Short answer: an AI agent is a system built on a language model that is given a goal, decides for itself which steps to take and which tools to use (search, a database, email, the ERP), carries them out, and picks the next step based on the result — until the task is done or it asks a person for help. The difference from ordinary automation is who decides the order of steps: with automation you do, with an agent the model does.
In 2026 the word "agent" is used for almost everything — from a website chatbot to a script that sends a report once a day. So it is worth separating three things that are often confused.
01. Chatbot, automation and agent: the difference
- A chatbot answers questions in a conversation. It waits to be addressed, replies, and that is the end of its job. It can be very useful, but it starts nothing on its own.
- Automation runs a predefined sequence of steps: when an enquiry arrives, extract the data, write it to the CRM, send a confirmation. One of the steps can be AI — reading an unstructured email, say — but the order is fixed and known.
- An AI agent is given a goal, not a sequence. "Prepare a quote for this enquiry" might mean the agent searches the catalogue, checks stock, looks at the customer's history, drafts the quote and asks for approval — and in another case skips half of that because it is not needed.
A fuller comparison of chatbots and process automation is in AI chatbot or process automation.
02. How an agent works
Every agent has the same basic parts:
- A model that understands the task and decides the next step.
- Tools — functions it is allowed to call: document search, database reads, sending email, writing to a system.
- Context — company data, rules, and the previous steps in the same task.
- Boundaries — what it may do on its own, and what must go through a person.
The fourth part decides whether an agent is usable in a business. An agent allowed to email a customer by itself and an agent that prepares a draft a person approves are two very different risks, even on the same model.
03. When an agent makes sense
An agent is the right choice when two conditions hold:
- The path to the goal is not known in advance. Each case needs a slightly different sequence of steps, so a fixed automation would need dozens of branches.
- A mistake is cheap or easy to catch. The agent drafts, researches, summarises — and a person checks before anything leaves the building.
Examples: preparing a reply to a complex enquiry that needs data from three systems, researching suppliers against set criteria, summarising contracts and flagging departures from standard terms, searching internal documentation with cited sources.
04. When you don't need an agent
Most business processes we measure run in the same order every time: a document arrives, data is extracted, entered, checked, sent. For those, an agent is needlessly expensive and harder to verify, because its behaviour varies from case to case.
Ordinary automation with one AI step — the one where a model reads or writes text — is cheaper, faster and predictable there. It is easier to test, easier to explain to the team, and easier to fix when something goes wrong.
That is why, for us, the choice between an agent and automation is not the starting point but the outcome of measurement. In the Design phase we decide, for each opportunity, the simplest solution that works — and an agent is only one of the possible answers.
05. What you need before an agent
- Data the agent can reach. An agent without ERP access cannot check stock, however good the model.
- Clear rules. What the standard discount is, who approves exceptions, which customers have special terms. If it is not written down, the agent will guess.
- Evaluation. A set of real examples used to check the agent works, before and after every change.
- Cost control. An agent makes more model calls than ordinary automation, so set a spending budget up front.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a conversation and waits to be addressed. An AI agent is given a goal, chooses its own steps and tools — search, a database, email, business systems — and carries them out until the task is done. A chatbot can be part of an agent, but an agent also works without a conversation.
What is the difference between an AI agent and automation?
With automation, you set the order of steps and it is always the same. With an agent, the model sets the order depending on the case. Automation is better for processes that always run in the same order; an agent for tasks where the path to the goal is not known in advance.
Are AI agents safe to use in a business?
Yes, if their permissions are limited. The safest pattern is that the agent prepares and a person approves anything that leaves the company or changes data in a system. If data must not leave the company, the agent can run on your own infrastructure — see AI security and compliance.
How much does it cost to build an AI agent?
It depends on how many systems the agent works with, how many tasks it handles, and where it runs after handover. Beyond the build there is monthly model spend, which is higher for agents than for ordinary automation. How the price is formed is set out on AI implementation cost.