AI that does the work, not just the talking

With a process you already run, not a technology you want to try. The best first candidates are high-volume, rule-shaped tasks where the data already exists in a system: reading incoming documents, triaging enquiries, drafting routine responses. Those deliver time back quickly and teach you what your data is really like, which is what makes the next project better scoped.

It depends entirely on the setup, which is why we treat it as a design decision rather than a default. Options range from provider APIs with no-training guarantees through to models running in your own environment, and most systems only need a narrow slice of data to reach the model at all. We agree what may leave your network before anything is built.

Assume it sometimes will and design for it. Extracted data gets validated against the records you already hold, confidence thresholds route uncertain cases to a person, and every automated action is logged so it can be traced and reversed. An automation that fails loudly and lands in a queue is far more useful than one that fails quietly.

Yes, and that is one of the more valuable places to put it, because Odoo already holds the structured data an agent needs to be useful: customers, products, stock, orders, invoices. Agents read and write through Odoo's API with a dedicated user and scoped permissions, so their access is auditable and limited to the records they should touch.

For the work described here, no. This is systems engineering: connecting models to your processes and data, handling failure cases and keeping humans in the loop where it matters. If a problem genuinely needs custom model training rather than integration, we will say so rather than dress up the wrong approach.