An AI agent is software in which a language model decides its own next step and acts through tools, such as reading an inbox, looking up an order or sending an email, until the goal is met or a limit stops it. A workflow follows steps written in advance; an agent chooses the steps as it goes.
- An AI agent is software in which a language model chooses its own next step and acts through tools until a goal is met or a limit stops it.
- A workflow follows steps written in advance. An agent decides the steps as it goes, which costs more time and money per task.
- Agents fit work where each case is a little different: support replies, lead follow-up, invoice chasing, reporting.
- Use a fixed workflow instead when the rules never change or a wrong action cannot be undone.
What is an AI agent?
An AI agent is a language model given a goal, a set of tools and permission to use them. It reads the situation, decides what to do next, does it, checks what happened and repeats until the job is done or a limit stops it.
Anthropic, which makes the Claude models, draws the line this way in Building effective agents:
"Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage."
The difference that matters to a business: in a workflow, you wrote every step. In an agent, the model picks the steps, inside the limits you set.
How an AI agent works
Every agent runs the same loop. The example column follows one support email.
| Step | What happens | Example: a support email |
|---|---|---|
| 1. Goal | A task arrives with instructions | "Where is my order?" lands in the inbox |
| 2. Read | The agent gathers context | It reads the email and the support policy |
| 3. Decide | The model picks the next action | Look up the order in the shop system |
| 4. Act | It calls a tool | Queries the order and gets the tracking status |
| 5. Check | It reads the result before going on | The parcel is out for delivery today |
| 6. Stop or repeat | It finishes, loops or hands over | Sends the reply, or queues it for a person |
Anthropic notes that agents need "ground truth" from their environment at each step, such as a tool result, and stopping conditions, such as a maximum number of steps. They can also pause for a person at checkpoints. Those limits are what make an agent safe to run on real work.
The parts behind the loop:
| Part | What it is |
|---|---|
| Model | The language model that reads and decides |
| Instructions | The goal, the rules and what it must never do |
| Tools | Connections that let it act: email, CRM, calendar, database |
| Memory | What it keeps between steps or between tasks |
| Limits | Step caps, spending caps and actions that need approval |
AI agent vs workflow vs chatbot
| Question | Chatbot | Workflow | AI agent |
|---|---|---|---|
| Who decides the next step | The user, one message at a time | The person who built it, in advance | The model, as it goes |
| Acts in other systems | Usually not | Yes, on fixed steps | Yes, it chooses which tools |
| Same input, same output | No | Yes | No |
| Cost and time per task | Low | Lowest | Highest |
| Best for | Answering questions | Stable, repeated rules | Cases that differ each time |
Anthropic puts the trade-off plainly: "Agentic systems often trade latency and cost for better task performance." Pay that price only where the variety of cases needs it.
The five types of AI agents
The classic textbook types, as IBM summarises them, describe how an agent decides. Most business agents built on language models mix several.
| Type | How it decides | Business example |
|---|---|---|
| Simple reflex | Fixed rule on what it sees now | Tag every email that contains "invoice" |
| Model-based reflex | Rules plus a picture of what it cannot see | Hold a reorder while a delivery is in transit |
| Goal-based | Plans steps toward a goal | Book a meeting that fits both calendars |
| Utility-based | Weighs options and picks the best trade-off | Pick the cheapest carrier that still meets the date |
| Learning | Improves from feedback over time | Rank leads better as deals close or fail |
Several agents can also work together as a multi-agent system, for example one that sorts incoming email and others that handle each kind.
AI agent examples by business function
For each function, the point is which part of the job the agent takes and what stays with a person.
| Function | What the agent does | What a person keeps |
|---|---|---|
| Support inbox | Answers from your policies and drafts the rest | Refunds, complaints and anything new |
| Lead follow-up | Replies to new leads, qualifies them and books the call | Proposals and prices |
| Bookings and reminders | Books, reminds and refills cancelled slots | Exceptions to your calendar rules |
| Invoices and collections | Issues invoices, matches payments and sends reminders | Disputes and payment plans |
| Reporting | Pulls your numbers into one weekly report | What to change |
| Content | Drafts posts, emails and product copy | Every piece before it goes out |
| Back-office data entry | Moves data between email, forms and your CRM | Records it flags as unclear |
For more worked cases, see AI agent examples and AI agents for small business.
Building with AI models? AIGROW API credit runs Claude, GPT and Grok models up to 75% below list price.
When not to use an AI agent
Anthropic's advice is to start by "finding the simplest solution possible, and only increasing complexity when needed." Skip the agent when:
| Situation | Use instead |
|---|---|
| The rules never change | A fixed workflow in a tool such as n8n, Make or Zapier |
| A wrong action cannot be undone, such as a payment | An agent that drafts, with a person who approves |
| You need the same output every time for an audit | A fixed workflow with logged steps |
| The task happens a few times a month | A person, or a saved prompt |
| The agent cannot reach the data or the tool it needs | Fix access first |
If the job is a fixed chain of steps, compare the tools in n8n vs Zapier before you build an agent.
How to start with one agent
- Pick one function with many similar cases, such as the support inbox.
- Write down what the agent may do alone, what needs approval and what it must never do.
- Give it read access first. Add actions one at a time.
- Set limits: a maximum number of steps per task and a queue for anything your rules do not cover.
- Review a sample of its work every week and tighten the rules.
You can build this yourself with the guide to building AI agents with n8n. Or AIGROW can run it for you: our AI agents for business take over one function at a time in your own tools, with a person checking anything your rules do not cover.
Sources
- Anthropic, Building effective agents, 19 December 2024
- IBM, Types of AI agents
The newest Claude, GPT and Grok models, up to 75% below list price
One API key for Claude, GPT and Grok models. It works with the OpenAI and Anthropic SDKs and with Claude Code.
Each model’s rate through AIGROW credit sits beside its maker’s list price, never above it.
Frequently Asked Questions
