What Is an AI Agent? A Plain Explanation for Business Owners

An AI agent is software you give a goal instead of a script. It reads the situation, decides the next step on its own, uses tools such as your calendar, your inbox or your customer records to act, checks what happened, and repeats until the job is done or it hands the case to a person. So if you are asking what is an AI agent, the short honest answer is: an AI model that can take actions in a loop, inside rules you set.

That difference matters for a service business, because most of your work starts with an enquiry that needs several small steps: reply, find out what the person needs, check that you can help, and agree on a time. A plain chatbot can manage the first step. An agent can manage all of them.

Below: how it works, how it differs from an LLM and a chatbot, one enquiry followed from ad to booked visit, and a checklist for telling a real agent from a relabeled chatbot.

What is an AI agent, in simple terms?

Think of a new employee. You do not hand them a flowchart for every possible conversation. You tell them the goal ("get new enquiries booked for a visit"), the rules ("we only work within 30 miles, we never quote a price before we have seen the job") and where things are ("the calendar is here, the service list is there"). Then you trust them to handle the conversation and to come to you when something is unusual.

An AI agent is the software version of that arrangement. OpenAI's practical guide to building agents puts it in one line: agents are "systems that independently accomplish tasks on your behalf." The guide names three core parts:

• A model: the large language model (LLM) that reads, writes and decides what to do next.
• Tools: connections to other systems that let it look things up or take action, such as reading a calendar, sending a text or updating a customer record.
• Instructions: the goal, the rules and the limits that define how it behaves.

Anthropic, the company behind Claude, draws the line in a similar place. Its guide Building effective agents separates workflows, where a model and tools are run "through predefined code paths", from agents, where the model directs its own process and decides which tools to use. In everyday words: a workflow follows the route you drew. An agent picks the route itself, inside the boundaries you drew.

How does an AI agent work? The loop in four steps

Every AI agent, whatever it is sold as, runs some version of the same loop:

1. Read the situation. A new message arrives or a lead form lands. The agent reads it together with the earlier conversation and the instructions you gave it.
2. Decide the next step. Should it answer a question, ask one, look something up or pass the case to a person? This is the part a script cannot do, because the right next step depends on what the person actually wrote.
3. Act with a tool. It sends the reply, checks the calendar, saves an answer to the lead record or creates a booking.
4. Check the result and go again. Did the person answer? Is the slot still free? Is the job done? Anthropic describes agents as needing "ground truth" from their environment at each step, such as the result of a tool call, to judge whether they are making progress.

The loop also needs a way to stop. Anthropic notes that agents can pause for human feedback at checkpoints or when they hit a blocker, and that it is common to add stopping conditions such as a maximum number of attempts. OpenAI's guide makes the same point for customer service: if the agent still does not understand what the customer wants after several attempts, it should escalate to a person.

That is the AI agent explained without the jargon. The model provides the judgment, the tools provide the hands, the instructions provide the rules, and the loop keeps it working until the task is finished or handed over.

AI agent vs LLM vs chatbot: what is the difference?

These three words are often used as if they meant the same thing. They do not.

An LLM (large language model) is the engine. Text goes in, text comes out. On its own, it only works with the text you give it: it cannot look in your calendar or send anything. IBM's explainer on AI agents points out that a standard LLM answers from its training data, while an agent uses tool calls to fetch up-to-date information.

A scripted chatbot is a decision tree with a chat window. It asks question one, then question two, and offers buttons for the answers. It works until the person types something the script did not expect.

An AI agent wraps an LLM with tools, instructions and a loop, so it can hold a real conversation and also get things done. OpenAI's guide is blunt about where the line goes: applications that use an LLM but do not let it control how the work is carried out, such as "simple chatbots, single-turn LLMs, or sentiment classifiers", are not agents.

A quick way to remember it: an LLM can talk, a chatbot can follow a script, and an agent can decide what to do next and then do it.

An AI agent example: one enquiry from ad to booked visit

Here is an illustrative example of an agent handling one enquiry with no script. The business is a made-up roofing company, but the pattern is the same for a clinic, a salon or a real estate office.

9:40 pm, Tuesday. Someone fills in a Facebook lead ad for roof repairs. The form holds a name, a phone number and one line: "Leak over the kitchen after the storm."

Seconds later. The agent reads the form and the owner's instructions: the service area, the jobs the company takes on, the questions to answer before a site visit (address, type of roof, how urgent it is) and the rule that prices are never quoted before a visit. It sends a text that acknowledges the leak, says the company can help and asks for the address.

9:52 pm. The reply is something no script would expect: "It's 14 Elm Road. Is that too far for you? Also, do you do gutters?" An answer and two questions in one message. The agent checks the address against the service area, confirms it is covered, answers the gutter question from the company's own service list and asks the next qualifying question: is it a tile roof or a flat roof, and is water still coming in?

9:55 pm. No answer. The lead has gone to bed.

8:15 am, Wednesday. Following the owner's follow-up rule, the agent sends one short message that does not repeat the first one: "Good morning. Is the leak still active? If so, we can come and take a look this week."

8:31 am. "Yes, still dripping. Tile roof." The agent now has what the owner asked for. It checks the calendar for free slots (on Google Calendar, for example, that is a single free/busy query for a time window), offers two times and books the one the lead picks.

8:33 am. The owner finds a summary in the team inbox: address, tile roof, active leak, also interested in gutters, site visit Thursday at 10 am. The full conversation is attached.

Notice what did not happen. Nobody wrote a flowchart for "asks about distance and gutters in the same message." Nobody had to remember to chase the lead in the morning. And the agent never mentioned a price, because the instructions said not to.

The same loop fits other businesses with a different goal and different questions. A physiotherapy clinic might ask about the injury and book a first assessment. A driving school might ask which license the person needs and book an intro call. A real estate agent might ask whether the person is buying or selling and book a valuation call. Only the instructions and the tools change.

What can an AI agent do for a small business, and what should it not do?

For a small or mid-size service business, the best first job for an agent is the one that slips most often when you are busy: inbound enquiries. In practice that is four tasks:

• Reply fast, at any hour. Leads from ads, forms, website chat and social messages get an answer while they are still thinking about you, not the next morning.
• Qualify. It asks the questions you would ask, in a normal conversation, and records the answers.
• Follow up on enquiries that go quiet. A polite nudge or two, then it stops.
• Book. A call, a consultation or a site visit goes straight into a real slot in your calendar.

What it should not do on its own matters just as much. OpenAI's guide recommends human oversight for high-risk actions, meaning actions that are sensitive, irreversible or high stakes. IBM calls it best practice to require human approval before an agent takes highly impactful actions. For a service business, that becomes a short list of cases that always go to a person:

• the customer asks for a human
• a complaint, or anyone who is clearly upset
• a complicated or unusual job that needs your judgment
• anything involving money, such as discounts, refunds or binding prices
• someone who is ready for a big purchase and deserves your personal attention
• emergencies and anything to do with health or safety

Agents also fail in predictable ways. IBM lists agents that keep calling the same tools over and over ("infinite feedback loops") among the risks, and Anthropic warns that the freedom that makes agents useful also brings higher costs and the potential for compounding errors. So you want a cap on follow-ups, a log of every action, and a way to take over any conversation.

How to tell if a tool is really an AI agent: a buyer's checklist

"AI agent" is now printed on many products that are scripted chatbots underneath. Before you pay, send it a few realistic enquiries yourself and check:

1. Does it handle the unexpected? Ask two questions in one message, or answer a question it has not asked yet. A script breaks; an agent adapts.
2. Does it act, or only talk? Can it see your real calendar and book a slot, or does it just send a link and hope?
3. Does it follow up by itself? Leave the conversation for a few hours. Does it come back once, sensibly, and then stop?
4. Does it know your business? Ask about your service area, your opening hours or a service you do not offer. It should answer from information you control, not make things up.
5. Can you set hard rules? For example "never quote a price" or "always hand over complaints". Push a little and check that the rule holds.
6. Does it hand over well? When it passes a conversation to you, do you get a summary and the full history, or only a notification?
7. Can you see everything it did? Every message, booking and handover, in one place your team can read.
8. Does it say it is an AI? In the EU, Article 50 of the AI Act, which applies from 2 August 2026, requires AI systems that interact directly with people to be designed so those people are informed they are dealing with an AI, unless that is obvious, at the latest at the first interaction. It is good practice everywhere else too.

If a product passes the first three tests, it behaves like an agent. If it also passes the last five, it is one you can trust with real customers.

Getting started: what to prepare before you switch one on

An agent is only as good as the instructions you give it. Before you connect one to your lead sources, write these down on a single page:

• The goal. For example: "Book a site visit with every qualified enquiry inside our service area."
• Your qualifying questions. Three to five, in order of importance, and which answers mean "not a fit for us".
• What it may say. Services, service area, opening hours, how a visit or consultation works, and what you never promise in writing (often prices).
• The handover rules. The list from the section above, adjusted to your business.
• Follow-up limits. How many messages, how far apart, and when to stop.
• Who gets notified, and where. One shared inbox beats five different apps.

Then test it on your own phone the way a customer would, awkward messages included, and read the first weeks of conversations yourself. Most fixes are one sentence added to the instructions.

Tools such as Konverto apply this pattern to inbound enquiries: they reply by SMS and email typically within a minute, ask your qualifying questions, send a follow-up if a lead goes quiet, book calls or site visits in your calendar and hand over to a person when a lead asks for one or raises a complaint. Judge Konverto, or any other tool, by the checklist above rather than by the word "agent" on the box.