AI Agents in Odoo: What You Can Wire Up Today, and What v20 Changes

Updated 9 September 2026: added the short answer, At a glance block, question headings, a comparison table of built-in versus external agents, the Odoo 19 permission finding from the Enterprise source, the current MCP specification date and “Goes well with” links. Facts re-verified against odoo.com, modelcontextprotocol.io, the ABS and the OAIC.

Odoo 19, the current production release, ships AI agents today: the read-only Ask AI assistant plus custom agents that act only through server actions an administrator has approved. External agents connect through Odoo’s per-user API, increasingly over the Model Context Protocol, and run with that user’s permissions. Native multi-step agents are Odoo 20 roadmap until Odoo Experience on 24 to 26 September 2026.

At a glance

• Who it is for: owners, CFOs and operations managers running or evaluating Odoo who want to know which AI agent features are real today and which are still slideware.
• Problem it solves: the gap between Odoo’s AI marketing, the Odoo 20 keynote hype and what an administrator can switch on or wire up this quarter.
• Platforms: Odoo 19 Enterprise (built-in AI app and agents); Odoo 19 Enterprise and Community (external agents through the API and third-party MCP servers); Odoo 20 (roadmap).
• What the built-in agents do: answer questions, find records, open views, draft text, sort documents and, for custom agents, call approved server actions.
• What external agents do: read records and draft quotations, notes and activities under their own user account; a person confirms anything that posts.
• Approval gate: Ask AI cannot write; custom agents write only through allowlisted tools; external agents write drafts that a named person posts.
• Audit trail: Odoo chatter, server logs and, on the third-party MCP servers we have reviewed, a per-call audit log and an agent badge on chatter messages.
• Typical cost: a single scoped agent pilot typically starts from a few thousand dollars ex GST for discovery and sandbox build, per our AI agents for ERP in Australia page.
• Typical timeline: 4 to 10 weeks from discovery to a drafts-first go-live, per the same page.
• What it never does: run as an administrator, confirm an order, post a journal entry or change a price or stock level without a person.
• Last verified: 9 September 2026 against the Odoo 19 AI documentation, the MCP specification, 28 July 2026 revision and the Odoo Experience 2026 event page.

Odoo now ships real AI agent features, Odoo 20 is about to promise bigger ones, and most of what’s written about either comes from people who haven’t opened the source code. We implement Odoo for Australian businesses and run AI agents against it in production, so this is the practitioner’s version: what you can switch on in Odoo today, what you can wire up alongside it, and what to leave alone until it ships.

First, a working definition, because AI agents in Odoo means two different things depending on who is selling. An AI agent in Odoo is software that uses a large language model to do useful work in your database: answering questions about your records, drafting content, sorting documents and, within strict limits, acting through tools an administrator has approved. In practice that covers two different things: the agent features built into the product, and external agents you connect through Odoo’s APIs. The capability is different and so is the risk, so it pays to be precise about both.

The demand side isn’t speculative. MYOB’s research found 78% of wholesale distribution businesses plan to use AI inside their ERP, and wholesale is exactly the kind of business Odoo wins in Australia. The question we hear in discovery calls has shifted from “should we?” to “what’s actually real?”

What AI agents does Odoo 19 ship today?

Odoo 19 includes a dedicated AI app, covered in the official Odoo AI documentation. We’ve checked what follows against the v19 Enterprise source code, because vendor marketing and shipped software aren’t always the same thing.

Ask AI is the headline: an assistant available across the database that answers questions in plain language, finds records and opens the right views. Its most important property is what it can’t do. The standard Ask AI agent cannot change your data. It will show you the report; it won’t post the journal entry.

Custom AI agents sit one level up. You create an agent, give it a system prompt and a response style, choose which large language model runs it, and attach knowledge sources it can search, like your product documentation or internal procedures. Then you give it topics: instruction sets with tools attached. The models on offer are Gemini and OpenAI, with your own API keys required on Odoo.sh and on-premise databases and optional on Odoo Online.

The tool design is the detail worth understanding, because it’s where Odoo quietly got the governance right. An agent’s tools are Odoo server actions, and only the ones an administrator has explicitly flagged as available to AI. The agent can’t grant itself new abilities. If nobody hands it a “schedule a follow-up activity” action, it can’t schedule one. That’s an allowlist, which is how agent permissions should work, and it’s the same principle we apply to every agent we build regardless of platform.

There is a second permission detail we found in the source and have not seen written up elsewhere. When a custom agent runs a tool, the tool’s code executes with the rights of the person chatting with it, not as an administrator. A public website visitor gets visitor rights; a portal customer gets portal rights. That is the correct default. The catch is that a tool’s own code can lift itself to administrator rights, and the lead-creation tool Odoo ships does exactly that so a visitor can leave their details. Anyone who copies that pattern into a “look up my order” tool has just made every order readable to anyone who types a number. Our build review now includes a search of every AI tool for an escalation call, and we suggest yours does too.

The same app also brings the assistive features: AI fields that fill themselves from context, AI server actions for automations, document sorting and data extraction, livechat agents for the website, AI drafting in email templates, and voice transcription. Our guide to AI in Odoo for Australian businesses walks through those feature by feature.

Two caveats. The AI apps are part of Odoo Enterprise, so Community sites don’t include them, though the API route below works for both editions. And what ships today is assistive: it answers, drafts, sorts and suggests. A native agent that runs a multi-step process across modules end to end isn’t in v19. That’s the v20 conversation.

Can external AI agents work with Odoo through the API?

Yes, and most of the practical agent value in 2026 comes from this second layer: an agent that runs outside Odoo, powered by whichever model you choose, working through Odoo’s external API under a normal user account. The Model Context Protocol has become the standard way to make that connection, donated by Anthropic to the Linux Foundation’s Agentic AI Foundation in December 2025 and now on its 28 July 2026 revision. Odoo SA does not publish an MCP server of its own as at September 2026; the connectors come from the community and the Odoo app store, where the better ones scope each key to read-only or read-and-write, run every call as the key’s owner and log it. We build ours against the API directly.

Here’s the worked example we’re asked about most. A wholesale business receives purchase orders as PDF attachments, in a different layout for every customer. The agent watches the inbox, reads each PDF, matches the customer record, checks products and pricing against the price list, checks stock, and creates a draft quotation in Sales tagged for review. A salesperson opens the draft, corrects anything odd and confirms it. The agent never confirms an order itself. When it misreads a smudged quantity, and eventually it will, the mistake dies at the review step instead of shipping. Our AI sales order agent post has the full spec.

Two properties make this safe enough for a sceptical CFO to sign off:

• The agent has its own user account. Odoo enforces access rights on API sessions, so the agent sees exactly what that user role sees. No administrator login, ever.
• Writes arrive as drafts. Anything touching the ledger, tax codes, pricing or stock is proposed by the agent and posted by a person. The approval gate is the design, not a training-wheels phase.

This is the architecture from our broader guide to AI agents for Australian businesses, and Odoo is a good host for it: the API is mature, permissions are granular, and everything the agent does is inspectable.

Does Odoo 20 change this, and what is still roadmap?

Odoo 20 gets unveiled at Odoo Experience in Brussels, 24 to 26 September 2026, with general availability expected after the event. Agentic AI is the expected headline: agents that execute multi-step work across modules natively instead of assisting inside one screen. We track what’s confirmed versus expected in our Odoo v20 roadmap breakdown.

Until it ships, it’s a roadmap. Nothing announced at a keynote should drive a 2026 implementation decision, and anyone selling you native “Odoo agents” as a present-tense product is describing slideware.

“The mistake we keep seeing is businesses pausing everything to wait for Odoo 20’s agents. The API is the agent surface today. Wire the workflow up now, run it drafts-first, and when v20’s native agents mature you’ll already have the data discipline and the approval habits they’ll need.”

Josh Craig, Director, Auboros

Our advice hasn’t changed: implement on the current version now, treat v20 as an upgrade decision once it’s real and stable, and let external agents cover the gap in the meantime. The work transfers. Clean data and approval workflows are exactly what v20’s agents will need too.

How do you run Odoo AI agents safely in Australia?

Adoption is accelerating. The Australian Bureau of Statistics reports 12% of Australian businesses used AI in 2024-25, up from 1% two years earlier, and most of them have no written rules about what the AI can touch. That gap is what turns a useful agent into an incident.

The Australian framework is workable. The National AI Centre’s Guidance for AI Adoption sets out six voluntary practices, including human oversight and accountability. The Privacy Act is not voluntary: customer records in your CRM are personal information, and the OAIC publishes guidance on using commercial AI products. From 10 December 2026, privacy policies must also disclose the kinds of substantially automated decisions that significantly affect people. If an agent helps decide credit terms, that needs documenting before the deadline, not after.

For an Odoo database specifically, safe operation comes down to five controls:

• Role-scoped access. The agent runs as its own user with the narrowest access rights that still let it work. Odoo’s permission model does the enforcement if you let it.
• Approval gates on writes. Ledger, tax, pricing and stock changes are drafted by agents and posted by people.
• An audit trail. Log what the agent read, what it proposed, who approved it and when. Odoo’s chatter and server logs make most of this cheap.
• Sandbox first. Prove the agent on a staging copy of your database before it sees production.
• A named owner per agent. Someone reviews its permissions and output quality on a schedule, so access never quietly grows. Our AI agent governance post sets out the policy that goes around these controls.

Where to start with AI agents in Odoo

Start with the workflow that burns the most hours and has an obvious human checkpoint: inbound order entry, reconciliation exceptions and supplier chasing are the proven first picks. Fix the data the agent depends on, run it against a staging copy until its drafts are consistently right, then go live with the approval gate permanent. Add a second agent only once the first has an owner, a budget and a clean month behind it.

None of this needs to wait for a keynote in Brussels.

Goes well with

• AI agents for ERP in Australia: which agent to build first over your Odoo database, what a drafts-first pilot costs and how the approval gates work.
• AI in Odoo for Australian businesses: the assistive features in the Odoo 19 AI app, one by one, and which are worth switching on.
• The AI reconciliation agent: the second workflow most Odoo finance teams hand to an agent after order entry, and how it protects the BAS.

Thinking about AI agents for your Odoo database?

Auboros is an Odoo partner in Brisbane that designs and runs AI agents over Odoo, drafts-first with human approval, for businesses across Queensland, NSW and Victoria. If you want to know what an agent could safely take off your team’s plate, book a free consultation. If the honest answer is “fix the data first”, you’ll hear that instead.