Odoo 20 turns AI from a writing assistant into something closer to an operational layer inside the ERP. The useful architecture is not simply “ask AI a question.” Odoo Agents can combine instructions, indexed Sources and controlled Tools so an agent can retrieve company knowledge, reason about a task and invoke approved actions.
For developers, the important question is not whether Odoo has AI. It is how to design an agent that is useful without allowing probabilistic reasoning to replace deterministic business rules.
The Odoo AI Agent model
Current Odoo documentation describes Agents using Skills and Sources. Skills define instructions and tools for a task. Sources provide indexed information the agent can retrieve.
Think of the layers as:
- Agent: identity and high-level behavior.
- Skill: task-specific instructions and allowed capabilities.
- Source: knowledge available for retrieval.
- Tool: an operation the AI may invoke.
That separation is useful because a sales assistant should not automatically have the same tools as an HR or accounting assistant.
Skills: constrain the job before expanding capability
A good Skill is narrow enough to evaluate.
Instead of one “Company Super Agent,” prefer skills such as:
- qualify an incoming lead;
- summarize an opportunity;
- answer questions from HR policy;
- find a product procedure;
- prepare a customer follow-up;
- classify a support request.
The narrower the responsibility, the easier it is to define expected answers, allowed tools and failure behavior.
Sources: ground the agent in company knowledge
Odoo currently documents source types including PDFs, web links, Documents-app documents and Knowledge articles.
A source-backed agent can answer from internal material such as:
- SOPs;
- product manuals;
- HR policies;
- sales playbooks;
- support procedures;
- compliance guidance.
Odoo also documents Restrict to Sources, which can constrain the agent to its active source material.
That is useful when the desired behavior is “answer from approved company documentation” rather than “use general model knowledge whenever useful.”
Tools: where AI becomes operational
Tools are the most powerful part of the design.
A tool lets the agent do something rather than only generate text.
Odoo 20’s AI server-action documentation describes a Manager/Worker pattern:
- the AI server action acts as the Manager and chooses an appropriate tool;
- the tool acts as the Worker and performs the operation.
The key engineering rule is that business validation belongs in the Worker.
If a tool creates a discount, changes an order, assigns a lead or triggers an approval, the underlying implementation should still enforce the same permissions and business rules expected from normal application code.
A prompt is not a security boundary.
A safe example: lead-triage agent
Imagine an agent that helps a sales team process new inquiries.
Sources
- qualification policy;
- product/service descriptions;
- territory rules.
Skills
- summarize inquiry;
- identify requested service;
- classify urgency;
- recommend next step.
Tools
- search existing customer;
- create a CRM activity;
- assign a permitted tag;
- draft a follow-up note.
Avoid giving the first version a broad “update anything in CRM” tool.
Start narrow, evaluate, then expand.
Production design checklist
Before exposing a tool, answer:
- Which model/records can it access?
- Does it read or write?
- Which authenticated user permissions apply?
- What fields can change?
- Which validations run?
- Can it trigger money, stock, payroll or external communication?
- Does the action need approval?
- How will the result be audited?
- What happens when the agent is uncertain?
Evaluate the agent with known-answer tests
Do not test only with impressive demo prompts.
Create a test set containing:
- straightforward questions;
- ambiguous questions;
- missing-information cases;
- conflicting source material;
- unauthorized requests;
- attempts to invoke the wrong tool;
- old/outdated source material.
Record expected behavior.
An AI workflow needs regression testing just like code—especially after changing prompts, sources, tools or models.
Native Agent vs MCP client
Odoo’s native AI Agent runs within Odoo’s AI architecture.
MCP allows compatible external AI clients to connect to Odoo through exposed tools.
Use a native agent when the workflow belongs naturally inside Odoo.
Evaluate MCP when an external AI environment needs controlled access to Odoo capabilities.
They can share concepts, but they are not the same architecture.
Common mistakes
- Creating one agent with too many unrelated responsibilities.
- Giving write tools before proving the read-only workflow.
- Encoding business rules only in natural-language instructions.
- Ignoring source freshness.
- Assuming an authenticated AI call is automatically safe.
- Testing only successful examples.
- Allowing tools to perform broad arbitrary writes.
- Treating an LLM response as authoritative financial/HR/legal approval.
Odoo 20 AI Agent configuration: what the current interface exposes
Current Odoo documentation adds several details that matter when designing an agent. Creating and customizing agents requires the AI application to be installed. An agent is configured with a System Prompt, one or more Skills, optional Sources, a provider and a response style.
Current master documentation lists ChatGPT/OpenAI and Gemini/Google as supported providers. Response styles include Standard, Nuanced, Snappy & Creative, and Slow & Rigorous. Match those settings to the job: a policy assistant should prioritize precision and source grounding; a marketing ideation assistant can tolerate more creativity.
Preconfigured Skills worth understanding
Current Odoo documentation lists Skills such as View Builder, Information retrieval, Update Records, Create Records, Generate Images, Web Search, Website Builder, and Create Leads when CRM is installed. Available tools vary with installed applications.
Do not create a custom tool until you confirm a standard Skill/tool does not already cover the requirement.
What happens if an agent has no Skills?
Odoo explicitly states that an agent without Skills can provide information but cannot complete tasks or make changes to the database. That makes a no-Skill, source-grounded agent a useful low-risk starting point.
System Prompt vs Skill Instructions
Use the System Prompt for stable global behavior: identity, audience, boundaries, escalation, tone and general safety rules.
Use Skill Instructions for task-specific behavior: when the Skill applies, exact workflow, which tool to use, constraints and expected output.
Avoid duplicating long instructions in both places. Conflicting instructions make behavior harder to evaluate.
A practical rollout pattern for Odoo AI Agents
A production rollout is safer when capability grows in stages.
Stage 1: read-only knowledge assistant
Give the agent curated Sources but no business-changing tools. Evaluate retrieval quality, missing-answer behavior and access boundaries.
Stage 2: read-only Odoo tools
Allow narrowly scoped search/reporting tools. Test company boundaries, record rules and ambiguous prompts.
Stage 3: reversible low-impact actions
Add tools such as creating an activity or draft note. Keep the operation auditable and easy to reverse.
Stage 4: controlled business actions
Only after the earlier stages are stable should you consider tools that affect operational records. High-impact actions should have deterministic validation and, where appropriate, human approval.
This staged approach makes failures easier to understand because you change one risk dimension at a time.
Agent governance checklist
For every production agent, document:
| Area | Decision |
|---|---|
| Owner | Who is responsible for behavior and Sources? |
| Users | Which roles can access it? |
| Sources | Which documents/links are authoritative? |
| Skills | What tasks may it perform? |
| Tools | Which operations are exposed? |
| Writes | Which records/fields may change? |
| Approval | Which actions need a person? |
| Evaluation | Which test set proves expected behavior? |
| Review | When are prompts/Sources revalidated? |
An AI agent without an owner eventually becomes an undocumented business process.
Observability matters
When an agent participates in an operational workflow, keep enough evidence to investigate unexpected behavior. Depending on the implementation, that can include the initiating user, tool invoked, record affected, timestamp and resulting business action.
Do not log secrets or unnecessarily retain sensitive prompt content. The goal is accountable operations, not indiscriminate data collection.
Where AI should not be the decision-maker
Some workflows can use AI for classification or preparation while keeping the final decision deterministic or human-controlled.
Examples include:
- posting accounting entries;
- changing payroll outcomes;
- approving large discounts;
- releasing stock with compliance implications;
- deleting records;
- changing user access;
- making legally consequential decisions.
The useful pattern is often AI proposes → validated tool/human approves → Odoo executes.
FAQ
What can Odoo 20 AI Agents use as Sources?
Current documentation lists PDFs, web links, Documents-app documents and Knowledge articles.
Can an Odoo AI Agent change records?
Agents can use tools capable of operations, but permissions and the implementation of those tools matter. Design write capabilities narrowly and enforce deterministic validation.
What is an AI server action?
Odoo documents AI server actions as managers that decide which available tool should perform the task. The tool performs the actual operation.
Is this the same as RAG?
Source-backed retrieval overlaps with RAG-style use cases. A dedicated RAG architecture may still be justified for specialized retrieval, scale, model or external-data requirements.
Does an Odoo AI Agent need the AI app?
Many AI features can exist without the AI app, but Odoo says creating and customizing agents requires the AI application.
Can Odoo AI Agents search the web?
Current documentation includes Web Search as a Skill. Enable it only when external information is appropriate for the use case.
Which AI providers are documented?
Current master documentation lists ChatGPT/OpenAI and Gemini/Google.
Final takeaway
The best Odoo 20 AI Agent is not the one with the most tools.
It is the one with a clearly defined job, trustworthy Sources, the minimum required capabilities, deterministic business validation and a test set that proves what it should—and should not—do.
Planning an Odoo 20 project? Octivex Tech helps businesses with Odoo upgrades, custom module migration, integrations and AI-enabled workflows. Explore our Odoo services or contact our engineering team to discuss your environment.