Artificial intelligence

Odoo 19 builds artificial-intelligence assistance into several places in the software. These are not variants of one and the same feature: they have neither the same prerequisites, nor the same degree of autonomy, nor the same consequences when they get it wrong. This chapter tells them apart before detailing the ones that are configured.

The four levels of assistance

The gap between the last two is the one that matters: an agent proposes an answer that a human reads, a server action applies a decision that nobody sees go by. The precautions are not the same.

The prerequisites

Three distinct conditions, to be checked in this order.

One clarification first, because it explains what follows: the assistance rests on two modules both named AI, and they do not install the same way.

In other words, a database may perfectly well offer writing assistance without any AI menu being visible: the foundation is there, the application is not.

The PostgreSQL extension

Agents index their reference documents so as to be able to refer to them. That indexing relies on the PostgreSQL extension vector, which the server must supply. Odoo checks before installing: the assistance foundation flatly refuses to install if the extension is missing. On automatic installation it abstains with no message and no visible trace; on manual installation the operation fails with an error.

The provider access key

Odoo does not host a model: it calls an outside service (OpenAI or Google) and passes it the elements needed for the request. That call requires an access key, whose conditions depend on the hosting. The procedure is in section 3.3.4.

The application

The writing-assistance features are used without installing anything. Configuring agents, topics or default prompts, on the other hand, requires adding the application, which never arrives on its own.

1. Open Apps.
2. Search for AI. The application presents itself as an ordinary application, with no need to remove the default filter.
3. Click Activate.
4. An AI menu appears, carrying the Agents entry. The Configuration entry, which gives access to the access keys and the default prompts, is only visible to a settings administrator.

The agents

An agent is an interlocutor configured for a specific use. It differs from the generic assistant in three ways: a permanent instruction that frames its answers, reference documents it relies on, and topics that give it tools.

Creating an agent

1. Open AI › Agents › Agents.
2. Click New.
3. Fill in the Agent Name, then the subtitle placed just below it, which sums the agent up for its users. This field has no label: it is recognised by its example text.
4. Choose the LLM Model from those of the configured providers.
5. Set the Response Style: Analytical, Balanced or Creative.
6. Write the System Prompt: the permanent instruction that frames the role and the tone.
7. Save.

Giving it reference documents

The Sources tab gathers the documents the agent relies on: catalogue, terms and conditions, internal procedure, website page.

1. Open the agent's Sources tab.
2. Click Add a source and choose a file or a web address.
3. Wait for the indexing.

Each row carries a status:

The Reprocess button re-indexes a source already processed: it is for web addresses, whose content changes without Odoo being told. An attached file does not need it, its content being fixed.

The topics

A topic is a skill: a set of detailed instructions together with the tools needed to exercise it. It is then attached to one or more agents, which avoids writing everything in the system prompt and allows the same skill to be reused elsewhere.

Odoo ships some. The list opened by AI › Agents › Topics is not empty on a fresh database: the installed modules deposit their topics there, with their instructions and their tools already attached.

Opening one of them shows what a finished topic is: several pages of instructions, setting out the rules to respect, the expected formats, the forbidden sequences and the fallback cases , and a dozen attached tools, of the AI: Search or AI: Get Fields kind.

Creating a topic.

1. Open AI › Agents › Topics.
2. Click New.
3. Fill in the Title, then the unlabelled subtitle placed below it, which sums the skill up in one line.
4. Write the Instructions: the procedure to follow, in detail.
5. Designate the AI Tools the topic makes available. Odoo ships a dozen, and any server action ticked Use in AI is added to them , see section 20.5.2.
6. Save, then attach the topic to the agent you want through its Topics field.

Putting the agent to the test

The Test button, at the top of the form, opens a conversation with the agent as it is configured. It is the way to check the behaviour before exposing it to users: rephrase the same question in several ways, check that it does cite its sources, verify that it refuses what falls outside its scope.

The default prompts

When a user triggers assistance, Odoo calls an agent and passes it an instruction. Neither is fixed: they depend on where the request comes from, and that pairing is configured through rules.

Each rule answers a question of the form "when users do this, which agent should be addressed and with what instruction". Odoo ships seven, one per entry point in the interface: eight if the Knowledge application is installed.

1. Open AI › Configuration › Default Prompts.
2. Open a rule, or click New.

The form holds five fields and one tab:

The eight entry points:

AI-assisted automation

Some decisions cannot be expressed as a fixed condition: classifying an incoming request by its content, choosing a treatment based on free text, extracting a piece of information from a message. A server action assisted by artificial intelligence handles those cases.

The principle

The mechanism keeps the decision and the execution strictly apart.

Creating a tool

1. Activate developer mode.
2. Open Settings › Technical › Actions › Server Actions.
3. Create the action performing the operation you want, or open an existing action.
4. Fill in the Model it operates on.
5. Choose the Type from the tiles offered. The choice is not free: only certain types can serve as a tool, see the table below.
6. Open the Usage tab and tick Use in AI.
7. Fill in the free text field that appears to the right of the box, it carries only the example text Short explanation of what this does. It is on that description, and on it alone, that the decision action will choose to call this tool.
8. If the type is Execute Code (or Multi Actions chaining a code action), fill in the AI Schema: the JSON-format description of the expected arguments, their type and their constraints. Leave it empty if the tool expects no argument.
9. Save.

Two types are eligible under conditions: Update Record, if the value comes from an entered value or a sequence and not from a computed expression, and Multi Actions, on the sole condition that all the actions it chains are themselves eligible.

Creating the decision action

1. Still from the server actions, create an action and choose the Type AI.
2. Select the Model concerned.
3. Open the Prompt tab, which only appears on this type of action.
4. Fill in the Tools field: the actions the decision is allowed to call. The list offered is limited to the actions ticked Use in AI and operating on the same model. A tool defined on another model will not appear.
5. Write the instruction in the AI Action Prompt field: the context, the decision expected, the criteria to apply.
6. Save.
7. Finally, attach this action to a trigger, through an automation rule (see chapitre 21) or a button on the document.

Writing the instruction. The field is a rich text editor, with two commands of its own, called up with / as in email templates:

• Field Selector, subtitled Insert a field: inserts the value of a field of the record being processed. That is how the instruction stays generic while relying on real data.
• Records Selector, subtitled Insert records: designates specific records the decision will be able to consult: a rate table, a reference list.

What you write is not all the model receives. Odoo automatically adds the current date and time, the identity of the record being processed, and the values of the fields your commands inserted. There is therefore no point restating the date in the instruction.

What happens at execution. The action processes the records one by one. For each, it submits the enriched instruction, receives the list of tools to call and executes them in order. The cycle stops when the decision judges it has nothing left to do.

Two checks block the saving of an inconsistent configuration: ticking Use in AI on an ineligible type refuses the save with the message The action '…' cannot be used as an AI tool., and an AI Schema that is not valid JSON is rejected likewise.

What the assistance does not do