An AI agent is software that uses a language model to pursue a goal by deciding its own next steps and taking actions in other systems: reading records, drafting messages, updating cases or initiating payments. A chatbot replies to you. An agent works on your behalf. That ability to act is what makes it useful, and what makes it a control question.
What are the parts of an AI agent?
Four, in most designs. A model that reads the situation and proposes the next step. Tools that let it act outside the conversation, such as a search over the case file or a call to the billing system. Context: the instructions, the sources it may rely on, and what it remembers of the task so far. And a loop that runs the cycle again until the goal is met or the agent stops.
How the model picks each step is covered in how an AI agent decides what to do next.
Why does it matter that agents act?
Because the consequences change. A chatbot's mistake is a wrong sentence a person can disregard. An agent's mistake may be a sent email, a closed account or a released payment. The model at the centre is the same kind of system that sometimes produces confident, wrong statements; see what hallucination is. Once its output drives actions, those errors leave the chat and enter your records.
So the useful question about an agent is not how clever it is. It is what it is allowed to do, on what evidence, and who answers for it.
How does an AI agent work in practice?
An agent gets a task, such as "work out why this customer was charged twice and fix it". It retrieves the account history, often through retrieval-augmented generation, and reads the relevant policy. It proposes a step, perhaps "refund the duplicate charge". In a governed setup, that proposed action goes to a control layer, which checks it against the agent's role, its certificate for refunds, the facts it rests on, the customer's limits and platform guardrails. The action is allowed, held for a person, or refused, and the verdict is filed.
For instance, a utility's billing agent spots a duplicate direct debit of 80. It is certified to execute refunds under the utility's limit, the duplicate shows in the signed payment record, and the refund goes through. The same agent finding a disputed charge of 2,000 drafts the case for a person instead.
How is an AI agent different from a chatbot?
A chatbot holds a conversation; its output is text for a person to read. An AI agent may hold a conversation too, but it also takes actions through tools, and it can work on a task with no conversation at all. Many products called chatbots are adding tools, which turns them into agents whether or not anyone notices the change. The test is simple: can it change a record or send something without a person pressing a button?
Where Pryme fits
Pryme governs AI agents: it decides what each one may do per capability, proves it on the customer's own cases, enforces it at the moment of action and records every verdict. Agents start at read-only and earn more depth one capability at a time. See the governed agentic operating system, explained.
Questions people ask
Is an AI agent the same as a bot or an automation script?
No. A script follows fixed steps written in advance. An agent decides its steps as it goes, which lets it handle messy, unstructured work but also means its behaviour has to be bounded and tested.
Can an AI agent work without a person?
It can, within limits. In regulated work, most agents draft or prepare actions for a person to approve until they have proven they handle a specific capability well.