GlossaryAI agent
AI agent
An AI agent is software that receives a goal, chooses the steps to reach it by itself and acts directly in the company’s tools.
Where an assistant drafts an answer that someone then has to copy elsewhere, an agent starts from a goal and acts. Asked to chase September’s invoices, it opens the accounting software, finds the ones still unpaid, checks that no payment arrived the day before, rereads the last exchange with each customer and prepares reminders that take it into account. Three elements define it: a goal rather than a script, tools that let it read and write in the company’s software, and a loop in which it acts, looks at the result and then decides on the next step. Agentic AI is the name for these systems as a whole.
How to recognise a real agent
The word no longer guarantees anything, and many products use it for a renamed assistant or automation. An agent is better recognised by five behaviours: it writes into a system, chooses the order of its steps, notices its failures, starts without being asked and remembers from one time to the next. A product that shows none of them is an assistant, one that shows two or three is an automation, and only two of them can be seen during a demo. Our detailed definition of an AI agent turns this into a buying method for a small business, and our article on agent washing gives the questions to ask the vendor about the other three.
Agent, automation and AI employee
An automation runs a script written in advance, whereas an agent decides while it runs. The deciding criterion is not how complex the task is but whether you can write out all its branches in advance, as our comparison of agents and automation shows. An AI employee is an agent that holds a role: every AI employee is an agent, but an agent built for a single task remains a tool that has to be configured and then maintained. The MCP protocol is one way of giving an agent tools, and it says nothing about what the agent is allowed to do with them.
What European law says
The European AI Act does not use the word “agent”. An agent falls under its definition of an AI system: a machine-based system designed to operate with varying levels of autonomy, which infers from the input it receives how to generate predictions, content, recommendations or decisions that can influence its environment (article 3). Since 2 February 2025, a company that uses one must take measures to ensure a sufficient level of AI literacy among its staff (article 4), and since 2 August 2026, a system intended to interact directly with people must be designed so that they know they are dealing with an AI (article 50). The GDPR adds that a person has the right not to be subject to a decision based solely on automated processing when it produces legal effects concerning them or similarly significantly affects them (article 22).
In practice, an agent can act alone on what stays inside the company, such as reading a file, filing a document or updating a record, and must go through a person for what leaves it, such as a message to a customer, a publication or a payment. That limit holds when it is built into the product rather than written in an instruction, as our list of what an agent should not be able to do explains.
Balt is an agent of this kind, deployed as an AI employee in Teams, Slack, WhatsApp and his web app for small and mid-sized businesses.

