Legal
Agent Policy
Last updated: July 29, 2026
1. Overview
This policy explains how merQato's AI agents behave, what they can and cannot do, and what happens when something goes wrong.
2. Draft-First, Human-Approved
Our agents never publish, send, book, or charge anything autonomously. Every agent action — including social media drafts, booking confirmations, and guest replies — is queued for human review and approval before it goes live.
3. AI Models We Connect To
Agents can run on:
- OpenRouter — a gateway to multiple AI model providers (e.g., community models, commercial models). Your OpenRouter API key is configured in the Operator Admin. We do not store your OpenRouter key.
- Local Ollama — runs AI models on your own hardware. No data leaves your machine.
4. What Agents Do NOT Do
- Our agents do not make binding commitments on behalf of your business
- Our agents do not process payments directly
- Our agents do not access your guests' personal data unless you provide it in a prompt
- Our agents do not store conversation history between sessions (each interaction is stateless)
5. Human Escalation
Every agent workflow includes a human approval step. If an agent cannot complete a task or generates unsatisfactory output, the request stays in queue for manual review.
6. Model Transparency
When you configure an agent, you choose the model. You can switch models at any time. Model performance and outputs vary — we do not guarantee specific quality levels for any particular model.
7. Rate Limits and Availability
AI model providers may impose rate limits. If a model is unavailable, agents will return an error message rather than making up a response. This is by design — we prefer a clear error over a hallucinated answer.
8. Feedback
If an agent produces incorrect or harmful output, flag it through the Operator Admin. We use that feedback to inform model configuration decisions — we do not retrain or fine-tune models without explicit agreement.
9. Contact
Questions about agent behavior? Email hello@merqato.digital.