“AI” is everywhere, and yet many small companies don't know what to do with it in practice. They hesitate between enthusiasm and suspicion. Both are legitimate.
What an AI agent actually is
An AI agent is a program that reads, understands and acts in your tools. It can read an e-mail, find information in your database, draft a reply, update a record, then ask you to approve. It doesn't replace a person. It takes away the tasks that pull them away from their real work.
Good first tasks
A good first task for an agent has four traits:
- It comes back often: every day or every week.
- It follows rules: you could explain it to an intern.
- It starts from text: an e-mail, a form, a document, a question.
- A mistake can be caught: a human can review before sending.
Typical examples: sorting incoming requests and drafting replies, summarising a meeting and creating the tasks, preparing a quote from an e-mail, answering internal questions from your own documents.
What to avoid at first
- Decisions that commit the company without human review.
- Rare tasks, where the agent would never learn your habits.
- Data whose hosting you don't control.
The safeguards
A useful agent is a well-framed agent. Before putting it to work, three questions:
- What may it do on its own, and what must it have approved?
- Where is your data, and who can access it? At Super Génial, applications are hosted by Infomaniak, in Switzerland, and the AI used is Claude, by Anthropic.
- How do you see what it did? Every action must leave a readable trace.
Start small, measure, extend
The method that works: a single task, a few weeks of trial, a simple measure (time saved, mistakes avoided). If it works, extend. If not, stop, having learnt something cheaply.
One last guideline: size. Below ten or so people, a task rarely comes back often enough for an agent to pay off. Start with a well-built internal tool instead, and the agent will come as the team grows.