AI agent
An AI agent is a software system that uses a language model to pursue goals and carries out actions on its own to reach them, such as buying, booking or filing applications.
Unlike a chatbot, an AI agent does not just answer; it acts. It calls tools and interfaces, fills in forms and triggers payments. To do so, it needs access to the accounts, data or payment methods of its principal.
The more independently an agent acts, the more it matters whether a specific action was covered by its instructions. An agent can make mistakes, be manipulated or exceed its limits. Secure architectures therefore separate the reasoning of the model from the decision whether an action is allowed.
Example
A customer instructs her agent to renew the monthly pass for her commute. The agent finds the offer, compares the price with the permitted range and pays. Only a mandate with an amount limit and a receipt make it traceable later that exactly this purchase was allowed.
How it differs
A bot runs fixed, programmed routines; a chatbot answers questions. An AI agent plans its own steps and picks tools as the situation requires. As currently understood, it is not a legal person; its actions are generally attributed to its principal.
In practice
- Give every agent its own identity and keys instead of letting it work with your personal credentials.
- Define a mandate with limits for every task and check actions before execution, outside the model.
- Keep receipts of every action, so you can show later what the agent did on your behalf and when.
Role in KYA
KYA does not verify the agent itself but the evidence of its authority: who the agent is, on whose behalf it acts, what it was allowed to do, and whether this action matched.
Related check: Agent and principal (subject_agent_binding). Do all parts name the same agent, is the principal named, does the agent hold the bound key?
Frequently asked questions
Is an AI agent the same as a chatbot?
No. A chatbot answers requests; an AI agent carries out actions such as orders or bookings. It uses tools and interfaces to do so, often through protocols like MCP. Usefulness and risk grow at the same time.
Who is liable for the actions of an AI agent?
An AI agent is not a legal person; its actions are generally attributed to its principal. What matters is whether the action was covered by the instructions given. That is exactly why verifiable evidence is needed. General information, not legal advice.
How can a shop tell whether an agent is authorized?
Through evidence the agent presents: a signed mandate, a decision before execution and a receipt afterwards. The shop or a verifying party can recompute this evidence independently, for example with KYA.