Frequently asked questions
Common questions about how our agentic AI works inside Odoo — security, control, and where it fits next to what's already built in.
Yes. Enterprise ships a handful of built-in AI features: content generation, OCR-based document digitization, predictive lead scoring, and similar. They're useful, and they're fixed — you can turn them on or off, not open them up. Our agentic AI framework is a different kind of thing: it runs your own agents, with your own logic, wired into your own workflows.
Because “the AI got it wrong” isn't actionable on its own — with a closed feature, that's all you get. With our agents, every run is logged step by step: which tool was called, what came back, why it routed where it did. “Wrong” becomes something specific and fixable, like a bad tool call, a stale prompt, or a routing mistake.
Most of the time, yes. An agent's instructions, model, and available tools all live in a versioned manifest, editable from the Agentic Admin UI. Changing a prompt is a manifest edit and a version bump, not a code deployment.
You don't debug it blind. The Replay Playground pulls up the exact recorded run, lets you tweak the prompt or a parameter, and re-executes it read-only so you can check the fix actually works before it touches production data.
No. Every agent reads and writes through the same ORM, record rules, and field-level access groups as a person clicking through the interface. A salesperson can't extract HR salary data through a chat window any more than through the Employees app.
No. Our LLM client factory supports any model or inference provider, configured independently per agent — OpenAI, Anthropic, ZAI, self-hosted models, and others. A customer-facing chat agent and a custom back-office agent can run on two different providers in the same deployment, each chosen for the task rather than the contract.
Every figure an agent states comes from an actual tool call against Odoo's live records — a search_read, a search_count— rather than the model's own recall. Separately, our Ground Truth Adherence metric checks every claim in a final answer against the actual tool results behind it before anyone treats it as fact.