In brief
An AI agent is a language model that is allowed to use tools itself: look up data, start calculations, read documents. That makes it far more useful than a chatbot. But an agent that may also post entries, send or delete things on its own can do a lot of damage in a few seconds. That is why we set a firm rule: AI proposes, humans decide.
How to draw the line
- Reading is allowed. Within the user's permissions, the agent may look up data in the accounts, search documents and find earlier decisions.
- Calculating and proposing is allowed. Working through scenarios, comparing options, drafting a document.
- Executing is not. Posting entries, paying, sending something to a customer or making a final change is a human action. That action simply does not exist as a tool for the AI.
- Evidence first, then conclusions. The agent first looks up what is already known and discusses it before it records anything.
One step further: fewer rights for smaller models
If the agent runs on a local model, you can restrict its rights even further, for example to read-only, until you are sure the model is reliable enough. That way you build trust step by step.
Keep a record of what the AI did
Record which tools the agent called, with which data, and what it proposed. If the AI reads data from a scanned document, also keep track of who or what read each figure. Then every outcome can be traced.
What to watch out for
- Set the rules in the system itself, not only in the instructions to the model. A model can ignore instructions; it cannot use a tool that does not exist.
- Let the AI flag doubt instead of guessing, especially with amounts.
- Start small: first reading, then proposing, and only executing with a human in between. This also keeps an AI pilot from stalling.
Would you like advice on AI agents and their limits? Request a free advice conversation.
Updated on