AI agents that work alongside you in your Odoo.

An agent prepares orders from email, WhatsApp or phone in Odoo, proposes journal entries, builds dashboards and answers questions about your data. Your team reviews and approves.

Schedule an introductory call

Which agents we have already built

Orders from email, WhatsApp and phone

Many orders come in as an email with a PDF, a WhatsApp message or a phone call. Someone then retypes them into Odoo. An agent reads that message, looks up the customer and checks that customer's order history. That way it works out faster what is meant by "the same as last time" or "the usual boxes". It then creates a quotation or sales order in draft. For a phone call, it starts from the transcript of the conversation.

If the agent is unsure about a product, a quantity or a price, it flags that on the order line. Your back office reviews the draft and confirms. The retyping disappears, the decision stays with your team.

Accounting: more than reading invoices

Odoo already reads vendor bills. An agent goes further: it proposes a journal entry based on how you booked that vendor before, with the account, the analytic account and the tax code. Your accountant approves the proposal or adjusts it.

It also looks for discrepancies. An invoice that differs from the order or the receipt, an invoice that came in twice, a bank transaction without an invoice, a tax code that differs from what you usually use. You get that list before your closing, with the journal entries attached to each discrepancy.

Dashboards and questions about your data

You ask in plain language what you want to see: revenue per customer per month, margin per product category, customers who ordered less this year than last year. The agent builds that as a spreadsheet in Odoo with live data, or finds the answer and shows where it comes from. A dashboard is a regular Odoo spreadsheet that you filter, adjust and share yourself.

Before you get a dashboard, we check the figures against the source in Odoo. The agent works with the access rights of an Odoo user: what that user is not allowed to see, the agent does not see either.

Building your website

An agent helps build your Odoo website: setting up pages, reworking texts, aligning the layout across all pages. You review and approve before anything goes online. What is online remains a regular Odoo website that you edit yourself in the editor.

In Odoo itself or from your own subscription

Odoo has AI built in. We configure it directly in Odoo, with an API key from OpenAI or Google Gemini. The AI then works in your regular screens: in the chatter, in fields and in automated actions. You pay for usage to the provider via API credits. Depending on the application, that stays limited.

If you already work with Claude or ChatGPT, you can also ask your Odoo questions and have dashboards built from that existing subscription, with the access rights of your Odoo user.

How it works

We start with the flow where your team retypes or searches the most today. We test the agent on real messages and journal entries from the past weeks, together with the people who do that work now. Only when they are satisfied with the results does the agent join your daily operations.

A message is not an instruction

An agent that reads emails and WhatsApp messages also receives messages from people with bad intentions. Such a message can contain hidden instructions, a forged sender or a forged approval. What an employee sees on their screen is then not always what the agent reads. Approval by your team alone is therefore not enough.

That is why an agent only gets the access rights its task requires. The agent for orders creates drafts, but confirms nothing, sends no emails and does not see your accounting. It sends nothing back to the sender and opens no links from messages. If a message does make the agent do something unintended, it can do little damage with those rights.

With every draft, Odoo shows the text as the agent read it, including what was hidden in an email. Every action of the agent is recorded in the document's history. Before the go-live, we also test the agent with messages that try to mislead it. Nobody can rule out every risk. What we do: limit what a misled agent can do.

Governance: an agent earns its mandate

Who gave that agent its mandate, who checks its work and what happens if it makes a mistake? We record those answers before an agent touches your customer data, invoices or orders.

An agent does not get its permissions all at once. For each task, we define a mandate ladder. It starts by flagging and proposing. If those proposals prove correct over an agreed period, it may create drafts that your team validates. Only after that may it execute on its own, and only within limits you define: a known vendor, a maximum amount, a fixed account...

We also define the human validation step per task and per amount. A proposed journal entry above a threshold goes to the accountant, a new vendor or a changed bank account number to the CFO. In Odoo, such a document waits for approval in its own status. The agent does not approve its own work, and whoever configures the agent does not validate its output. The segregation of duties you have today stays intact that way.

Then the correction loop: Odoo tracks what the agent proposed, what it eventually became and who changed it. That way you see per agent and per task how often it gets it right. Those figures determine whether it may move up a step. If corrections increase, it drops back to proposing until we know the cause. Your accountant or auditor sees in the history of every document what the agent did, on whose mandate and who validated it.

What you had better not give to an agent

An agent that processes orders or proposes journal entries does not, as a rule, fall under the strict rules of the European AI Act. An agent that judges people does. Some applications are even prohibited.

Prohibited since February 2025: inferring emotions in the workplace, for example from an employee's voice on the phone or from a job interview on video. Disadvantaging people based on a score for their behavior, or classifying them via biometric data by, for example, origin, religion or sexual orientation, is not allowed either. An agent that processes phone calls therefore works with what is said, not with how it sounds.

Resumes are the clearest example of high risk. An agent that filters applications, ranks candidates or gives a score falls under the strictest rules. The same applies to an agent that helps decide on promotion or dismissal, assigns tasks based on behavior or personal characteristics, or monitors employee performance. Such an agent profiles people, and then it remains high risk, even if it only prepares. Copying data from a resume into a candidate's record can be a limited, administrative task, as long as the agent does not filter, rank or assess. We record that assessment in writing.

Assessing the creditworthiness of individuals is high-risk too. Proposing a credit limit for a company does not fall under this, but for a private individual or a sole proprietorship it does. Fraud detection is exempt.

For high-risk systems, the rules apply from December 2, 2027. Anyone deploying such an agent must then organize human oversight by people with the right authority, keep the logs for at least six months and inform employees and their representatives in advance. The GDPR already applies today: rejecting a candidate on the basis of a fully automated decision is not simply allowed.

So for resumes, employee evaluation and credit decisions about individuals, we do not build an agent that filters, ranks or assesses people. If you still want to use AI there, we first look at what that requires together with your legal advisor. We do not give legal advice ourselves.

What it costs

The cost depends on the sources the agent reads, the Odoo apps it works with and the checks you want. After a short analysis, you get a price for the build. On top of that comes the usage of the AI model, through API credits or your existing subscription. That depends on your volume, and we confirm an estimate in advance.

Does the agent confirm orders or journal entries itself?

Not at the start. An agent begins by making proposals and only gets a wider mandate when its results warrant it, within limits you set. How that works is explained above under governance.

What if the agent makes a mistake?

That happens, just as with a colleague. That is why the agent creates drafts, flags where it is unsure and shows where an answer comes from. We use your team's corrections to adjust the agent.

What happens to our data?

The agent sends the text of a message or a question to an AI model, and only what is needed for the task. If the AI runs in Odoo, that text goes to OpenAI or Google through your own API key. With paid API access, they do not use your data to train models. If you work from your own Claude or ChatGPT subscription, the terms of that subscription apply: with a business subscription there is no training on your data, with a personal subscription you check that setting yourself.

Does this also work on an Odoo that someone else implemented?

Yes. The agent works with the data that is already in Odoo: customers, products, orders and journal entries. We first look at your environment and then tell you what is feasible.

Does the agent talk to my customers?

The order agent does not: it sends nothing back to the sender. If you want an agent that replies to customers, the customer knows they are talking to an agent. Such an agent gets stricter limits, because it can send data outside your company.

What does the European AI Act require from us?

For most agents in an SME, it comes down to transparency and oversight: knowing which agents you use, what they do and who checks them, and teaching your people to work with them. The mandate ladder and the history per document are a good basis for that. Agents that judge people, such as with resumes or creditworthiness, fall under stricter rules. What you should not hand to an agent is explained above.

Who owns an agent's instructions and limits?

You do. We document what an agent does, which permissions it has and which mandate it was given. You can review and change that, even without us.

What if the AI provider goes down or we want to switch?

If the provider goes down, the work returns to your team, as it is today. Because the agent works with drafts, nothing is left half done. If you switch provider or model, we test the agent again on the same messages and journal entries before it continues.

What if the volume grows?

Then the usage of the AI model grows with it. You set a budget limit or an alert with the provider. We also show you what a processed order or journal entry costs, so you can weigh that against the work your team does today.

Does my team need training for this?

Briefly, but yes. Your key users learn to validate a draft and to enter a correction properly, because those corrections adjust the agent and determine whether it gets a wider mandate. We include that in the support.

Where does your team do the most retyping today?

Schedule an introductory call, online or on site at your place. Or call us at +32 53 89 50 21.