← Blog

Use cases · 9 November 2026

One AI Agent per Department, in Teams, That Learns and Doesn't Forget

Un agente de IA por departamento, en Teams, que aprende y no olvida

If you run a small or mid-sized business, you know the scene: someone is looking for the latest version of a document, another person interrupts a colleague to ask how something is processed, and there is a procedure that only the person who is on holiday knows how to do.

The information exists. The hard part is finding it, and making sure that what people learn isn't lost. This is what we have set up at Haz Energía, and they use it every day.

One agent for each department

Desde Microsoft Teams, cada persona llega al agente de su departamento: Licencias, Ingeniería, Puntos de acceso, Subvenciones, Informática o General, con SharePoint según sus permisos

Nobody has installed anything: the assistant is just another contact in Microsoft Teams. The Licensing agent speaks the language of Licensing and doesn't mix its conversations with those of Engineering.

Adding someone is a matter of adding their email address to the department's list. The rest of the company has a basic agent for general questions.

SharePoint: what you can see, and nothing more

The agent acts in SharePoint with the permissions of the person writing to it, and that is enforced by code, not by an instruction to the model.

  • It finds documents by name or by content.
  • It reads Word, Excel, PDF and photos, including scanned ones.
  • It creates documents from your templates and uploads them alongside the rest.
  • It doesn't delete, doesn't move and doesn't change permissions.

It can also work with Microsoft 365 email, monday.com or GitHub, but each integration is granted deliberately.

It learns, remembers and doesn't make you repeat yourself

Cómo aprende un proceso: se lo enseñas, pregunta lo que falta, te lo resume, dices sí y lo aplica, y tras revisión humana llega a toda la empresa

Most assistants start from zero in every conversation. With this one, every conversation can leave something saved: a fact, a preference, a process or a correction. And what is saved never needs to be explained again.

What it learns

How

What you get

Facts and preferences

It saves them in the department's memory

It remembers how you like a report or which criteria you apply

Processes

You teach it once and say “yes”

Next time you tell it “handle this” and it already knows how

Corrections

It updates that same process, as a new version

It improves with use, with no contradictory copies

Each area's knowledge

It is proposed daily to the shared catalogue

After human review, every agent has it

What your most experienced person knows stops living only in their head.

And what it doesn't learn: it doesn't save a process without your “yes” (12 correct out of 12 in our tests), it doesn't save passwords or customers' personal data in processes, and nothing reaches the rest of the company without a person approving it.

Tests before every update

Before deploying, the system tries to make the agent reveal secrets and subjects it to manipulation attacks. If a critical test fails, nothing is deployed. It is built on Hermes Agent, an open-source project created by Nous Research; we add Teams, departments, permissions, learning with review, and maintenance.

When we wouldn't recommend it

  • You don't use Microsoft 365. There are other channels, but it would be a different project.
  • You're looking for something that needs no supervision. It can make mistakes, and someone has to review what it learns.
  • Two areas need strict separation. That requires deploying an independent agent.
  • You're three people at the same table. It probably isn't worth it yet.

Start with one department

Do the math: people × minutes a day spent searching or asking × working days × cost per hour. With 10 people and 20 minutes a day, as an example, that comes to more than 70 hours a month.

Pick one department and one process that hurts today. We'll analyse it with you and tell you whether an agent makes sense there or whether something simpler would do.

👉 Let's talk · AI agent platform

← Back to the blog