AI agents in your organisation: what they are and how you stay in control
Short answer
An AI agent is a language model that does not just answer, but takes actions itself: querying a system, creating a document, sending an email. That enables real automation, but also real damage if the model makes a mistake or is manipulated. Stay in control with minimal rights per agent, human confirmation of irreversible actions, a separate identity for each agent and a complete audit log.
From chatbot to agent
A chatbot returns text; a person decides what happens with it. An AI agent is given tools: access to systems and the ability to take actions. It can split a task into steps, look up information, create something and pass the result on.
Examples from organisations:
- reading incoming invoices, checking them against the order and preparing them for approval;
- answering a customer question with data from the CRM and preparing a draft reply;
- working through a case file and drafting a first version of a report;
- classifying a service desk ticket, completing it and assigning it.
What can go wrong
| Risk | Example |
|---|---|
| Too many rights | An agent meant to read invoices can also prepare payments or export customer data |
| Manipulation | An incoming email contains hidden instructions that the agent follows. See Prompt injection |
| Mistakes at scale | A wrong interpretation is not made once, but a hundred times a day |
| Unclear responsibility | Nobody knows who approved an action, or why the agent carried it out |
| Data leaving the building | An agent at an external AI service sends context to that service at every step |
OWASP lists “excessive agency” among the main risks of LLM applications, precisely because agents can do more and more.
Five limits for every agent
- Minimal rights. Each agent gets only the systems, actions and data its task needs. Read access where reading is enough.
- Its own identity. Do not let an agent work under an employee’s account, but under its own service account, so you manage its rights separately and can trace its actions. See Access control for AI.
- A human in the loop for irreversible actions. Paying, sending outside the organisation, deleting, finalising something: an employee confirms.
- Complete logging. Which agent did what, with which data, based on which instruction, and who approved it.
- An off switch. An agent must be stoppable immediately, per role or for the whole organisation.
Start small, then expand
Start with agents that prepare rather than execute: a draft, a proposal, a queued task that a person completes. Measure how often the employee changes the proposal. Only when that rarely happens is it time to hand over one more step. See also Where to start with AI automation.
With werqly, agents and automations run on a dedicated node in your server room, connected to your systems through connections you manage yourself. In werqly AI Control you decide per role who may start which automations, and the audit log records every step. At no step does context go to an external AI service. See What is private AI?
Frequently asked questions
Questions on this topic
What is the difference between a chatbot and an AI agent?
A chatbot gives an answer and a person decides what happens with it. An AI agent can take actions itself: querying systems, creating documents, sending messages. That lets it take over more work, but also do more damage when it makes a mistake.
May an AI agent make decisions on its own?
Technically it can, but it is wise to have a person confirm irreversible or important actions. Decisions about people are also subject to the GDPR and, for high-risk applications, to the human oversight requirements of the AI Act.
How do I stop an AI agent from being able to do too much?
Give each agent its own identity with only the rights its task needs, have sensitive actions confirmed, log everything and make sure there is a way to stop the agent immediately.
Who is responsible for what an AI agent does?
The organisation that deploys the agent. Record for each agent who owns it, which actions it may take and who approves them. A complete audit log shows afterwards what happened and why.
Sources
- OWASP — Top 10 for LLM Applications 2025 (LLM06 Excessive Agency)
- AI Act — Regulation (EU) 2024/1689, Article 14 (human oversight)
This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.