On this page
Where a chatbot answers a question and stops, an AI agent keeps going: it plans the steps, uses tools to carry them out, looks at what happened and decides what to do next, until the goal is reached or it needs a person. The engine is a large language model. What turns it into an agent are three things around it: tools it may call, memory of what it has done, and a loop that keeps it going.
The word agent has been part of AI research for decades. Language models made it practical. In 2022 the ReAct paper showed a model alternating between reasoning and acting with tools; in 2023 experiments such as AutoGPT chained tasks on their own, impressively and unreliably. In November 2024 Anthropic introduced the Model Context Protocol, an open standard for connecting assistants to tools and data that most AI vendors have since adopted; see MCP. Since then agents have moved from demos into real work, mostly in tightly scoped tasks.
Chatbot, workflow or agent
| Chatbot | Automated workflow | AI agent | |
|---|---|---|---|
| What it does | Answers questions | Runs fixed steps you defined | Decides its own steps towards a goal |
| Example | Answers FAQs from your site | New invoice in, booked in accounts | Handles a supplier dispute: reads the thread, checks the order, drafts a reply |
| Predictability | High | Very high | Lower: it chooses its path |
| Best for | Information | Repetitive, well-defined work | Varied work with judgement in between |
Anthropic's own advice to builders is to start with the simplest option that works. Many tasks sold as agents are better done as a workflow with one AI step: cheaper, faster and easier to check. If you can write the steps down in advance, you need a workflow, not an agent.
One supplier complaint, worked through in a loop
Goal: answer supplier X's complaint about delivery 4471
Plan
read the email, find the order, check the delivery
Act
searches the ERP, opens the delivery note
Observe
two boxes short, recorded on the note
Repeat
drafts a reply with the note attached
Example of a scoped task. The agent may read and draft; a person sends.
Five business tasks agents handle today
- Inbox triage. The agent reads incoming email, recognises quote requests, complaints and invoices, drafts replies from your own information and files the rest. A person approves what goes out.
- Lead follow-up. A form comes in; the agent looks up the company, checks your CRM for history, drafts a first reply and proposes a call slot.
- Back-office checks. Matching delivery notes against invoices, chasing missing documents, preparing a monthly report from several systems.
- Operations. On a transport platform like the one we built for DA Logistics, an agent could prepare the decisions waiting each morning, from leave requests to expiring inspections, for a planner to confirm.
- Research. A briefing on a prospect or market from public sources, every claim linked to where it came from.
What a run costs
The model is the cheap part. The real cost is connecting the agent to your systems through APIs or MCP, setting its limits, and reviewing its work in the first months. Estimate your own usage with the AI cost calculator; a longer walk-through is in our guide AI agents explained.
The limits that keep it safe
An agent that may send email can send the wrong one; one that may change records can change the wrong one; and a model can hallucinate halfway through a task. Good agents get the narrowest permissions possible, a person approving anything external or irreversible, and a log of every step. In the EU, the AI Act has required AI literacy among staff using these tools since February 2025, and some uses carry extra duties; check yours with the EU AI Act checker.
Starting with an agent in four steps
- Pick one recurring taskVaried enough to need judgement, contained enough to check: supplier disputes, lead follow-up, document chasing.
- Give read access firstEvery write action, such as sending or changing a record, is a separate decision.
- Put a person at the gateThe agent prepares; a person approves anything external or irreversible. Loosen only with evidence.
- Log and reviewEvery step written down. Review a sample weekly for the first two months.
What teams ask about agents
What is agentic AI?
AI that acts with some autonomy, choosing its own steps and tools towards a goal. It is the property; an AI agent is the system that has it.
Can an agent replace an employee?
Not as an unsupervised colleague. It can take over well-scoped parts of a job, preparing work that a person checks and sends.
Which tools can an agent use?
Anything you connect and allow: email, calendar, CRM, accounting, a database, a browser. Each connection is a permission you grant deliberately.