What is an AI agent? A plain guide for business operators
An AI agent is software that decides its own next step to finish a job. Here is what that means, where the term gets stretched, and how to judge one for your business.
An AI agent is software that is given a goal and works out the steps to reach it. It reads what is in front of it, decides what to do next, uses tools such as your email, your accounts system or a website, checks the result and carries on until the job is done or it gets stuck. The important word is decides. A normal program follows steps a person wrote in advance. An agent chooses some of its own steps.
OpenAI puts it simply in its practical guide to building agents: agents are systems that independently accomplish tasks on your behalf. The same guide says that a simple chatbot, or a tool that only answers one question at a time, is not an agent. That distinction matters, because a great deal of software now carries the agent label.
Anthropic, which makes the Claude models, draws a useful line in its guide to building effective agents. It separates workflows from agents. In a workflow, an AI model is used inside steps that a person has laid out in code: read this invoice, pull out the total, put it in this field. In an agent, the model directs its own process and picks which tools to use. Anthropic's advice is to start with the simplest approach that works and only add the freedom of an agent when the task needs it. Most useful business systems today are a mix of the two.
Under the bonnet, an agent usually has four parts. There is a large language model, the engine that reads and writes text and makes the decisions. There are tools, which are the connections that let it act, such as sending an email or updating a record. There are instructions, which tell it what the job is and what it must never do. And there is memory or context, the information it can see about this customer, this order or this conversation. When people compare agents, they are mostly comparing how good each of these parts is and how tightly they are controlled.
A concrete example helps. A customer emails to say an order has not arrived. A chatbot might reply with a link to the tracking page. An agent could look up the order, check the courier's status, see that the parcel is lost, follow your rules on replacements, raise the replacement order and write back to the customer with the new delivery date. A person might still approve the replacement. The difference is that the agent carried the job through several systems instead of answering one question.
The term is stretched, and buyers should know it. In June 2025 the research firm Gartner warned of what it called agent washing: rebranding existing chatbots, assistants and robotic process automation tools as agents without real agentic capability. Gartner estimated that only about 130 of the thousands of vendors claiming agentic AI were real. The same release predicted that over 40 per cent of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear business value and weak risk controls.
Adoption is real but uneven. In its 2026 State of AI survey, McKinsey reported that the share of large organisations scaling AI agents rose from 27 per cent to 40 per cent, while smaller organisations stayed flat at 22 per cent. About two in ten respondents said they had reached the scaling stage with agents across their organisation. In other words, many firms are trying agents and fewer are running them widely.
Agents also make mistakes, and their mistakes are different from a person's. They can misread an instruction, act on false information or be tricked by text hidden in a document or web page. Because an agent can take actions, a mistake can travel further than a wrong answer from a chatbot. This is why serious deployments limit what an agent can touch, keep a log of what it did and put a person in charge of decisions that carry real cost or risk.
The practical question for an operator is not whether a product calls itself an agent. It is whether the software can take a defined job off your team's desk, at a cost you can measure, with errors you can catch. Some jobs fit that test today. Many do not, and for those a simpler tool or a capable person is still the better answer.
Why it matters to your business
Agents change the unit you buy. Instead of paying for software that helps a person do a task, you may pay for software that does part of the task itself. That affects hiring plans, budgets and who is accountable when something goes wrong. Knowing the difference between a workflow, a chatbot and a true agent stops you paying agent prices for a relabelled tool, and stops you giving a tool more freedom than the job needs.
What to do on Monday
- List three repetitive jobs in your business that cross more than one system, such as email plus your accounts package. These are the jobs where an agent, rather than a chatbot, could help.
- For any product pitched to you as an agent, ask the vendor to show it completing one of those jobs end to end, including what happens when it hits something it does not understand.
- Ask what the agent is allowed to change on its own, what it must ask a person about, and where you can see a record of every action it took.
- Start with the simplest version that does the job. If a fixed workflow with an AI step works, you may not need a fully independent agent yet.
More on this topic: AI agent news