What is an AI agent? A plain explanation for business owners

An AI agent is software that you give a goal and a set of tools, and that then works out the steps on its own, uses the tools and checks what happened. That is the main difference from a chatbot, which only answers, and from an automation, which only does what you wrote down in advance.

The people who build these systems draw the same line. Anthropic, the company behind Claude, separates two kinds of system in its article Building effective agents. Workflows, it says, are systems where AI models and tools are orchestrated through predefined code paths. Agents are systems where the model directs its own process and tool use. In plain terms: a workflow follows your script, an agent writes part of the script as it goes.

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

The table sets the three side by side. It is our summary of the definitions in the sources listed at the end, not a rating of any product.

What does the same job look like done three ways?

Take an example. A six-room guest house in Wilderness, which is a made-up business, gets this email at 21:40: "Do you have a family room for the 12th to the 15th of January, and is breakfast included?"

As a chatbot

The chatbot on the website can answer the breakfast question, because that is written in its notes. It cannot see the booking calendar, so it says "please phone us to check availability". The guest still has to wait for the owner, which is the gap you wanted to close.

As an automation

The rule says: when an email arrives with the word "availability", send a template reply and add a line to a spreadsheet. It fires every time, which is its strength. But the guest asked about a family room and breakfast in one sentence, and the template answers neither. If the email arrives in a form the rule did not expect, nothing happens.

As an AI agent

The goal is "answer enquiries and line up a booking for the owner". The agent reads the email, sees the two questions, checks the calendar for the 12th to the 15th, finds the family room free, looks up the rate and breakfast policy in your price list, and writes a reply. It then checks that the reply names the right dates and the right room before it is shown to anyone. In draft mode, the owner reads it in the morning and taps approve. For more on this use case, see AI for guest houses.

The three are not rivals. A good setup often uses an automation for the predictable part, such as logging every enquiry, and an agent for the part that needs judgement, such as reading a messy email.

How does an agent actually work?

Anthropic describes agents as typically just AI models using tools based on feedback from their surroundings, in a loop. The agent takes a step, looks at what came back, and decides what to do next. The same article says it is crucial for agents to gain "ground truth" from the environment at each step, such as the results of a tool call, to judge their progress.

OpenAI, in its agents documentation, puts it this way: "Agents can plan and complete tasks using tools, work with other agents, and maintain context across steps." In our reading, the two descriptions agree on the main points: the agent uses tools, takes several steps and keeps track of where it has got to.

What tools can an agent use in a small business?

A tool is just something the agent is allowed to look at or do. An agent has no access to anything until someone connects it. The usual list for an owner-run business:

• Email: read an enquiry, draft a reply, file the thread.
• Calendar: check what is free, hold a slot for the owner to confirm.
• WhatsApp: answer a customer who messaged you. Sending messages through the official business platform has its own rules and costs, covered in what the WhatsApp Business API costs.
• Accounting software: look up whether an invoice is paid, draft a reminder. Moving money should stay with a person.
• A spreadsheet or customer list: find a past customer, log a new lead, update a status. Our CRM agent and admin agent are built around this kind of work.
• Your own price list and rules: the document the agent reads so it quotes your rates, not invented ones.

Each connection is a permission you choose to give. An agent that can only read your calendar cannot cancel a booking. Start with the fewest tools that let it do one job.

What can go wrong with an AI agent?

Three things, all avoidable with care.

• It can be confidently wrong. An agent writes fluent text whether or not it has the right facts. If your price list is out of date, it will state the old price calmly.
• It needs clear rules and good information. "Be helpful" is not a rule. "Never promise a date, offer a deposit link, or quote below the rate card" is. The agent is only as good as the notes and data you give it.
• It can take a wrong step you did not foresee. Because it chooses its own steps, you cannot list every outcome in advance the way you can with a rule. That is the price of flexibility.

There is also a legal side once an agent handles customers’ personal details. POPIA and AI agents explains what a South African owner should check.

What does an AI agent cost to run?

We cannot give one honest figure, because cost depends on the job. Anthropic notes that agentic systems often trade latency and cost for better task performance, and that you should consider when that trade makes sense. An agent does more work per task than a rule does, and that work is not free.

The costs fall into four broad parts:

• A set-up cost, for connecting the tools and writing the rules.
• A monthly fee for the agent or the platform it runs on.
• Usage costs, such as the AI model itself and, for WhatsApp, per-message charges.
• Your time, especially in the first weeks, reading and approving drafts.

For one reference point from us: Never Miss a Lead, our agent for calls and email, with WhatsApp listed as coming, is open to a founding waitlist and is not on sale yet. Its published price is R2,395 a month excluding VAT, or R1,495 a month excluding VAT for the first 50 founding members. See pricing. Custom solutions are priced per build and scoped on a call. Whether any agent pays for itself in your business is a separate question, which is AI worth it for a small business in South Africa helps you work out.

Do you need an agent, or is a rule enough?

Anthropic recommends finding the simplest solution possible and only adding complexity when needed. That is good advice for a small business too. A quick test:

1. Write the job as a ruleIf you can finish the sentence "when X happens, always do Y" with no exceptions, use an automation. It is cheaper and easier to trust.
2. List the exceptionsIf the job keeps needing "unless" and "it depends", such as which room, which date, which customer, an agent may earn its place.
3. Check the information existsAn agent needs your prices, hours and rules written down. If they live in your head, write them first. That step helps whichever tool you pick.
4. Pick one job and one resultChoose a single result you can measure, such as the share of enquiries answered the same day. Run it in draft mode and judge it on that.

If the job is mostly answering the same ten questions, a chatbot or a template may be all you need. If you want help with a specific job, we build custom solutions, scoped on a call. You can book a call to talk it through.