GUIDE FOR SMES

How to install a local AI agent in a business?

A local AI agent is not just about installing a model. You need to choose useful tasks, protect data, organise rights and plan for backups and rollbacks.

1. Start with a measurable use

DCE analysis, searching in technical folders, minutes, administrative classification or action tracking: the first scope should be defined. You measure the time saved, the oversights avoided and the quality of the responses.

2. Audit the machine and network

The inventory checks processor, memory, graphics card and VRAM, disks, BIOS, network, encryption and backups. An existing machine can often be used for the pilot. The final choice mainly depends on the local model and the number of simultaneous processes.

3. Separate system, data and backups

The system and models, the company's documents and the backups should not depend on a single volume. An encrypted external copy remains necessary, as a disk in the same machine does not protect against theft or a general failure.

4. Install a maintainable foundation

An Ubuntu LTS distribution, containerised services, a firewall, limited remote access, logs and controlled updates provide a reproducible base. Secrets remain in a secure vault.

5. Organise documentary memory

Authorised documents are indexed by project and by rights. Responses must cite their source. Data from different companies or services remain separate, and deletion should be possible.

6. Keep humans in control

A summary can be automatic; sending, deleting or making a consequential decision should follow the planned validations. Governance defines who can do what, the agent's limits and the records kept.

7. Deploy progressively

After a pilot with a few users, the results are checked before adding new workflows. Significant developments remain separate projects from the foundation's maintenance.