August 19, 2026 · updated August 31 · 11 min · Odoo AI

Odoo 19 AI agents: give AI tools without handing over the ERP keys

Odoo 19 documentation now describes agents that can understand natural-language requests, consult sources and use tools inside Odoo. The important development is not merely conversational: depending on their assigned topics, agents can open a view or trigger a business action, such as creating a CRM lead.

For Belgian or French organisations, this shift from answering to acting changes project scoping. A useful prompt does not replace access rights, business rules or data controls. Before deploying an agent in Odoo 19, teams must define what it knows, what it may do and how its actions can be evidenced.

Understand the architecture before selecting a use case

In Odoo’s documented model, an agent has a purpose and system prompt. Topics provide instructions and group the tools it can call. Sources — PDFs, web links, Documents files or Knowledge articles — provide indexed context. Without a topic, an agent can provide information but cannot perform actions.

This separation should become the governance boundary: a source expands what the agent may know; a tool expands what it may change. These are different risks and should not be approved in one step.

Three levels of operational risk

Answering from a controlled corpus

An internal support agent can be limited to approved procedures by enabling source restriction. Owners, versions, expiry dates and document access still need management. An answer based on an obsolete procedure is an error at scale.

Guiding users inside Odoo

Natural-language search and opening views may look low-risk, but they can expose sensitive records if the access matrix is too broad. Tests should use real roles — sales, accounting, HR and multi-company managers — and include negative controls.

Triggering business actions

AI server actions separate the “manager”, which selects a tool and its arguments, from the “worker”, a standard server action that executes the logic. Odoo states that a selected tool executes unconditionally unless its own code prevents it. Limits, allowed states, existence checks, segregation of duties and idempotency must therefore live in the tool, not only in the prompt.

Documents: a strong pilot if an exception queue remains

Document automation can classify files, add tags, create activities or create business records such as vendor bills. It is configured at folder level, allowing a progressive workflow. A sensible pilot begins with sorting and tagging, measures errors, and only then permits business-record creation with human validation and reconciliation.

The documented multi-document PDF example captures the principle: detect the file, label it “To Split”, then stop. A well-designed agent also knows when not to proceed.

Helpdesk and Live Chat: formalise the handoff to a person

Odoo 19 documentation now details two support-oriented uses. In Helpdesk, agents, automations and AI fields can summarise, categorise or prepare a ticket for handling. In Live Chat, an agent can answer, qualify the conversation and trigger lead creation when the request requires custom pricing, intervention or an account change, or when its confidence is insufficient.

This handoff should not be treated as a sentence in a prompt. Teams must define escalation criteria, the minimum information to collect, the owner of the human queue, its response target and duplicate prevention. Odoo describes a lead-creation topic that gathers contact details one at a time, confirms them and calls the tool exactly once. That conversational discipline should be tested as a business rule.

AI fields: generated data still needs governance

Odoo can add AI fields through Studio or as property fields. Their values can be generated from the record context and fields referenced in the prompt. A scheduled action, enabled by default, fills certain empty text and property AI fields once a day.

Before using these values for a priority, date, amount or category, the project team must decide whether the field is a suggestion or an authoritative value. It should also control permitted inputs, possible values, recomputation after changes and human corrections. In an Odoo migration, custom AI fields, their prompts and the scheduled action belong in the customisation inventory and regression tests.

Underside analysis: govern capability, not just the model

Model choice matters, but ERP risk lies elsewhere: tool scope, source quality, user permissions, coded rules and logs. We recommend a register for every agent listing its business owner, purpose, companies, sources, tools, sensitive data, human approval, metrics and shutdown procedure.

In Belgium and France, workflows involving accounting, HR data or customer contacts should be reviewed alongside the organisation’s access and data-protection governance. Odoo should not be presented as providing automatic legal compliance: the organisation remains responsible for its use case, configuration and controls.

Production roadmap

This approach complements a robust Odoo security and access model and an Odoo 19.4 change-testing plan.

Official sources

Underside helps organisations scope, prototype and secure Odoo automation. A short pilot should produce as much evidence of control as productivity gain.

Back to blog