Odoo 20: governing agents that execute ERP processes

Article created on September 25, 2026 · Publication analysed: September 24, 2026 · Source: Odoo

Odoo 20 lets users describe an automation in natural language and then allows an agent to execute it in the ERP. The announcement also covers an accounting assistant, website generation, and connections to external AI tools while retaining access rights. For organisations, the main change is not conversational: AI is entering the business execution chain.

1. From suggestions to action in the ERP

According to the official announcement, an agent can confirm a sales order when its matching email arrives, assign a lead, create a project when a contract is signed, or report low stock. Odoo says the agent explains the automation before it is activated. External AI tools can also read and write database information while preserving existing access rights.

In accounting, the assistant can query reports about receivables or cash flow and audit a reconciliation to produce a summary and recommendations. Odoo also announces account and analytic predictions. These capabilities must be separated: answering, recommending, preparing an entry, and validating a transaction do not create the same level of risk.

2. What this changes for a Belgian or French organisation

For an SME, Odoo 20 shortens the distance between a business rule expressed by a manager and its automation. Benefits may emerge quickly in CRM qualification, stock alerts, or task preparation. Each automation still needs an owner, a defined scope, and a way to stop it.

For mid-market organisations, large enterprises, or public administrations, separation of duties is central. Existing permissions are a useful foundation, but they do not prove that a sequence of actions is appropriate. Finance, procurement, sales, and production teams must decide which actions remain recommendations, which require approval, and which may be executed automatically. In Belgium and France, logs, data minimisation, and evidence of approval matter as much as the chosen model.

3. Underside analysis: governance must cover the process

Odoo 20 confirms an evolution already visible in agents, RAG, and MCP: the ERP is becoming both a context source and an action surface. Governance limited to the model provider is therefore insufficient. It must cover the trigger, data read, tools called, records changed, exceptions, and business outcome.

Sovereignty also depends on the ability to understand the system and regain control. A local model does not compensate for opaque automation; conversely, a precisely bounded cloud service may be more controllable when flows, subprocessors, permissions, traces, and exit procedures are documented. For RAG connected to Odoo, citations and document versions must remain verifiable. For Apple Enterprise or mobile devices, user identity and device posture should form part of the authorisation decision.

4. A controlled deployment path

Begin by inventorying current automations and selecting one frequent, reversible, low-risk process. Define authorised data, technical identity, least-privilege permissions, approval thresholds, quality metrics, logs, and rollback. Test normal cases as well as ambiguous emails, duplicates, missing data, permission changes, and AI-provider outages.

Before expanding into accounting, procurement, or production, measure correct-action rates, human rework, time saved, cost, errors prevented or introduced, and the ability to explain every decision. Production rollout should follow process criticality, not demo convenience.

Read the official Odoo 20 announcement · Read the official release notes