What is an AI agent?
An AI agent is software that takes a goal, chooses the tools it needs and acts until the goal is met. A chatbot stops at an answer. An agent carries the task through, and asks a person when a step needs approval.
The difference is who chooses the steps.
The same language model can sit at any of these levels. What changes is how much it is allowed to decide and do, and that is a design choice.
- Answers
Chatbot
Text in, text out. It can explain your refund policy. It cannot issue the refund.
- Follows a script
Chatbot with tools
It can call a tool, but only along a path a person wrote in advance. Useful where the steps never change.
- Chooses its steps
AI agent
It is given a goal, decides which tools to use and in what order, checks the result and carries on until the goal is met or it needs a person.
- Works as a team
Multi-agent system
Several agents with separate jobs hand work to each other: one reads the request, one checks the rules, one writes the reply.
What an agent is built from.
The model gets the attention. In practice most of the engineering goes into what surrounds it: the context it reads, the tools it can reach and the limits it works within.
A model
The language model does the reasoning: it reads the request, plans the steps and writes the output.
Context
What the agent reads before it answers: your policies, records and code. Poor context is the most common cause of a wrong answer.
Tools
The actions it may take, such as looking up a record, booking a slot or sending a message, usually exposed over the Model Context Protocol.
A loop
Plan, act, check, repeat. The loop is what lets an agent recover from a failed step without a person stepping in.
Limits
Which actions it may take alone, which need approval and which are off the table. Set per task by the people accountable.
A record
Every step written down: what it read, what it did and who approved it. Without a record nobody can review the work.
When an agent fits, and when it does not yet.
Projects that fail these tests tend to stall after launch. The signs are easy to check before any money is spent.
Build an agent when
- The path changes from case to caseSomeone has to read, decide and route each time. That judgement is the part an agent can take on.
- The facts live in a system of recordA CRM, a student system, a ticket queue or a repository the agent can read and trust.
- The work comes in volumeEnough cases each month that the time saved outweighs the cost of testing and watching the agent.
- Someone owns the outcomeA named person approves what the agent may do and answers for it.
Wait, or use something simpler, when
- The steps never changePlain automation is cheaper, faster and easier to audit.
- The truth lives in inboxesFix where the facts are kept first. An agent reading bad data is worse than no agent.
- Nobody will approve its actionsFind the owner before the engineers, or the project stalls at launch.
- Every action must be a person’sThat is a drafting assistant, and it should be designed as one from the start.
Most failures are about context and control.
Agents rarely fail because the model is weak. They fail because they read stale or conflicting information, act without clear limits, or cannot show where an answer came from.
Over 40%of agentic AI projects will be cancelled by the end of 2027, for cost, unclear value or weak risk controls.Gartner prediction, 25 Jun 2025
AI agents, asked plainly.
For the terms themselves, see the glossary. For the standards agents rely on, see the field notes on the agentic web.
What is an AI agent in simple terms?
An AI agent is software that takes a goal, chooses the tools it needs and acts until the goal is met. A chatbot stops at an answer. An agent carries the task through, and asks a person when a step needs approval.
What is the difference between an AI agent and a chatbot?
A chatbot replies to a message. An AI agent plans and acts: it can read a record, call several tools in sequence, write the result back and continue until the goal is met.
Do I need an AI agent or just automation?
If the work follows fixed rules, use plain automation. An agent is worth building when the path to the answer changes per case and a person normally has to read, decide and route.
Is ChatGPT an AI agent?
On its own it is a chat interface. Given tools, a goal and a loop that lets it act and check its work, the same model becomes an agent. The architecture decides, the brand name does not.
What are the risks of using AI agents?
Wrong answers from stale or conflicting context, instructions hidden in the content an agent reads (prompt injection), actions taken without authority and costs that grow unnoticed. Each has a control: grounded context, limits per task, approvals and a full log.
What does an AI agent cost?
It depends on how many systems the agent must reach, how much it may do without approval and how carefully it must be tested. We map the operation first and quote for that, so the figure is tied to real work.
Should we build, buy or work with a partner?
Buy when the task is generic. Build in-house when you have the engineers and the data is yours alone. Work with a partner when you want the first operation running soon and want to own the result afterwards.
When should you not build an AI agent?
When the steps never change, when the facts are not kept in a system the agent can read, when no one will own its actions, or when rules require a person to take every action.
Have an operation in mind? Check it with us.
Tell us what the work is. We will say plainly whether an agent fits, and what it would take to run one safely.