It's Tuesday and you're in the committee. Someone from sales says they use an AI assistant to draft proposals. Someone from support uses it to respond to tickets. No one requested it and no one approved it, but it works and saves time.
Then the official proposal comes: "let's activate AI within Odoo". It sounds reasonable. And right there it's worth stopping with three questions that are almost never asked before saying yes:
- Who pays for each query?
- What information about your clients leaves your system?
- Who decides what the assistant can see?
If any of these lacks a responsible party, you are not adopting a tool. You are accepting an expense without an owner and an open door with a good interface.
The problem: AI does not live on your server.
When someone asks an AI agent something in Odoo, the system constructs the query and sends it to a language model that is outside your company. The AI module of Odoo Enterprise currently connects with two providers: OpenAI and Google.
This has two consequences for the person signing the budget:
- The cost is per use, not per license. Each query consumes a billing unit from the provider. An agent that everyone loves costs more than one that no one opens.
- The risk depends on what is sent, not on what you have stored. Your database can contain everything. What matters is which fragment travels in each query.
The alternative to not deciding is not "not using AI". It's that your team uses it on their own, with personal tools, pasting customer information into windows that you don't see. Governing it within the ERP is the option with the most control, as long as it is configured wisely.
1. Decide who has the key
The API key is the link between your Odoo and the provider: whoever has it pays. Odoo allows you to enable your own key, both for OpenAI and for Google.
☐ Enable your own OpenAI key
☐ Enable your own Google key
With your own key, the consumption appears in your account with the provider. You see the actual spending, set a monthly cap, and revoke it in a minute if something goes wrong. Without it, you depend on the conditions that Odoo defines for that service.
Functional Pro-Tip
Use a different key for testing and another for production, each with its own spending limit. This way, a misconfigured agent in testing doesn't eat up the month's budget.
What management needs to do: request that AI spending has a cost center and a monthly cap from day one.
2. Control what each agent can read
An agent is defined by three things: its instructions, the model it uses, and its sources, which are the files and pages with which it responds. The instructions give personality. The sources are what protect you or expose you.
- Limit it to its sources. The agent responds only with what has been loaded, instead of improvising. Fewer invented answers in front of a client.
- Monitor the status of each source. Each document appears as in process, indexed, or failed. An agent with half of its database unindexed responds with half of the truth and with total certainty.
- Respect access. Each source calculates whether the user asking has access to it. Before loading sensitive documents, confirm in your own instance how far that restriction goes.
Functional Pro-Tip
Before loading a source, ask yourself: "Would I put it in an email to a third party?". If the answer is no, it is not yet a source. Start with manuals, policies, and catalogs. Contracts and personal data wait for someone responsible.
What is up to management: appoint a person responsible for approving which documents feed each agent.
3. One agent, one job
The most common cost error is a "one-size-fits-all" agent, with the most powerful model, open to the entire company. It's like asking the legal director to file invoices.
It is better to create small agents with a clear task: one for support questions, another for business answers, another for summarizing meetings. Each with the model that corresponds to it:
- Repetitive inquiries: a lightweight and economical model usually suffices.
- Analysis or delicate writing: justifies a larger model, with few agents and few users.
- Retired models: Odoo notifies when an agent's model has become obsolete and suggests its replacement. Review that notice, because an agent on a discontinued model stops working.
What management needs to do: demand that each agent has a business owner and a measurable outcome.
From intuition to decision
AI within the ERP is neither a threat nor a magic wand. It is a variable expense with access to your information, and it becomes profitable when it has three things: a key that you control, carefully chosen sources, and agents with measurable work.
Our recommendation is boring and that's why it works: a 30-day pilot, one agent, one team. At the end, you will have real consumption data, resolved inquiries without escalating to a person, and cost per inquiry. With that, you decide with your own numbers, not with someone else's demo.
Key Insight
Before approving, ask your team or your partner to respond in writing: whose key it is, what sources will be loaded, who approves them, what the monthly cap is, and how the outcome is measured. If any answer is "we'll see it later," it is still not time to activate the agent.
Do you want to learn about AI or consult with an expert in controlled AI in Odoo?
Talk to a Vauxoo consultant and define how to use AI in your operation without losing control of your data or your budget.
Talk to a consultant