom OpenAI-compatible endpoints — each with your own API key, vaulted through the platform integrations service and resolved at execution time rather than stored in the agent definition. Self-hosted inference is available through the custom provider type: a custom OpenAI-compatible endpoint pointed at a model server you host yourself.
Your content is not training data
We do not use your conversations, knowledge bases, or agent configurations to train foundation models — ours or a provider's. Your content is processed to serve your agents, nothing else. Whether a connected LLM provider itself trains on data you send it is governed by your agreement with that provider; we recommend checking each provider's own policy, and using an enterprise or zero-retention tier where that matters.
Human oversight
AI-generated output can be inaccurate or inappropriate for a given context. You are responsible for reviewing and validating agent output before relying on it, and every conversation lands in a shared team inbox where a person can read the thread and take over. We say plainly, in this policy and on our Security and FAQ pages, where automated handling ends and a human decision begins.
Transparency about what is live
We label AI capabilities as Live or Roadmap and keep that label accurate as the product ships — see the table below. We would rather name a boundary than blur it, including on this page.
Where AI is used — live vs. roadmap
The control plane — configuring providers, knowledge, tools, and agent relationships — is real today. Consuming those inside a live conversation is the agent runtime upgrade we are shipping in the open, and this table says exactly where that line sits right now (11 provider types supported).
LLM provider registry & connection testing
Register and verify OpenAI, Anthropic, Google, Cohere, Mistral, Groq, Together, HuggingFace, Azure OpenAI, AWS Bedrock, or a custom OpenAI-compatible endpoint, with credentials vaulted and resolved at execution.
Chat execution against OpenAI-compatible endpoints
The live chat path calls OpenAI-compatible endpoints today (including Azure OpenAI and custom/self-hosted endpoints); native adapters for the remaining provider families are on the roadmap.
Knowledge base ingestion, chunking & embedding
Upload documents, choose a chunking strategy (fixed-size, recursive, semantic, markdown, or code-aware), and embed into a vector collection with a visible index lifecycle.
Semantic search validation
Run the questions your users will ask against a knowledge base and inspect the retrieved chunks and relevance scores before anyone relies on them.
RAG grounding inside a live agent answer (with citations)
Automatically injecting retrieved chunks into a live conversation turn, with citations back to source, lands with the agent runtime upgrade.
MCP tool registry, discovery & risk classification
Register MCP servers, test connections, discover tools and resources, and classify each tool by risk level in a policy definition — the vetted inventory the runtime will obey.
Agent-invoked tool calls inside a live conversation
An agent actually calling a registered MCP tool mid-conversation is part of the agent runtime upgrade and is not live yet.
Agent-to-Agent (A2A) registry & structured messaging
A registry of known agents with trust levels, declared connections, and structured messages with priority and delivery status — a record of intended hand-offs.
Live cross-agent delegation in production conversations
An agent automatically delegating to another agent mid-conversation is on the runtime roadmap, not shipped.
Execution tracking, token counts & cost
Every run records status, input/output, the model called, token counts, and computed cost for audit and billing.
Automated low-confidence handoff to a human
Detecting a low-confidence or sensitive turn and automatically assigning a human mid-conversation is designed and on the roadmap; today a person supervises from the shared team inbox.
What we won't do
- Train foundation models — ours or a connected provider's — on your Customer Data.
- Choose an LLM provider on your behalf without your explicit configuration.
- Claim a capability is live when it lands with the agent runtime upgrade — see the table above and our FAQ for the current boundary.
- Auto-invoke an MCP tool or delegate to another agent inside a live conversation until that runtime capability actually ships.
Questions about how AI is used in Botmanor? Read our Privacy Policy, our Security posture, our Subprocessors list, or contact us.