"Agentic AI" is the buzzword in ERP land in 2026. Software vendors promise agents that run your entire administration while you are on holiday. The reality is more nuanced. Some things already work surprisingly well today; others you had better not hand over just yet. A level-headed overview, without the sales pitch.
What already works
Today, AI agents are reliable in tasks where they read, summarise and prepare. Think of:
- Looking things up and summarising. "Which invoices are outstanding, what was discussed with this customer last month?" An agent searches your Odoo data faster than you can set a filter.
- Preparing drafts. A draft quotation based on a previous order, a draft reply to a customer question, a proposal for a purchase order. The human approves, the agent does the typing.
- Flagging. Deviating margins, customers going quiet, stock dropping below the minimum. Agents are tireless at the kind of checking work that people let slide after week three.
The pattern is always the same: the agent does the groundwork, a human takes the decision. That is not a limitation but a design choice, and for now a sensible one.
What you should not automate (yet)
There are also tasks you had better not let an agent loose on independently in 2026:
- Autonomous postings. An agent that independently creates and confirms journal entries sounds efficient until the first wrongly posted VAT entry. Let agents prepare postings, not finalise them.
- Pricing decisions. Discounts, price changes and payment arrangements have a direct impact on your margin and your customer relationship. That remains human work.
- Irreversible actions. Deleting records, cancelling orders, sending emails to customers without review. Anything you can't undo with a single click deserves a human checkpoint.
The reason is not that the technology is stupid. Language models can make mistakes, and do so with great confidence. In a summary that is annoying; in a posted invoice it means a correction entry and a conversation with your accountant.
Veilig beginnen: leesrechten eerst
The sensible route is the same as with a new employee: first observe, then assist, and only then work independently. In concrete terms:
- Start with read-only rights. Let the agent answer questions and report, but change nothing. Through an MCP integration such as DAADit's free hosted MCP server you set those rights per user.
- Make sure there is an audit trail. Every action by the agent must be traceable: what was retrieved, what was created, and when. In Odoo the chatter helps with that; per record you can see who or what changed something.
- Keep a human in the loop. Let the agent create drafts that someone explicitly approves. Only once a task has run flawlessly for months should you consider going a step further.
Praktische tip
Make a simple list with two columns: "the agent may do this itself" and "the agent may only prepare this". Start with everything in the second column and only move tasks across after demonstrably good results. That one sheet of A4 is your entire AI governance, and for most SMEs that is exactly enough for now.
Want to talk through what agentic AI could mean for your Odoo environment? Contact DAADit Group for an honest answer, even if that answer is "wait a little longer".