News · 19 November 2026

AI agents: from chatbot to digital co-worker.

For a long time, AI in a company meant a chat window: you ask a question and get an answer. Agents go one step further. They are given a goal, plan the necessary steps themselves and carry them out in the connected systems. That changes which work can be automated, and it places new demands on technology and organisation.

Engineer in front of two monitors showing a workflow graph of connected AI agents

What separates an agent from a chatbot

A chatbot answers. An agent acts. It breaks a task into steps, calls tools for them, checks the intermediate result and decides how to continue. Those tools are perfectly ordinary interfaces: the ERP, the CRM, the calendar, the ticket system or a database.

This is made possible by language models that handle such interfaces reliably, and by open standards such as the Model Context Protocol, which connect existing systems in a uniform way. An agent is therefore less a new product than a new way of operating the software you already have.

Where agents deliver value today

Agents deliver most where many small, similar cases occur that someone currently carries back and forth between several systems by hand: pre-qualifying enquiries, preparing quotes, recording orders, classifying and answering support cases, reconciling data, compiling reports.

Agents have also arrived in software development itself. They write code, run tests and fix errors over several passes. That speeds up the work considerably, but it does not replace responsibility for architecture, security and operations.

What matters when introducing them

An agent is only as good as the environment it works in. Three things are decisive. First, clean interfaces: what the agent cannot reach through a clearly defined function, it will not do reliably. Second, clear permissions: an agent may only do what an employee in that role would be allowed to do, and every action must be logged traceably. Third, a human in the right place: for payments, contracts or customer data, an approval step belongs in the process.

Then there are data protection and cost. Which data may leave the company determines the choice of model and of where it runs. And because agents work in several steps, they use more compute than a single request. Both can be planned if they are considered from the start.

Start small, measure properly

It has proven effective to start with one narrowly defined process that measurably costs time today. Such a pilot shows within a few weeks how reliably the agent works, where it needs help and what it saves. Only then is it worth extending it to further processes.

Support from inaffect

Agents that hold up in operation.

inaffect AG supports you in this area: from choosing the right process and integrating LLMs into your existing systems to agentic workflows that comply with Swiss FADP and GDPR and are looked after in operation. Built with AI. Owned by engineers.

Our AI services

We look forward to your inquiry.

One field is enough – we get back to you within 24 hours.