What is an AI agent? A plain definition for companies
An AI agent is a language model given instructions and tools, which decides for itself which steps to take to reach a goal and checks the result of each one. A chatbot answers a question; an agent can also search your documents, call other software and produce a file. It pays off on open-ended tasks where the steps cannot be written in advance, and it costs more time and money than a fixed workflow.
Ask three vendors what an AI agent is and you get three answers. Here is the plain version, built on what Anthropic and OpenAI say in their own documentation.
What is an AI agent?
An AI agent is a language model that has been given instructions and tools, and that decides on its own how to use them to reach a goal. OpenAI’s Agents SDK documentation describes agents as “LLMs equipped with instructions and tools” and adds that they run in a loop that continues until the task is complete.
Anthropic draws a line between two kinds of systems. In a workflow, the model and the tools are orchestrated through code paths written in advance. In an agent, the model directs its own process and its use of tools. An agent starts from a command or a discussion with a person, plans, then works independently, and can pause to ask for feedback when it hits a blocker.
What does an agent do that a chatbot does not?
A chatbot takes a message and returns a message. An agent takes a goal and acts: it searches a base of documents, reads a file, queries a database, writes a document, looks at what came back and decides on the next step. According to Anthropic, at each step the agent should read the real result from its environment, such as the output of a tool, before going further.
The chat window does not tell you which one you are using. Anthropic’s own example, customer support, “combines familiar chatbot interfaces with enhanced capabilities through tool integration”: the screen looks like a chat, and behind it the agent can pull customer data and trigger an action such as a refund.
What tasks can an agent handle in a company?
Anthropic names two where agents work well: customer support and coding. In coding, solutions can be checked by automated tests, so the agent iterates on the test results.
In a company, the same pattern applies to tasks where the work is reading, cross-checking and writing. A few illustrations:
- Finding in past project files the three assignments closest to a new tender, then drafting the Word document that cites them.
- Reading a request that arrived by mail, looking up the price list and preparing a quote for someone to approve.
- Going through a repair history to prepare the reply to a customer who is contesting an invoice.
These illustrate a kind of task. None of them comes from a particular deployment.
What are the limits of an AI agent?
Anthropic is direct about them. Agentic systems “often trade latency and cost for better task performance”. Their autonomy brings higher costs and “the potential for compounding errors”: a wrong choice at step two carries into steps three to ten. Anthropic recommends extensive testing in sandboxed environments, with guardrails, and notes that it is common to add stopping conditions such as a maximum number of iterations.
One more limit follows from how agents work. An agent only knows what its tools return. With access to well-organised documents it can be useful; with a shared drive nobody has tidied for ten years, its answers reflect that. A person still has to read what it produces before it leaves the company.
When is an agent worth it, and when is it not?
Anthropic’s advice is to find the simplest solution possible and add complexity only when needed. Often that means no agent: a single model call with retrieval and a few examples is enough. Workflows suit tasks that are well defined and predictable. Agents suit open-ended problems where you cannot know in advance how many steps will be needed.
A practical test: can you write the steps on a sheet of paper before the task starts? If yes, a workflow or an automation tool will be cheaper and more predictable. If the steps depend on what the agent finds along the way, such as which documents exist, which case this one resembles, what the customer wrote, an agent earns its cost.
And with us?
We publish this blog, so you should know where we stand. We are a forward deployed agency: three engineers from Télécom Paris, specialised in AI, who build custom AI agents at the client’s site, and who also take on machine learning and data missions. Our clients work in automotive (a network of garages), environmental engineering and consulting.
Our agents read your documents, cite their sources, and write Word and PowerPoint files. They are also reachable from Claude, ChatGPT or Copilot. Sources are read in read-only mode from SharePoint, OneDrive or Dropbox, in the folders the client chooses, and no file is kept; other tools can be connected on request. Nothing goes out without validation by a person at the company.
The first call lasts 30 minutes, is free and commits you to nothing. Access to your tools is granted in two clicks, with no password, and two weeks after the call a first agent is running. For the detail of what an agent can do for a company, see the page on AI agents for business.
Sources, consulted on 12 November 2026
Questions
What is the difference between an AI agent and a chatbot?
A chatbot replies to what you type. An agent has tools, such as a search in your documents, a database query or a file writer, and chains several calls until the task is done. Many agents are reached through a chat window, so the two often look alike on screen.
Can an AI agent act without human approval?
Technically yes, but it is rarely a good idea. Anthropic describes agents that pause for human feedback at checkpoints, and recommends stopping conditions and guardrails. With us, nothing is sent without a person at the company validating it.
Does every company task need an agent?
No. Anthropic advises looking for the simplest solution first, which is often a single model call with retrieval and no agent at all. An agent is for tasks whose number of steps cannot be predicted.
Where does an AI agent get its information?
From the tools it is given. If it can read your documents, it works from them; if it has no access, it answers from what the model learned in training, which is why connecting the right sources matters.