What is an AI agent? Definition and examples

An AI agent is a program in which a language model completes a task on its own: it plans the next step, calls a tool for it, such as a search, a database, or an ERP’s interface, reads the result, and decides again, until the task is done or a person has to take over. A chatbot writes an answer. An agent does the work.

What an AI agent consists of

Part	What it does	Example
Model	Understands the task, plans, decides	Claude, GPT, or Gemini
Tools	Reads and acts in other systems	ERP API, PIM search, sending email, MCP servers
Context	What the agent needs to know for the task	Instructions, customer history, item master
Loop	Plan, act, check, repeat	Until the order is created
Limits	What it may do, when a person decides	Read yes, post only after approval

The software that holds these parts together is called a harness. It matters more for an agent’s quality than the model does, see What is an agent harness?.

AI agent, chatbot, and workflow: the difference

	Chatbot	Workflow	AI agent
Result	An answer	A fixed process, done	A task, done
Who sets the steps	Nobody, there are none	The developer, upfront	The model, at runtime
Uses other systems	Usually not	Yes, hard-wired	Yes, through tools
Handles the unexpected	In conversation, yes	No	Yes, within the set limits

Anthropic’s guide Building effective agents draws the same line: workflows follow predefined code paths, agents direct their own process. Both are agentic systems, and often the workflow is the better choice. When to use which is in AI agent, RPA, or workflow.

An example, step by step

An order arrives by email, as a PDF attachment:

1. The agent reads the email and the PDF and recognizes customer, items, and quantities.
2. It looks the items up in the PIM. One number no longer exists.
3. It finds the successor item in the PIM.
4. It checks price and availability in the ERP.
5. It creates the order as a draft and flags the changed line.
6. A person reviews the flag and approves.

None of these steps is hard-coded. The agent decided to look for the successor because the number was missing. Only the limit is hard-coded: nothing posts without approval. Twelve jobs of this kind are in 12 jobs for AI agents in ERP, PIM, and CRM.

Where an AI agent fails in a business

• Missing context: the agent does not know a customer’s special terms, because they live only in one employee’s head.
• Bad data: it checks against an item master with gaps and finds false errors. See 9 product data errors AI agents catch.
• Too much at once: an agent is asked to run a whole process when only two steps in it need judgment.
• Limits in the prompt instead of in code: “please don’t post anything” is a request. A missing write permission is a limit.

How an agent reaches your systems

Through the systems’ interfaces, meaning APIs, or through the Model Context Protocol (MCP), the open standard agents use to call other systems’ tools. Several ERP vendors now ship their own MCP servers. Both routes are compared in MCP or API.

FAQ

Is ChatGPT an AI agent? In chat it answers questions. As soon as it opens websites, fills in forms, or edits files in agent mode, it works as an agent. The difference is not the model but whether it uses tools and decides its own steps.

What is agentic AI? The umbrella term for systems in which a language model plans steps and uses tools: fixed workflows with model calls as well as autonomous agents.

Is an AI agent safe? As safe as its limits. Its own identity with the smallest role it needs, approvals for everything that touches money or customers, and a log of every step. That belongs in code, not in the prompt.

Do we need our own AI models? No. Business agents use the large models through their APIs. What is company-specific are the context, the tools, and the limits.

How an agent sorts a real email and when it hands over to a person shows in the free AI Inbox Triage. AI agents for business that work inside your ERP, PIM, and CRM are what we build, at a fixed price.