What is an AI agent? Types, examples and how businesses use them

An AI agent is software that takes a goal, works out the steps, uses tools such as email, a database or an ERP to carry them out, and checks the result. A chatbot answers questions. An agent does work. In a business, a useful agent has one narrow job, a fixed list of actions it may take, and a person who approves anything risky.

#How an AI agent works

Whatever the branding, every agent runs the same loop.

1. Receive a goal or a trigger. A supplier email arrives, a ticket is opened, or a manager types "reconcile the March statement."
2. Read the context. The agent pulls what it needs: the email body, the customer record, the purchase order, last quarter's policy document.
3. Decide the next step. A language model picks from the actions it has been given, such as "look up vendor," "draft reply" or "flag for review."
4. Act through a tool. The agent calls an API, queries a database or writes a draft. Anything outside its list is off limits.
5. Check the result. Rules or a second check confirm the output makes sense. If not, the agent retries, asks for more information or hands the case to a person.
6. Log everything. Inputs, decisions, tool calls and outcomes are recorded so someone can reconstruct what happened.

Here's a concrete case. A distributor gets supplier statements by email as PDFs. An agent reads each one, looks up the supplier in the ERP and compares open invoices against the statement lines. It lists the differences: a missing credit note, a duplicate invoice, a payment nobody applied. Then it drafts a query email for the AP clerk, who approves or edits it. The agent never sends or posts anything on its own.

#AI agent vs chatbot vs workflow

The terms get mixed up, so here's the short version.

A chatbot holds a conversation. It may answer from a knowledge base, but it usually can't change anything in your systems. A workflow follows a fixed path you designed in advance: if the invoice total is over $5,000, send it to the controller. An agent decides which step comes next, within limits you set.

Most good business systems mix all three. Fixed rules handle what's predictable, and an agent handles the step that needs reading or judgment. We compare them in detail in AI agents vs RPA vs workflow tools, and the related idea of giving a system more autonomy is covered in what is agentic AI.

#The classic agent types

The classic taxonomy comes from academic AI textbooks and describes how an agent decides. It turns up in nearly every "types of AI agents with examples" search result. It's worth knowing because it explains why some agents are simple and others are hard to trust.

Two more textbook terms come up in course notes. A rational agent is one that picks the action expected to give the best result, given what it knows. A knowledge-based agent keeps a store of facts and reasons over it, which is close to how a RAG assistant answers from your documents. The model-based agent above is the "model based agent in AI" you will see in definitions.

Some lists stretch this to seven types by adding hierarchical agents (a manager agent directing specialist agents) and multi-agent systems (several agents cooperating). Those describe architecture more than decision style.

Real systems blend the categories. A modern business agent uses a language model to interpret input, rules to enforce limits, and reviewer feedback to improve. Think of the textbook types as vocabulary. Nobody picks one from a menu.

#Practical types of business agents

When we scope an agent for a client, nobody asks "is this utility-based?" They ask what the agent does all day. These are the job types that come up most.

If you're choosing where to start, begin with extraction or triage. Their errors are visible and cheap to catch.

#AI agent examples by business function

These describe common designs. They're patterns, not claims about any named company.

#Finance

An AP agent reads incoming invoices, matches them to purchase orders and receipts, flags price or quantity differences, and proposes a general ledger code based on how similar invoices were coded before. A rule decides whether the invoice needs approval, based on amount and vendor. Our page on finance operations shows how we set this up.

A collections agent reviews overdue receivables, checks whether a payment is sitting unapplied, and drafts a reminder in the tone you use for that customer segment. The AR lead approves each batch.

#Customer support

A support agent classifies a ticket as billing, bug or account access, pulls the customer's plan and recent orders, and drafts a reply from your help center. For refunds it applies your policy limits, and anything above the limit goes to a person with the reasoning attached. More on this in customer support automation.

#Sales and revenue

A lead research agent takes a new inbound lead, finds the company's size, industry and likely contact roles, and writes the CRM notes before a rep opens the record. A follow-up agent drafts emails after demos from the call notes, and reps review them. We don't let an agent send first-touch outreach unsupervised, because one bad batch can damage a sender domain's reputation.

#Operations and logistics

A freight audit agent compares carrier invoices to rate agreements and flags overcharges. An order-entry agent turns customer POs sent as PDFs into sales orders and routes mismatches to a coordinator. Property managers use agents to triage maintenance requests by urgency and trade before dispatch.

#HR

An onboarding agent collects documents from a new hire, checks them against a list, creates accounts through the IT ticketing system and chases missing items. It doesn't make hiring decisions.

Key point: What makes an AI agent useful is the limits around it, more than how autonomous it is. A narrow job, a short list of permitted actions, rules that check its output and human approval on consequential steps are what separate a working agent from an expensive demo.

#What AI agents are bad at

Agents are probabilistic. They can return a confident wrong answer without raising an error, and that has practical consequences.

• Exact arithmetic and compliance logic. Put these in code. A model can propose a tax code, but a rule should apply the rate.
• Open-ended permissions. An agent that can touch any record will eventually touch the wrong one.
• Rare, high-stakes decisions. If one mistake costs more than the time saved across a year of cases, keep a person in charge and use the agent for preparation.
• Processes nobody can describe. If two experienced people handle the same case differently, agree on the process first.
• Low volume. Twenty cases a month seldom justify the build and testing effort.

Sometimes the honest answer is that a process needs a plain workflow tool. That's cheaper for you, and it's the right call.

#How to start with an AI agent

1. Pick one process with high volume, messy input and a manageable cost of error.
2. Write the agent's job description in one sentence, plus the list of actions it may take.
3. Collect a few hundred real examples, including the awkward ones.
4. Measure accuracy on those examples before anything goes live.
5. Run in shadow mode, with the agent working alongside your team, and compare results.
6. Launch with approvals on. Relax them only where the data shows review rarely changes the outcome.
7. Name an owner for monitoring and updates.

Our AI agents page describes how we scope and constrain them, and how a project runs covers the stages. If you have a process in mind, book a strategy call and we'll tell you whether an agent fits or a simpler tool will do.

#Frequently asked questions

#Is ChatGPT an AI agent?

A plain chat conversation isn't an agent, because it produces text and stops. Chat products add agent features when they can browse, run code or call tools on your behalf. For business use, what matters is whether the system can take actions in your systems and under what limits.

#What are the 7 types of AI agents?

The textbook list has five: simple reflex, model-based reflex, goal-based, utility-based and learning agents. Seven-item lists add hierarchical agents and multi-agent systems. In practice, business agents are described by job, such as extraction, triage, drafting or orchestration.

#What is the difference between an AI agent and a chatbot?

A chatbot converses, and an agent acts. An agent can read a document, query your systems and write a record or draft, under rules you define. Many support tools combine both, with a chat front end and an agent behind it.

#How much does an AI agent cost?

It depends on the number of systems involved, how messy the input is, how much review you need and the volume of cases. Running costs include model usage, hosting and monitoring. Measure model cost per case on your own samples before committing to a budget.

#Are AI agents safe to connect to our systems?

They can be, when access is narrow. Give the agent only the actions it needs, use read-only access where possible, require approval for payments, deletions and external messages, and log every action. We cover this in more depth in our post on AI guardrails.

#What is an intelligent agent?

An intelligent agent is the textbook name for anything that perceives its environment, decides and acts toward a goal. A thermostat qualifies in the simplest sense. In business, the phrase usually means a software agent that reads inputs, applies rules or a language model, and acts in your systems, such as an invoice agent that reads, matches and posts a bill. The practical question is how much it may do alone and where a person approves.