A chatbot holds a conversation: it answers questions, from scripted replies or a language model, and hands off to a person when it gets stuck. An AI agent is given a task and uses tools to finish it, such as checking an order system, updating a record or drafting an email for someone to approve. The large language model, or LLM, is the engine inside an agent and inside any chatbot that writes its own answers; on its own it only reads and writes text, and it becomes a chatbot or an agent depending on the instructions, data and tools built around it.
So the useful question is not which label a vendor uses but what the system can read and what it can change. This guide compares the three, then covers cost, risk and fit for a small team. For chatbot pricing and launch questions, see our AI chatbot FAQ.
How do LLMs, chatbots and AI agents compare?
| LLM (the model) | Chatbot | AI agent | |
|---|---|---|---|
| What it is | A model that generates text from the text it is given | An application that holds a conversation, scripted or model-driven | A system that pursues a goal with a model, tools and instructions |
| What goes in and comes out | A prompt in, text out | A question in, an answer out | A task in, completed work out |
| Acts in your systems | No, not without tools around it | Rarely; it answers and hands off | Yes, through the tools it is allowed to use |
| What it knows about your business | General knowledge from training, plus whatever is in the prompt | The approved content you load into it | Whatever its tools and knowledge sources can reach |
| Typical failure | Confident, wrong text | A wrong answer given to a customer | A wrong action taken in a system |
| Controls that matter | Good prompts and reviewing output | Approved content and an easy handoff to a person | Scoped access, approval queues, logs and spend limits |
| What you pay for | Tokens processed | A subscription, or a charge per conversation or answer | Build work, model and tool usage per task, and supervision |
What is the difference between an LLM and an AI agent?
An LLM is the model: give it text and it returns text. An agent is a system built around a model. OpenAI's practical guide to building agents lists three parts, the model, the tools it can call and the instructions that set its behavior, and defines agents as systems that independently accomplish tasks on your behalf. The model decides what to do next; the tools are how anything actually happens.
The same guide draws the line plainly: applications that integrate LLMs but don't use them to control workflow execution, such as simple chatbots, single-turn LLM calls or sentiment classifiers, are not agents. An agent uses the model to manage the workflow, recognizes when the job is done, and can stop and hand control back to a person if it fails.
Is a chatbot an AI agent?
Usually not. Older chatbots followed scripts and buttons; newer ones use an LLM to answer in plain language from approved content. Either way, a chatbot's job ends with an answer. The line moves when a chatbot gets tools. A website assistant that can look up an order, book an appointment or change a delivery address is acting on your systems, and it needs the same controls as any agent: its own identity, limited access, approval for anything with money or legal weight, and a log of what it did.
Product names won't tell you which one you have. Ask what the system can read, what it can change and who approves the changes.
AI agents vs models: what are you paying for?
A model is billed by usage. As of October 2026, OpenAI's API pricing lists gpt-4.1-mini at $0.40 per million input tokens and $1.60 per million output tokens, and larger models cost more per token. A single model call is cheap; what you build around it is not.
A chatbot is usually bought as a subscription or charged per conversation. One Microsoft example: Copilot Studio, which builds both chatbots and agents, charges 1 Copilot Credit for a scripted answer and 2 for a generated one, with credits at $0.01 each pay-as-you-go as of October 2026, according to its billing rate table.
An agent adds tools and steps. The same table charges 5 credits for each agent action, and Microsoft prices agents in its Foundry Agent Service per call for inference plus tool usage, with container compute on top for agents you host there. The larger costs are the build, the integrations and someone reviewing what the agent does.
When does a business need an agent instead of a chatbot?
When the work doesn't end with an answer. OpenAI suggests reserving agents for workflows with complex decisions, rules that have become hard to maintain, or heavy reliance on unstructured data such as emails and documents, and says a deterministic solution may be enough otherwise. In a small business that looks like this:
- Chatbot: answering hours, prices, policies and order status from approved content, and collecting details for a person to follow up.
- Agent: reading an inbound request, checking two systems, drafting the reply and queuing it for approval, or chasing missing documents and recording who replied.
- Neither: a handoff that happens the same way every time, such as a web form that becomes a CRM record, which a plain workflow handles more cheaply.
Whichever you build, plan where a person steps in. OpenAI names two triggers: the system exceeding its failure limits, such as misreading a request several times, and high-risk actions such as cancellations, large refunds or payments.
Which fits a small team?
Start with a chatbot or an internal assistant that answers from approved content: it is quick to launch, and its worst failure is a wrong answer someone can correct. Move to an agent when you can name a process with clear inputs, a few systems to check, and a person who will approve its actions while it earns trust. Skip both for work that follows fixed rules, and choose the model last, once you know the instructions, data and tools the job needs.
How NetSys helps
Our AI agents service starts by asking whether a job needs an agent at all, or whether a workflow or a question-answering assistant would do. For agents, we write a charter covering the tools, data scope, stopping rules and owner, give the agent its own identity rather than a staff login, keep read access as the default, and put anything touching money, client records, outbound messages or system settings in an approval queue. We build in Copilot Studio for agents inside Microsoft 365 or on Anthropic Claude for work that crosses other systems, test against instructions hidden in emails and documents, and start every agent in shadow mode. For a short definition to share with your team, see our AI agent glossary entry.
Frequently asked questions
What is the difference between a chatbot and an AI agent?
A chatbot answers questions in a conversation and hands off when it can't. An AI agent completes a task by using tools, such as looking up records, updating systems or sending drafts for approval. If the system can change something in your business, treat it as an agent, whatever it is called.
What is an LLM?
A large language model is the AI model that reads and generates text. It powers AI agents and model-driven chatbots but does nothing in your systems by itself. Providers charge for it by the number of tokens processed, and the instructions, data and tools around it decide what it can do for you.
Can a chatbot become an AI agent?
Yes, once you give it tools that act, such as booking, ordering or updating records. At that point it needs agent controls: its own limited identity, approval steps for high-risk actions and a log of every action it takes.
Do AI agents need a person to approve their actions?
For anything sensitive or irreversible, yes, at least until the agent has a track record. OpenAI's guidance names high-risk actions such as payments, large refunds and cancellations, along with repeated failures, as the points to bring a person in.
What affects the cost and support of a chatbot or an agent?
A chatbot's cost follows conversation volume and the content you maintain for it. An agent's cost follows the number of systems it connects to, the model and tool calls per task, testing and supervision. Both need someone to review answers or actions and update them as your business changes.
Discuss ai agents for small business for your business.
Tell us about your current systems, the result you need and your timeline. We will discuss the work, responsibilities and pricing before you decide on an engagement.