An AI agent is software that gets a goal in plain English, decides what to do next, does it in another system through a tool (booking a calendar slot, updating the CRM, looking up an order), and checks the result before the next step. It keeps going until the job is done or a rule tells it to hand the case to a person.
How does an AI agent work?
An AI agent works in a loop. It reads the situation, picks one action, runs it through a tool and looks at what came back. The loop ends when the goal is met, when a rule says stop, or when the agent decides a person should take over.
-
Perceive the request
-
Decide the next step
-
Act through a tool
-
Check the result
-
Finish or hand off
Here's that loop on an everyday request. A homeowner texts an HVAC company at 9 p.m.: "AC is blowing warm air, can someone come tomorrow?"
- Perceive. The agent reads the message and sorts it: service request, not an emergency (no mention of smoke, a burning smell or a vulnerable person in the home).
- Decide. Its instructions let it book diagnostic visits, but only for addresses inside the service area. Next step: look up the customer.
- Act. It calls
find_customer(phone), finds the address, then callsget_open_slots(service="diagnostic", date="tomorrow"). - Check. Two windows are open. It offers both, the customer picks the morning one, and the agent calls
create_job. The job system confirms, so the agent replies with the arrival window and the company's policy on technician arrival notices.
If any step fails (no record for that number, nothing open tomorrow, the customer mentions a gas smell), the instructions spell out what happens instead: offer the next day, take details for a callback, or page the on-call tech right away.
Researchers call this pattern reasoning and acting in turns. The best-known write-up is the 2022 ReAct paper from Princeton and Google, and nearly every agent framework today runs some version of it, including the business AI agents we design.
What are the parts of an AI agent?
An AI agent has five parts, and the language model is only one of them. When an agent misbehaves in production, the problem is usually somewhere else.
| Part | What it is | What goes wrong when it's weak |
|---|---|---|
| Model | The large language model (LLM) that reads requests and picks the next step | Misreads requests, picks the wrong tool, answers slowly |
| Tools | The actions it's allowed to take: API calls to your calendar, CRM, help desk or order system | It can only talk, or it has far more power than the job needs |
| Instructions and knowledge | Written rules plus the approved content it answers from: policies, prices, service areas | Makes up answers or contradicts your own policies |
| Memory and state | What it keeps track of within a conversation and, if you allow it, across conversations | Asks the same question twice or loses the thread mid-task |
| Guardrails and logs | Limits, confirmations, handoff rules and a record of every step | Mistakes nobody catches until a customer complains |
That's why we treat an agent as a software project, not a prompt. On our agent projects, more hours go into tools, rules, test cases and monitoring than into picking a model. The scene below shows why: a support agent searches approved help content, applies its guardrails and hands a sensitive case to a person with a summary attached.
Illustrative example Fictional business, people and data. Not a real client, conversation or result.
AI agent vs chatbot vs automation: what's the difference?
An AI agent understands open-ended requests and takes action. A chatbot mostly answers questions. A workflow automation runs fixed steps you defined ahead of time. Each one fails in its own way, so picking the right one matters more than the label.
Chatbot, automation or agent
| Criterion | Rule-based chatbot | Workflow automation | AI agent |
|---|---|---|---|
| Input it handles | Buttons and expected keywords | Structured events: a form, a new row, a webhook | Free text or speech, in the customer's own words |
| Who decides the steps | You, in a decision tree | You, in a fixed sequence | The agent, inside limits you set |
| Acts in other systems | Rarely | Yes, the same actions every time | Yes, chosen case by case |
| When something unexpected happens | "Sorry, I didn't understand" | Stops or throws an error | Asks, adjusts or hands off |
| Predictability | High | Very high | Lower; needs testing and monitoring |
| Best for | A short FAQ that rarely changes | Moving data between apps | Varied conversations that end in a task |
Rule-based chatbot
- Input it handles
- Buttons and expected keywords
- Who decides the steps
- You, in a decision tree
- Acts in other systems
- Rarely
- When something unexpected happens
- "Sorry, I didn't understand"
- Predictability
- High
- Best for
- A short FAQ that rarely changes
Workflow automation
- Input it handles
- Structured events: a form, a new row, a webhook
- Who decides the steps
- You, in a fixed sequence
- Acts in other systems
- Yes, the same actions every time
- When something unexpected happens
- Stops or throws an error
- Predictability
- Very high
- Best for
- Moving data between apps
AI agent
- Input it handles
- Free text or speech, in the customer's own words
- Who decides the steps
- The agent, inside limits you set
- Acts in other systems
- Yes, chosen case by case
- When something unexpected happens
- Asks, adjusts or hands off
- Predictability
- Lower; needs testing and monitoring
- Best for
- Varied conversations that end in a task
A rule of thumb we share with clients: if you can draw the whole process as a flowchart, automate it. If one box on that flowchart says "figure out what they mean," that box is where an agent goes. The flowchart side is covered on our workflow automation page, and the narrower bots-versus-agents question gets its own breakdown in AI agent vs chatbot.
What are some examples of AI agents in business?
The most useful AI agent examples follow a pattern: a clear trigger, a few tools, a definition of "done" and a rule for when a person takes over. These are typical of what US small and midsize businesses ask us to scope.
| Business | Trigger | Tools the agent uses | Done when | Hands off when |
|---|---|---|---|---|
| HVAC or plumbing company | After-hours call or text | Job management system, on-call schedule, SMS | Visit booked, or emergency routed to the on-call tech | Gas smell, no heat in freezing weather, caller asks for a person |
| Dental office | Patient wants to reschedule | Practice management calendar, patient lookup | Appointment moved and confirmation sent | Patient describes pain or swelling, billing dispute |
| B2B services firm | Inbound web form or email | CRM, reps' calendars | Lead qualified and meeting booked | Big deal, custom pricing question |
| Online store | "Where's my order?" | Order system, carrier tracking, return policy | Status given or return label sent | Damage claim above a limit you set |
| Accounting team | Vendor invoice comes in by email | Inbox, document extraction, accounting software | Draft bill created for approval | Totals don't match the purchase order |
None of these agents is general-purpose. A narrow scope is what makes an agent testable, and being testable is what makes it safe to put in front of customers. Phone agents follow the same structure; how AI receptionists work walks through a single call from first ring to booking.
What types of AI agents are there?
Textbooks sort agents by how they make decisions. Russell and Norvig's Artificial Intelligence: A Modern Approach describes simple reflex, model-based, goal-based, utility-based and learning agents. The taxonomy is older than language models, but it still describes the reasoning styles.
If you're buying, it's more useful to sort agents by the job they do:
- Voice agents and AI receptionists answer and place calls, book appointments and transfer urgent callers.
- Sales agents reply to new leads in seconds, ask qualifying questions and book meetings on a rep's calendar.
- Customer support agents answer from your help content, look up accounts or orders, handle simple actions and escalate with context.
- Back-office agents sort a shared inbox, read documents or pull a report together from several systems.
When several of these work on the same customer and pass the case back and forth, you've got a multi-agent system. That's one form of what people now call agentic AI.
What do AI agents still do poorly?
AI agents are weakest where mistakes are expensive and the right call depends on judgment nobody has written down. Saying so before a project starts saves you from a failed one.
- Long chains of decisions. Every step carries some chance of error, and errors stack up. Ten-step autonomous tasks need checkpoints.
- Professional judgment. Clinical, legal and financial advice stays with qualified people. An agent can gather the details and route the case.
- Processes nobody has defined. If three people on your team handle the same request three different ways, the agent won't learn any of them well.
- Systems it can't reach. Information that lives in someone's head, or in software with no API, is invisible to it.
- Speed and cost at volume. Every reasoning step costs time and model usage. On a phone call, a two-second pause is noticeable.
How much autonomy should an AI agent have?
Give an agent the least autonomy that still takes the work off your plate. We think about it in four levels and only move up when reviewed transcripts back it up, never because a demo looked confident.
Four levels of agent autonomy
-
Step 1
Suggest
The agent drafts a reply or an action, and a person approves each one. A good fit for the first weeks of any agent.
-
Step 2
Act with confirmation
It acts after the customer or an employee says yes: "Want me to book Thursday at 3:30?"
-
Step 3
Act within limits
It acts on its own inside hard limits, like refunds under a set amount or bookings for certain appointment types only.
-
Step 4
Act and report
It acts and reports afterward. Save this for low-risk actions that are easy to undo.
Most customer-facing agents we build sit at level 2 or 3. Our guide on how to build an AI agent covers how to set those limits and test them before launch.
Do you need an AI agent?
An AI agent is worth scoping if most of these are true for the work you have in mind:
Is this a job for an agent?
If the last two aren't true yet, start there. No agent can fix a process nobody has defined. If you want to know what a first agent would involve, our breakdown of AI agent development cost explains what drives the build cost and the monthly running cost.
We've been building software since 2012, and we design agents around the integrations, handoff rules and monitoring that keep them dependable. If there's a conversation you'd like an agent to take over, missed calls or inbound leads for example, tell us about it and we'll tell you whether an agent is the right tool or a simpler automation will do the job.