Skip to Content
Q&A

AI colleagues: how we build hiring, the portal view, usage per customer and the incident list

Four Odoo modules around one question: who can see what an AI colleague does?

Back to all Q&A
November 4, 2026 by
AI colleagues: how we build hiring, the portal view, usage per customer and the incident list

In short: We build our AI colleagues as something you hire, not something you configure. Four modules carry that idea: hiring and setup, a view in the customer portal, usage per customer, and one list of everything that goes wrong. This article describes how we built them and what you can take from that for your own plans. It describes how we work and is not a product offer.

Two employees and three friendly robots with name badges in an office, one employee shaking hands with a robot

When you let AI do work in your company, the practical questions follow quickly. Who gave it access? What did it do today? Who pays for which usage? And who notices when it goes wrong? We built four modules for Odoo 19 around those questions. What Odoo itself offers is covered in our article on AI agents in Odoo 20. This article is about what we build ourselves.

What we mean by an AI colleague

An AI colleague is an AI agent that you hire the way you would hire a temporary worker. After the order, it sets itself up in your Odoo with your company details, the rights you give it, a fixed budget and a work schedule. It sends a daily update, makes proposals that you approve or reject with one click, and asks first when it cannot reach something.

Hiring: set up completely, or not at all

The hire is a record of its own, with one running hire per customer, colleague and connected environment. A second attempt reuses the existing hire instead of creating another one.

Each customer gets its own copy of the colleague, with that customer's list of models, privacy list, read scope, budget and schedule. All of this is checked beforehand. Is the connection missing, is an app absent, or is it a production environment while the colleague needs to change data? Then setup is refused, with the reason, and nothing is left half finished.

Stopping is just as tidy. The schedule goes off, the copy is archived and nothing is removed at the customer. A scheduled job follows the subscription, so ending it really does stop the colleague.

What a colleague may do

  • Read: only through the list of models, the read scope and the privacy list of that copy. Users, keys and settings are on a hard refusal list.
  • Create and update: only with a signed mandate that covers that model and that type of action, within a daily ceiling and a validity period.
  • Delete: not included.
  • Post, pay or submit: remains a human step. Financial records stay in draft.
  • Accountability: every attempt, including a refused one, ends up in a trail that a person reads: what, for whom, which records and why.

The portal: a mirror, not a control panel

The portal module puts hired colleagues next to the customer's orders and invoices. Per colleague the customer sees whether it is running, on which environment, since when, what it did today and, if something went wrong, why and what the next step is. The status is in plain language.

We deliberately left out buttons that change the contract. Stopping or cancelling stays with the subscription and with a person. The introduction and onboarding of a new colleague also take place here.

Usage per customer: no second price model

Our own database serves several customers. Without extra work, all AI usage would be one total, and you cannot explain one total to a customer. The usage module links an agent, and if needed a single schedule, to the customer it serves. Usage is then attributed to that customer in the same work units that the meter and the invoicing already use.

What cannot be attributed to anyone stays visible as a separate residual item. It is not quietly spread over customers.

The customer sees its own usage in the portal, in the same units as on its invoice. Per customer there is an included bundle, an overage policy and an optional hard stop. Measuring always stays on.

One list for everything that goes wrong

The incident module replaces scattered signals with one list, with four kinds of entry:

  • Broken: a run was aborted or a tool keeps failing.
  • Made up: a report claims work that left no recorded effect.
  • Access: a hired colleague hit a boundary, such as a model outside the connection. This can also be seen in advance, and on an approval screen the customer decides whether the colleague may have access.
  • Opportunity: the colleague saw work for a colleague that has not been hired at that customer. This is not a defect and does not count towards the errors.

Per kind, agent, hire and model there is one incident with a counter. A notification only goes out when an incident is new or reopened.

What this means for you

Whichever way you use AI agents, these steps are worth taking before one starts working in your administration:

  1. Decide who grants rights, per environment. Treat reading and writing as two separate decisions.
  2. Keep financial actions with a person. The agent prepares, an employee posts.
  3. Agree where usage is visible and what happens at the limit of the budget.
  4. Make one person responsible for the list of incidents.
  5. Test what happens when the agent is refused. It should say why, not stop silently.

We designed around two pitfalls in particular: a setup that stops halfway, and a single usage total that nobody can explain. Check how your own plans deal with both.

This article says nothing about price, delivery time or availability. Whether any of this suits your situation is something we discuss in a conversation. How we use AI tools in general is on our AI tooling page, and examples of processes are under AI automation.

Frequently asked questions

Can an AI colleague delete records or post financial records?

No. Deleting is not included, and posting, paying or submitting remains a human step. Creating or updating requires a signed mandate for that model and that type of action.

What does the customer see of an AI colleague?

In the portal the customer sees per colleague whether it is running, on which environment, since when, what it did today and, if something went wrong, why and what the next step is.

What happens when an AI colleague cannot reach something?

It asks first. In the incident list this appears as an access incident, and on an approval screen the customer decides whether the colleague may have access.

Finally

An AI colleague only works if someone can see what it does, what it uses and where it gets stuck. That is what these four modules are for. Want to discuss what this could mean for your company? Contact us for a conversation. Prefer to look at your own processes first? Take the free Business Scan.

in Odoo

Question not answered?

Our Odoo consultants are happy to think along with you. Book a no-obligation intro call.

Book an appointment
Odoo for startups and scale-ups: a growth programme from €1,000 per app, or a monthly plan
Why fast-growing companies outgrow Excel and loose tools, what a first Odoo setup really costs, and how DAADit helps startups and scale-ups go live in weeks