Skip to Content

AI agent or chatbot? What an agent is and what it costs

How an agent differs from a chatbot, when your business needs one, and what decides its cost

Written by Sean Veldboer, Consultant at CBOS
Published 10 min read

An AI agent is software given a goal and the tools to reach it, so it can decide its next step and take it. A chatbot answers your question and leaves the doing to you, whether that's changing an order or sending a reply. ChatGPT is a chatbot with an agent mode, added in July 2025, that can work through tasks on the web.

What an agent costs depends less on the AI model than on what it's allowed to do, which of your systems it connects to and how much work it handles.

The difference shows up quickly in an ordinary request. Let's say a customer messages your sales team on a Thursday to move Friday's delivery to Monday and add two more cases. A chatbot on your website could explain your delivery policy and give the customer a number to call. An agent connected to your CRM and ERP could handle the request itself, in these four steps:

  1. Check the customer's account in the CRM, including whether it's on hold.
  2. See whether the extra stock is free for a Monday delivery in the ERP.
  3. Change the order and confirm the new date with the customer.
  4. Hand the conversation to a person if anything falls outside its rules.

For your business, the decision is whether to let the AI act in your systems. That permission is what makes an agent useful, and it also decides most of the risk and most of the cost.

What is an AI agent, in simple terms?

An AI agent is a language model with three things added: a goal, a set of tools and a loop. The goal is the job, such as "resolve this order change". The tools are the connections it may use, such as reading stock levels, updating a record or sending a message. The loop is how it works, by looking at the situation, choosing a step, taking it and checking the result before choosing the next.

This loop is what separates an agent from software that follows fixed instructions. Anthropic, which makes the Claude models, describes agents as systems where language models "dynamically direct their own processes and tool usage". In plainer terms, the steps aren't written down in advance; the agent works out their order from what it finds.

The tools decide what an agent can reach. An agent with no tools is a chatbot with ambitions, and an agent with a login to your whole database is a risk your IT team will rightly refuse. Most of the design work for a business agent goes into the space between those two, deciding which tools it gets and what each one may do.

What is a simple example of an AI agent?

The order change above is a simple business example, because the agent reads from two systems, makes one change and knows when to stop. Other common examples follow the same pattern of reading, deciding and acting:

  • A service agent that checks a claim's status in your system, answers the customer and opens a follow-up task for the claims team.
  • A finance agent that matches supplier invoices against orders and prepares the payment, leaving the approval to a person above a set amount.
  • A sales agent that reads a new enquiry, finds the account in the CRM and drafts a quote for a salesperson to check.

In each one the agent does the preparation, and a person keeps the decisions that carry real consequences. That split is usually where a business starts, because it lets you see how the agent behaves before you give it more to do.

Is ChatGPT a chatbot or an AI agent?

ChatGPT is a chatbot with an agent mode, so the answer depends on which one you're using. In its normal mode you ask, it answers, and even when it searches the web for you, you decide what happens next.

OpenAI added ChatGPT agent in July 2025, and it works on what OpenAI calls "its own computer". It can browse websites, work through tasks on them and produce documents such as presentations. It combined two earlier OpenAI tools, Operator for web tasks and deep research for gathering information, and at launch it came with OpenAI's paid plans, switched on as agent mode from the tools menu.

For a business, the limit is where ChatGPT agent acts. It works on the open web and in whatever apps you connect to it, under OpenAI's rules. It doesn't know your stock levels, your customers' credit limits or who in your company may approve a discount, unless someone connects those systems and sets those permissions.

A general agent can do useful work on public websites, but changing an order in your Sage or Odoo system, safely and under your rules, takes an agent built around that system.

What's the difference between an AI agent and a chatbot?

A chatbot holds a conversation, while an agent can also carry out the work the conversation is about. The two often use the same language models, so the difference lies in what each is connected to and allowed to do.

ChatbotAI agent
What it does with a requestAnswers itAnswers it and can complete it
Systems it touchesUsually none, or read-onlyYour ERP, CRM, email or other systems, with set permissions
Who takes the actionThe person reading the answerThe agent, within its rules, with sign-off where you ask for it
What can go wrongA wrong answerA wrong answer and a wrong action
What it needs before going liveContent and toneContent, permissions, sign-off rules, an audit trail and testing

The line blurs at the edges. TechTarget points out that a chatbot with a few limited actions, such as resetting a password, can fairly be called an agent. Vendors stretch the label too; Gartner calls it "agent washing" (chatbots and older automation tools rebranded as agents). It estimated that only about 130 of the thousands of vendors claiming agentic AI were offering the real thing.

The practical test is to ask what the tool is allowed to change in your systems. If the answer is nothing, it's a chatbot, whatever the brochure calls it.

Do you need an AI agent, or would a simpler workflow do?

Many jobs that look like they need an agent are better served by a workflow with fixed steps. Anthropic draws this line in its guidance on building agents: workflows follow "predefined code paths", while agents decide their own steps. It recommends "finding the simplest solution possible, and only increasing complexity when needed", because agents trade extra cost and time for better results on harder tasks.

A CBOS project for a healthcare claims administrator shows the simpler route. The platform collects medical-scheme statements, reads each scheme's layout, reconciles the payments against the bank and sends only the differences to a person. The steps are the same every time, which is why a fixed workflow does the job.

An agent earns its place when the next step depends on what the request says. The order change from the opening can arrive in a hundred different wordings, with different accounts, stock positions and exceptions. Writing a fixed rule for each one would take longer than building the agent.

A useful rule of thumb is that if you can draw the process on a whiteboard and it doesn't change from one case to the next, you should build a workflow. If the drawing keeps growing branches, an agent may be worth it.

How does an AI agent connect safely to your ERP and CRM?

An agent connects to your ERP and CRM through tools that each do one thing, so it never needs a login to the whole database. One tool might read a customer's balance, another might reserve stock, and a third might change a delivery date. Each tool can be limited to the branches, records and amounts that the person or process behind it is allowed to touch.

These limits are set in the design, before anything is built. On CBOS's AI capabilities page, that stage decides "what it may do on its own and where a person signs off". Its invoice example sends any payment over R15 000 to a person for approval. Every action is recorded, so you can see what the agent did, when and on whose behalf.

The same limits are built into the AI agents CBOS builds. They connect to a company's ERP and CRM and act on them, with permissions set per user, sign-off where the business wants it and the data stored and processed in South Africa. They're designed and built like any of the bespoke systems CBOS makes, which connect to your ERP, CRM, accounting and older applications.

An agent usually starts by preparing work for people to approve, and takes on more as it proves itself. My earlier article on creating functional agentic AI covers that move from advising to acting in more detail.

How much do AI agents cost?

The cost of an AI agent depends far more on the work around the model than on the model itself. Six things decide it:

  • What the agent is allowed to do. One that only reads and answers is the cheapest kind, and every action it may take needs rules, sign-off points and testing, so each new permission adds to the build.
  • The systems it connects to. Each connection is its own piece of build work, and a modern ERP with a documented API is quicker to connect than an older system that has to be reached another way.
  • The state of your data. Duplicate customers, inconsistent product codes and missing fields confuse an agent as much as they confuse a new employee, and cleaning them up is often part of the project.
  • How much work it handles. Running costs grow with the number of tasks and the number of steps in each one, because each step carries the context of the steps before it. Anthropic found that agents typically use about four times as many tokens (the units AI models are billed in) as a chat does.
  • Testing and oversight. An agent's behaviour has to be tested across many realistic cases before it goes live, and someone has to review the exceptions it hands over.
  • Change after go-live. Models are updated, your processes change and new exceptions turn up, so an agent needs maintenance like any other system.

How you pay for those costs depends on whether you buy an agent or have one built. Off-the-shelf agents sold as a service are usually priced per conversation, per resolved task or per user, which suits standard jobs such as answering common customer questions. An agent built around your own systems has a build cost and a running cost, and suits work that depends on your data, rules and approvals.

In both cases scope moves the cost most, because one well-chosen task, done properly, costs a fraction of an agent meant to do everything.

Why do so many AI agent projects get cancelled?

The same cost drivers explain most of the failures. In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of "escalating costs, unclear business value or inadequate risk controls".

Each of those reasons goes back to a decision made at the start. Costs escalate when the scope keeps growing, and value stays unclear when the task chosen has no result you can measure. Risk controls fall short when permissions are bolted on after the agent is built, and none of these is a problem with the model itself.

The safer path starts with one task where people lose real time, such as order changes or invoice matching. Decide what the agent may do on its own and where a person signs off, and measure the result before giving it more. That's also how a CBOS project starts, with a discovery call about what you're trying to achieve and why the current process causes friction.

Frequently asked questions

Is ChatGPT an AI agent?

Normal ChatGPT is a chatbot that answers questions. Its agent mode, added in July 2025, is an agent, because it can browse websites and complete tasks on its own computer. It works on the open web and in the apps you connect to it, and it only reaches your company's systems if someone connects them and sets the permissions.

How much does ChatGPT agent cost?

At launch, ChatGPT agent came with OpenAI's paid plans. OpenAI changes its plans and usage limits often, so its pricing and help pages have the current details.

What are the 5 types of agent in AI?

The classic five, from AI textbooks, are simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents and learning agents. They describe how an agent decides, from reacting to a rule up to improving from experience. Business agents built on language models today are best described as goal-based, with some learning from feedback.

What's the difference between agentic AI and an AI agent?

"Agentic AI" is the broader approach of AI that pursues goals and takes actions, while an AI agent is one system built that way. A business might run several agents, each with its own job and permissions, under one agentic AI setup.

Can I create my own AI agent for free?

For personal use, yes. Several AI providers offer free tiers, and open-source tools let you build simple agents for things like sorting email. An agent that acts on company systems is different, because it needs secure connections, permissions per user, sign-off rules, an audit trail and testing.

How do I create an agent in ChatGPT?

On a paid plan, you switch on agent mode from the tools menu in a chat and describe the task, and ChatGPT then works through it on its own computer. OpenAI's menus change often, so its help pages have the current steps.

Can AI agents run on a phone?

Most business agents run on servers, and you reach them from your phone through an app, WhatsApp, Microsoft Teams or a web chat. Your phone carries the conversation, while the agent's work happens in the systems it's connected to.

About the author

Sean Veldboer

Consultant at CBOS

CBOS designs and builds platforms, portals, mobile apps and integrations for companies in South Africa and around the world, and has done since 2014. More about CBOS

Have a project like this in mind?

Book a discovery call and we'll look at how your operation runs, then put the scope and price in a written proposal.

Is ChatGPT safe for your company's data? What staff and employers need to know under POPIA
What happens to what staff paste into AI, what POPIA expects of the employer, and what to do if it's already happened