The short answer
A chatbot answers questions. An automation runs the same fixed steps every time something happens. An AI agent is given a goal, works out the steps itself, uses your tools to carry them out and checks the result. Pick the simplest one that handles your real cases: a chatbot when people need answers, an automation when the work follows a fixed path, and an agent only when the right next step depends on what the last step found.
Most businesses that ask us for an agent need one of the other two, or a mix of all three. A fixed workflow is cheaper to build, cheaper to run and far easier to fix when something goes wrong. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs, unclear business value or weak risk controls (Gartner, June 2025). Choosing the right tool at the start is how you stay out of that 40%.
What is the difference between an AI agent, a chatbot and an automation?
The difference is who decides what happens next. With a chatbot, the customer drives: they ask, it answers, and the conversation is the product. With an automation, you decided every step in advance: when an invoice lands in the inbox, read it, check it against the purchase order, post it to the ledger. With an agent, the software decides: you give it an objective, and it chooses which tools to call and in what order, based on what it finds along the way.
Modern versions of all three can use the same language models, which is why the labels blur. A chatbot built on Claude or GPT can sound very capable, and an automation can include an AI step that reads a messy email. Neither one is an agent until it plans its own steps and takes actions in your systems.
| Chatbot | Automation workflow | AI agent | |
|---|---|---|---|
| Main job | Answers questions | Runs the same steps every time | Plans steps and acts toward a goal |
| Who decides the next step | The user | You, when it is built | The agent, within limits you set |
| Takes actions in your systems | Rarely, one or two simple ones | Yes, fixed ones | Yes, chosen case by case |
| When something unexpected happens | Hands over to a person | Stops or alerts a person | Tries another route, then escalates |
| Typical Zeebrix build cost | $10,000 to $40,000 | $12,000 to $30,000 | $18,000 to $40,000 |
| Typical time to go live | 4 to 8 weeks | 4 to 6 weeks | 4 to 8 weeks |
| Easiest to fix when it goes wrong | Medium | Easiest | Hardest |
When is a chatbot enough?
A chatbot is the right tool when the job is answering, and the answers already exist somewhere: your website, help articles, policies, price lists or product manuals. Built properly, it finds the relevant passage, answers with a link to the source, and says it doesn't know when the answer isn't there.
It stops being enough when people want things done, not explained. "When is my order arriving?" is a chatbot question if the bot can look up one order status. "Move my delivery to Friday and refund the shipping" is not. Most bots can safely take one or two simple actions, such as creating a support ticket or booking a slot in a calendar. Past that point you are building an agent, whether or not you call it one.
Our guide to AI chatbot costs for small business compares ready-made chatbot tools with a custom build. These are the jobs where a chatbot usually does the whole job:
- Customer questions about products, policies, prices and opening hours
- Internal help desks over HR policies, IT guides and manuals
- Lead capture on your website, with a handover to sales
- First-line support that creates a ticket when it cannot answer
When is a fixed automation the better choice?
Choose an automation when the work follows the same path almost every time. Invoices, timesheets, new lead forms, order confirmations and weekly reports are classic examples. Something triggers the workflow, it runs a set list of steps, and it behaves the same way on every run. Tools like Zapier, Make and n8n handle the simple versions well. A custom build makes sense when the volume is high, the data is sensitive, or the steps cross systems those tools can't reach.
AI fits inside an automation as one step, not as the boss. A language model is very good at reading a messy email or a scanned invoice and turning it into clean fields. The workflow around it then does the predictable part: check the numbers, post the record, and alert a person when the model is unsure. You get most of the value of AI with very little of the unpredictability, which is why we look at this option first on every AI automation project.
Typical jobs for an automation with an AI step:
- Invoice and receipt processing into your accounting system
- Sorting and routing support tickets or enquiries
- Copying lead details from forms and emails into the CRM
- Scheduled reports pulled together from several systems
When do you actually need an AI agent?
You need an agent when the task branches: the right next step depends on what the previous step returned, and there are too many paths to draw as a flowchart. Think of researching a prospect across several sources, resolving an exception that doesn't match any rule, reconciling records that disagree across three systems, or onboarding a customer whose paperwork is incomplete in a different way each time.
That flexibility is the point, and it is also the risk. Every tool an agent can call is a permission you have given it, and it can take a wrong action with full confidence. So a production agent needs limits that a chatbot or a workflow doesn't: a cap on steps, cost and time; a person's approval before anything it cannot undo; a full log of its decisions; and a trial period where it proposes actions and your team approves them before it acts alone. Our AI agent development page explains how we build each of these.
Jobs where an agent earns its extra cost:
- Research that pulls from many sources and decides what matters
- Handling exceptions that fall outside your normal rules
- Reconciling data that disagrees across several systems
- Requests from customers or staff that need different steps each time
A 5-question test to pick the right one
Answer these in order. The first clear yes usually tells you what to build, and the last question applies whatever you choose.
- Is the job mostly answering questions from information you already have? Build a chatbot.
- Does the work follow the same steps at least 8 times out of 10? Build an automation, with an AI step if the inputs are messy. That ratio is our rule of thumb, not a law.
- Do the remaining cases matter enough to automate, or can a person handle them in a few minutes? If a person can, stop at the automation.
- Does finishing the job mean choosing between several tools or paths based on what you find? That is an agent.
- Can a wrong action be undone cheaply? If not, keep a person approving those actions, whatever you build.
Most businesses end up with a mix
In practice the three work together. Take a trades business. A chatbot on the website answers questions about service areas and prices and books a quote visit. An automation turns each booking into a job in the scheduling system and sends the confirmations. Once the volume justifies it, an agent chases unpaid invoices: it reads each reply and decides whether to send a reminder, offer a payment plan or flag the account to the owner.
Building it in that order is cheaper and safer. Each layer proves itself before the next one depends on it, and the agent only takes the work the cheaper tools can't handle. Our guide to AI automation for trades businesses walks through this kind of setup in more detail.
What does each one cost to build and run?
These are Zeebrix fixed-price ranges as of November 2026, for a build connected to your own systems and data. Ready-made tools cost far less to start, and for simple needs they are the right answer.
Running costs are similar for all three in a small or mid-sized business: usually $200 to $1,500 a month for model usage and hosting. Agents tend to sit higher in that range because each task takes several model calls, and an agent without limits can loop and run up the bill, which is one more reason for step and cost caps. For the full breakdown, see how much an AI agent costs to build.
- Chatbot on your own data: $15,000 to $40,000, live in 4 to 8 weeks
- Chatbot added to an app or platform you already run: $10,000 to $25,000
- Automation workflow with an AI step, one process: $12,000 to $30,000, live in 4 to 6 weeks
- Custom AI agent with two or three system integrations: $18,000 to $40,000, live in 4 to 8 weeks
- Several agents running a whole process end to end: $50,000 to $120,000, 10 to 16 weeks
How to spot agent washing before you buy
Gartner uses the term agent washing for vendors that relabel chatbots, AI assistants and robotic process automation as agents. It estimates that only about 130 of the thousands of vendors claiming agentic AI offer the real thing (Gartner, June 2025).
Three questions cut through it. Can you show me, in a log, the steps it chose on a real case? What does it do when a tool returns an error or an unexpected result? Which actions need a person's approval, and where is that enforced? A real agent has clear answers to all three. A relabelled chatbot usually has no decision log at all, because it never made any decisions.
The same test works in reverse. If a vendor is selling you an agent for a job that follows the same steps every time, ask why a cheaper workflow wouldn't do. A good partner will tell you when it would.
How we help you choose
Every project starts with a free discovery call where an engineer looks at the job and a sample of your real cases. If a chatbot or a workflow will do, we say so and quote that instead, because it costs you less and goes live sooner. If the job really needs an agent, we quote a fixed price for a fixed scope and build the limits described above into it from day one.
US clients work with our San Jose team and Australian clients with our Perth team.
Frequently asked questions
Is a chatbot an AI agent?
Not usually. A chatbot answers questions in a conversation. It becomes an agent only when it can plan several steps and take actions in your systems on its own, such as updating records or sending messages. Many products sold as agents are chatbots with a new label.
Is ChatGPT an AI agent?
ChatGPT is mainly a chat assistant, although OpenAI, Anthropic and Google have all added agent modes to their assistants that can browse and complete tasks. For business use, the question that matters is whether a tool can act inside your own systems with permissions and logs that you control.
Can we start with a chatbot and add agent features later?
Yes, and it is often the best path. Start with a chatbot that answers well, then give it one or two safe actions, such as creating a ticket or booking a time. If those actions grow into multi-step work across several systems, that part becomes an AI agent built on the same knowledge base.
Is an AI agent more expensive to run than an automation?
Usually, yes. An agent makes several model calls per task because it plans, acts and checks its work, while an automation with one AI step makes one call. For most small and mid-sized businesses either one runs at $200 to $1,500 a month including hosting, but an agent needs step and cost limits so one bad case cannot loop and run up the bill.
Can Zapier or Make do this instead of a custom build?
For simple, fixed workflows, yes, and they are often the cheapest answer. Both now offer AI steps and basic agent features. A custom build makes sense when the volume is high, the data is sensitive, the steps cross systems those tools can't reach, or you need approval rules and logs that you control.
Not sure which one your job needs?
Show us the job and a few real cases. After a free discovery call we tell you whether a chatbot, an automation or an AI agent fits, and send a fixed-price quote for that.
Book a free discovery call
