Most Odoo automations are deterministic. A record changes, a domain matches, and a server action does exactly what it was told: update a field, send a template, create an activity. That reliability is valuable, but it breaks down the moment the next step requires judgement. Which helpdesk ticket sounds urgent? What should this uploaded PDF become? How should a customer message be summarised before a follow-up is created? In Odoo 20, automated actions and scheduled actions can summon an AI agent to handle that judgement step inside the automation itself.
That is why this feature matters more than a chat window. Odoo's release notes describe it under agentic automation: agents can be summoned by automated and scheduled actions, and you can check the automation chats to see how an agent handled a task and ask for changes based on the errors it faced. The agent becomes a step in a business process that starts on its own.
What changed in Odoo 20
Previous automation could move data, but it could not interpret unstructured input. It could copy text, evaluate a domain, or run Python code written in advance. It could not read a supplier document, decide which records it implied, or adapt when input did not fit a fixed template.
Odoo 20 agents add those capabilities to the same trigger framework. According to the release notes, an agent can create records, including from an uploaded file of instructions such as a PDF. It can update records, answer questions about a file, generate images and enriched buttons for websites and mailings, work from files linked in a chat, and take dictated requests by voice. Odoo selects the model automatically once a provider is chosen, and conversations are stored for up to 30 days.
The vocabulary changed too. What Odoo called topics — collections of instructions and tools that define how an agent behaves in a context — are now called skills. A skill is the reusable behaviour you want a trigger to invoke: read this kind of document, apply these rules, and stop when a decision needs a person.
Two trigger types, two very different jobs
Automated actions: react when something happens
An automated action fires because a record was created, updated, deleted, or changed in a way the rule watches. Put an agent behind that trigger and the system can react to messy, human input at the moment it arrives. A new lead with a long free-text enquiry can be summarised and categorised. A helpdesk ticket can be read for urgency and routed with a short explanation. A document attached to a record can be questioned, extracted, or turned into follow-up records.
The design rule is to keep the trigger narrow. The automation should decide when the agent runs, using ordinary domains and watched fields. The agent should decide how to handle the content it was given. If you let the agent decide both, you lose the predictability that made automation trustworthy in the first place.
Scheduled actions: sweep, review, and chase
A scheduled action fires because time passed. That suits agents well for periodic work: review records that have stalled, summarise a set of open items, check whether follow-ups are overdue, or prepare a digest for a person to approve. The work is important but not urgent, and a batch rhythm makes the output easier to review.
Scheduled agent runs also make cost visible. Every run consumes IAP credits, because Odoo 20 states that all AI features need credits bought via IAP. A rule that fires on every small field change can spend quietly all day. A scheduled sweep runs at a known cadence, which makes budgeting and sampling easier. Start scheduled, prove the value, then consider event-driven triggers for cases that truly cannot wait.
What an agent can actually do inside the run
It helps to separate the agent's verified abilities from the workflow you design around them:
- Create records. Ask an agent to create records for you, including by uploading a file of instructions like a PDF. In an automation, that means a document or message can become structured work instead of sitting in an inbox.
- Update records. An agent can be asked to update existing records, which is the capability that turns classification into action: set a category, add a summary, or prepare the next step.
- Answer questions about a file. While previewing a file, a user can ask questions about it. In a process, that same ability supports extraction and checking before anything is written.
- Generate images and buttons. Agents can generate images and enriched buttons for websites and mailings, which makes them useful in marketing and website workflows, not only operations.
- Reprocess sources. You can ask an agent to reprocess its sources and use the most up-to-date version, which matters when instructions changed after the first run.
- Filter by time period. When building views, agents can filter by time period such as weeks or quarters — useful for scheduled reviews of a defined window.
Agents also show live, specific updates on what they are doing instead of a generic thinking message, and their responses are interactive: questions, permission requests, and suggestions can be answered by clicking buttons. If an agent reaches its maximum tool call limit, it prompts for confirmation before proceeding. In an automated setting those controls are the difference between a helpful assistant and an unsupervised script with a language model attached.
Reviewability is the feature you will rely on
The release notes highlight automation chats for a reason. When an agent runs from an automated or scheduled action, you can check the chat to see how it handled the task and ask for changes based on the errors it faced. That creates an audit trail a business user can read: what the agent saw, what it decided, what it changed, and where it struggled.
Build review into the process from day one. For the first weeks, treat agent output as a draft: let it create tasks, summaries, and proposed updates, and have a person sample the automation chats daily. Look for patterns rather than single mistakes. Is the agent misreading one document type? Is a skill missing a rule your team applies instinctively? Is the trigger firing on records the domain should have excluded? Each pattern points to a fix in the trigger, the skill, or the source data.
Odoo also notes that agents can update themselves to best serve their use cases. Treat that as a capability to govern: if behaviour can evolve, your review has to check not only outputs but whether the behaviour still matches the process you approved.
Designing a safe agent automation
- Start with one narrow transition. Pick a single model, a single trigger event, and a single outcome, such as summarising new tickets in one team. Broad agents fail broadly.
- Let the automation own the gate. Use the rule's model, trigger, watched fields, and domains to decide exactly which records reach the agent. Test the domain as a search before you trust the agent with it.
- Write the skill like a work instruction. State the input, the steps, the allowed tools, the output format, and the stop conditions. Skills are where business judgement lives now.
- Separate drafting from committing. Where stakes involve money, stock, or customer commitments, let the agent prepare the record and let a person confirm.
- Cap the blast radius. Respect the tool call limit as a safety control. When an agent asks to continue past it, that is a signal the task was bigger or messier than expected.
- Review chats on a schedule. Assign an owner to read automation chats and log recurring errors. Unreviewed automation drifts.
- Budget the run rate. Count how often the trigger fires before you enable it. AI is usage-priced through IAP credits, so frequency is a cost decision.
Where this quietly changes everyday work
The obvious demos are document intake and ticket triage, but the quieter wins are real too. The AI accounting assistant can provide insights based solely on the accounting data in your database; scheduled reviews can surface questions for the finance team instead of waiting for month-end. The AI website assistant can create pages, edit snippets, add images, and build forms and menus, so a content workflow can prepare drafts a marketer approves.
Odoo 20 also adds a native MCP connector, letting an external AI system connect to your database within the user's existing access rights. Keep that separate: MCP is about outside tools reaching in, while agents inside automated actions are about Odoo's own processes continuing without a person opening a chat. Both matter, but they need different governance.
The limits to respect
An agent does not make a bad process good. If your stages are unclear, your required fields are empty, or your team disagrees about what urgent means, the agent will automate that confusion at a steady pace. Clean the model, the domains, and the definitions first. Also remember that stored conversations last up to 30 days, so review inside that window if you want the chat evidence.
Finally, do not confuse fluency with certainty. An agent's summary can read beautifully and still miss the one clause that mattered. Keep humans on the irreversible steps, keep triggers narrow, and keep reading the chats. Used that way, agents inside automated actions are Odoo 20's quiet superpower: judgement, placed exactly where your existing automations used to stop.
Further reading
- Odoo 20 Release Notes — Odoo — agentic automation, skills, and the IAP credit requirement.
- Odoo Pricing — Odoo — Agentic AI on the Custom plan, and IAP credits listed as not included.
