What is an AI agent? How it works, with examples

What is an AI agent?

An AI agent is software that's given a goal by a human and then largely chooses its own route: gathering information, using other systems, assessing intermediate results and deciding the next step. Unlike a chatbot, which answers a question within a conversation, an AI agent carries out a task in several steps and can consult systems and sometimes change them. How much independence it gets is a setting the organisation chooses, from only preparing to working fully independently with checks after the fact.

An AI agent is a software system that’s given a goal and works out for itself which steps are needed to reach it. It gathers information, uses other systems, assesses intermediate results, and chooses the next step based on those. The goal comes from a human; the agent largely decides the route itself.

That’s the essential difference from most AI applications you already know. A tool that summarises a text or answers a question delivers a single outcome for a single request. An agent works on a task made up of several steps, such as “complete this customer file” or “prepare the quote for this request”.

Most agents you come across today run on a large language model. That model handles understanding the instruction and reasoning about the next step. Around it sit the parts that turn a language model into an agent: access to tools and systems, a memory for the task at hand, and instructions on what is and isn’t allowed.

Anyone who mainly wants to know what this means for an organisation will find the applications, costs and approach on the page about AI agents for business.

Can a mistake by an AI agent carry through without the final result looking wrong?

Yes, that can happen: an AI agent takes several steps in sequence, so a mistake from an early step can carry through to the final result without it looking wrong. Logging and checkpoints matter more for an AI agent than for an ordinary AI application because of this.

What forms of structure occur in an AI agent?

Three forms exist, broadly: a single-purpose AI agent for one tightly scoped task, an AI agent within an AI workflow with a fixed process flow, and collaborating agents where a coordinating agent divides the work. For most organisations, the AI agent within a workflow is the sensible starting point.

Source: article by Dave Willemstein. Updated: October 2026.

How does an AI agent work?

An AI agent’s operation comes down to a loop that repeats until the goal is reached or the agent gets stuck.

1. GoalA human or system gives an instruction
2. PlanThe agent decides the next step
3. ActCall a tool or system
4. AssessIs the result right, is the goal reached?
5. Finish or continueDeliver, or back to planning

What happens at each step:

• Goal and instructions. The agent gets an instruction plus fixed rules: which sources it may use, which actions are forbidden, and when it must bring in a human.
• Planning. The language model works out what the logical next step is. For a simple task that’s one action; for a complex one, a sequence.
• Acting through tools. A tool is anything the agent can call: a search function, a database, a CRM, a calendar, a calculation module, or another AI model. This is where the agent moves from talking to doing.
• Assessing. The agent looks at the result of its action. Did the search return anything useful? Does the answer fit the question? If not, it tries a different route.
• Memory. During the task, the agent remembers what it’s already done. Some agents also keep information between tasks, which is useful but also raises questions about what data you’re retaining. Those questions are covered separately on this site, in an article on AI agents and data protection.

Because an agent takes several steps in sequence, a mistake in an early step can carry through to the final result without that result looking wrong. That’s why logging and checkpoints matter more for agents than for an ordinary AI application. I covered this in the article on what AI agents can and cannot do: agents do well on reversible steps with a clear final check, and break down on long chains where a mistake carries through quietly.

AI agent versus chatbot

The difference between an AI agent and a chatbot is one of the most common questions, because they can look alike from the outside. You can ask both something in plain language. The difference is in what happens next.

	Chatbot	AI agent
What it does	Answers a question within a conversation	Carries out a task in several steps
Access to systems	Usually limited to a knowledge base	Can consult systems, and sometimes change them
Who decides the route	The conversation, question by question	The agent, based on the goal
Consequence of a mistake	A wrong answer	A wrong action, possibly with knock-on effects
What you need to arrange	Accuracy of answers	Access rights, checkpoints, logging and ownership

A chatbot that tells a customer how to change their address is a chatbot. A system that actually changes the address, sends the confirmation, and updates the file is an agent. In practice, the line isn’t always sharp: many chatbots are gradually gaining agent features, and many agents have a chat window as their front end.

Types of AI agents

Several ways of classifying agents exist. For an organisation, two are the most useful: by degree of independence, and by structure.

By independence. How much an agent is allowed to do on its own is a setting you choose. An agent can only prepare (gathering information for a human), propose actions that a human approves, act with confirmation at irreversible moments, or work fully independently with checks after the fact. My advice is almost always to start low and only increase autonomy once you know what mistakes occur.

By structure. Broadly, three forms exist:

• Single-purpose agents that carry out one tightly scoped task, such as preparing a reply to a customer email.
• Agents within an AI workflow, where a fixed process flow sets the main line and the agent only handles the steps that need interpretation, such as reading a document or classifying a request.
• Collaborating agents, where several agents each handle part of a task and a coordinating agent divides the work. That’s more powerful, but also harder to test and control.

For most organisations, the second form is the sensible starting point: fixed rules where you can, an agent where you must.

Examples

A few generic examples to make it concrete:

• Customer service: an agent reads an incoming email, looks up the order, drafts a reply, and puts it to a staff member.
• Sales: after a client call, an agent turns the notes into a summary, updates the CRM, and drafts a follow-up email for approval.
• Finance: an agent compares incoming invoices against purchase orders and flags discrepancies for accounts payable.
• Internal knowledge: an agent answers colleagues’ questions based on handbooks and policy documents, with the source cited.

A fuller list, worked out by department with how it works and the integrations needed, is on the twelve AI agent examples page.

Considering an agent yourself? The next question is how to have one built: what to set out upfront, how to choose a delivery partner, and what it costs. That’s covered on the page about having an AI agent built. Want to talk through whether an agent is the right fix for your process first? Get in touch for a no-obligation conversation.