AI Agent vs AI Chatbot vs AI Automation: Which Does Your Business Need?
How an AI agent, an AI chatbot, and AI automation differ, what each costs per task to run, and a two-question test to pick the right one.
On this page
Ask three vendors what your business needs and you'll get three answers. One sells an AI chatbot, one sells AI automation, and one says everything is an AI agent now. The words get used so loosely that a lot of businesses pay agent prices for a job a simple workflow could do, or put a chatbot on a problem nobody actually types about.
This guide sorts it out. You'll get a plain definition of each, a two-question test to pick the right one for a task, what each costs to run per task, and the math for when an AI agent is worth it. AI voice agents are in here too, since a lot of work still starts with a phone call.
Key takeaways
- AI automation runs steps you can write down ahead of time. An AI chatbot answers people who type to you. An AI voice agent does the same on the phone. An AI agent works out its own steps to reach a goal.
- Two questions pick the right one. Does a person start it by talking to you, and can you write every step down before it runs?
- Running costs are far apart. An AI step in an automation costs about $0.0002 to $0.004, and a six-step agent task costs about $0.16 on Claude Sonnet 5.5.
- Build an AI agent only when the task needs judgment, the rules are too tangled to keep up, or the work runs on messy documents and text. Otherwise a workflow is cheaper and more predictable.
- Most real systems use more than one. A chatbot or voice agent takes the request, and automation or an agent does the work behind it.
What's the difference between an AI agent, an AI chatbot, and AI automation?
AI automation follows fixed steps, an AI chatbot holds a conversation, and an AI agent works out its own steps toward a goal. Chatbots and voice agents are how people reach you. Automation and agents are how the work gets done behind the scenes.
That second point is the one most comparisons miss. These aren't four rungs on one ladder. Two of them are front doors and two of them are back offices, and you'll often need one of each.
| AI automation | AI chatbot | AI voice agent | AI agent | |
|---|---|---|---|---|
| What starts it | An event, like a form, email, or new record | A person typing | A person calling | A goal or a request |
| How it decides | Fixed steps, with AI for one step like sorting | Answers from your content and follows set flows | Like a chatbot, out loud and in real time | Picks its own next step and tool |
| Best at | Repeat work that's the same every time | Questions, booking, and tickets on your site | Answering, qualifying, and booking calls | Judgment calls on messy inputs |
| Main risk | Breaks quietly when an app changes | Wrong answers when your content is thin | Slow replies feel like dead air | Stacked mistakes and higher cost |
| Common tools | n8n | Botpress, Dify, Langflow | Vapi, Retell AI | n8n, Claude, ChatGPT |
AI automation
AI automation is a workflow with fixed steps that runs when something happens, like a new form, an email, or a paid invoice. The AI part is usually one step in the middle, such as reading an email and deciding which team it goes to. Anthropic calls this kind of system a workflow, where models and tools run "through predefined code paths." Every run takes the same route, so it's cheap, fast, and easy to check.
AI chatbot
An AI chatbot is a conversation in text, on your website, in your help center, or inside your team's tools. A good one answers from your own content, asks a few questions to qualify a visitor, books a call, or opens a ticket and hands off to a person.
OpenAI's guide to building agents lists "simple chatbots" among the AI apps that aren't agents, because the model answers but doesn't run the work. Once a chatbot starts looking up records and taking actions in your systems, it's doing agent work, and you should plan and test it like an agent.
AI voice agent
An AI voice agent is the same idea on the phone. It answers calls in a natural voice, qualifies the caller, and books them into your calendar. The big difference is time. A two-second pause is fine in a chat window but feels broken on a call, so voice needs fast models, handling for interruptions, and a clean way to transfer to a person.
AI agent
An AI agent gets a goal instead of a script. It decides which step to take next, which tool to use, and when it's done. Anthropic describes agents as systems where models "dynamically direct their own processes and tool usage." That's both the upside and the cost. An agent handles cases you didn't plan for, but it costs more per task, and its mistakes can stack up over several steps.
A two-question test to pick the right one
Ask who starts the work and whether you can write every step down. Those two answers point to one of the four, or to a pair that works together.
- Does a person start it by talking to you? If they type, put an AI chatbot at the front. If they call, put an AI voice agent there. If nobody's talking, like when a record changes or an email lands in an inbox, go straight to the second question.
- Can you write every step down before it runs? If you can, it's AI automation, even when one step needs AI to read or sort something. If you can't, check the three signs below before you call it an agent.
Put the two answers together and you get this grid.
| You can write the steps down | The steps depend on the case | |
|---|---|---|
| A person is talking to you | AI chatbot or voice agent with set flows | AI chatbot or voice agent with an agent behind it |
| An event or system starts it | AI automation | AI agent |
Three signs a task needs an AI agent
OpenAI's guide names three kinds of work where agents beat rule-based automation.
- Judgment calls full of exceptions, like deciding whether to approve a refund.
- Rules that have grown so long and tangled that updating them is costly and error-prone, like a vendor security review.
- Work that depends on reading messy text and documents, like processing a home insurance claim.
If a task doesn't clearly fit one of these, OpenAI says "a deterministic solution may suffice." Anthropic makes the same point and recommends the simplest solution that works, which "might mean not building agentic systems at all."
Check it against your own data
You don't have to guess which side of the line a task is on. Pull the last 50 times it happened and try to write the rule for each one.
If 45 of them fit five or six rules, build AI automation and send the other five to a person. If you keep writing new rules and still can't cover half of them, that's agent work. Most tasks land closer to the first case than people expect.
What each one costs to run per task
Running costs range from a fraction of a cent for an AI step in a workflow to about $0.16 for an agent task. The build is usually the bigger number, and our cost guides break that down for each one. These figures come from the worked examples in those guides, at Claude's list prices at the time of writing.
| Option | One unit | Running cost per unit | Full breakdown |
|---|---|---|---|
| AI automation | One AI step, like sorting an email | About $0.0002 on Claude Haiku 5.5 to $0.004 on Sonnet 5.5, plus n8n from about $22 a month for 2,500 runs | AI automation cost |
| AI chatbot | One resolved conversation | About $0.05 in model fees on a custom build, or $0.50 to $2.00 on a per-resolution tool | AI chatbot cost |
| AI voice agent | One two-minute call | About $0.28 on a voice platform, or about $2 on a per-call plan | AI voice agent cost |
| AI agent | One six-step task | About $0.16 on Sonnet 5.5, or about $0.35 counting the time a person spends fixing misses | AI agent cost |
An agent costs more because it makes several model calls per task and reads more at each one. Anthropic puts it plainly. Agentic systems "trade latency and cost for better task performance." That trade is worth it when the performance matters, and wasted when it doesn't.
What it costs to pick the wrong one
Picking wrong costs money in both directions. Here's a worked example of each.
An agent where a workflow would do
Say you sort 3,000 inbound emails a month into the right team's queue. As one AI step in an automation on Sonnet 5.5, that's $0.004 an email, or $12 a month. Built as a six-step agent at $0.16 a task, it's $480 a month.
That's $144 a year against $5,760 for the same sorted inbox. And the workflow sorts the same way every time, while the agent has six chances per email to wander off.
A workflow where an agent belongs
Now say you review 600 client documents a month, and fixed rules clear 70% of them. The other 180 go to a person who spends 10 minutes on each, at $35 an hour. That's 30 hours and about $1,050 a month in staff time.
Suppose an agent takes those 180 and fully handles 60%, sending the rest to a person. Each one it clears saves about $5.83 of staff time. Averaged across every exception, that's 0.6 × $5.83 minus $0.16 in model fees, or about $3.34 per exception.
The agent also has fixed costs. Use the example from our agent cost guide, a $30,000 build spread over 24 months plus $1,500 a month for upkeep, which comes to $2,750 a month.
Break-even exceptions per month = monthly fixed cost ÷ saving per exception
= $2,750 ÷ $3.34
≈ 820At 180 exceptions a month, the agent doesn't pay back. At 1,000 a month, it saves about $590 after its fixed costs. So the same task can be a bad agent project at one volume and a good one at another.
Below the break-even point, there's a cheaper middle option. Use AI automation with one AI step that sorts each exception and writes a short summary for the person who handles it. That cuts the reading time without the agent's fixed cost, and you can move up to an agent once the volume is there.
How they work together in one business
Most working systems use more than one of the four. Take a 25-person accounting firm in tax season as an example.
- A website AI chatbot answers "what do I need to send you?" from the firm's own checklist, books calls, and opens a ticket when a client is stuck.
- An AI voice agent picks up overflow calls, answers the same questions, and books appointments into the calendar.
- AI automation files every upload into the right client folder, updates the practice software, and sends a receipt.
- An AI agent reads each client's documents against last year's return, lists what's missing, and drafts the request for a staff member to approve.
Only one of those four is an agent, and it's the one doing the reading and judging. Everything else follows fixed steps or set conversation flows, which keeps the running cost low and the behavior predictable.
Add a person where mistakes are expensive
OpenAI's guide names two moments to bring in a person. One is when the agent keeps failing, for example when it still can't tell what a customer wants after several tries. The other is before high-risk actions, which it describes as sensitive, irreversible, or high-stakes, like canceling orders, approving large refunds, or making payments.
That applies to all four, not just agents. A chatbot should hand off when it can't answer. A voice agent should transfer the call. An automation should alert someone the moment a step fails, instead of failing quietly for a week.
Common mistakes when choosing
- Buying an "agent" that's really a chatbot. Gartner calls this agent washing, where vendors rebrand chatbots, assistants, and robotic process automation (RPA) as agents. It estimated that only about 130 of the thousands of vendors offer real agentic AI. Ask what actions the product takes in your systems and what it does when it isn't sure.
- Starting with the agent. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, because of rising costs, unclear business value, or weak risk controls. Start with the workflow and add judgment where the exceptions pile up.
- Putting a chatbot on a phone problem. If most of your requests come in by phone, a chatbot on your site won't catch them. Count calls and chats for a month first. Our guide to an AI receptionist for small businesses covers which calls are worth handing to AI.
- Leaving no way out to a person. Every one of the four needs a handoff, and it should be tested as hard as the happy path.
- Testing on the demo instead of your own cases. Run last month's real requests through it and count the misses before you switch anything on.
Conclusion
Use AI automation when you can write the steps down, an AI chatbot when people type to you, an AI voice agent when they call, and an AI agent only when the task needs judgment that rules can't hold. Run the 50-case test on your own data and check the break-even math before you pay for an agent.
If you'd rather have it built for you, we build AI agents, AI automation, AI chatbots, and AI voice agents in your own accounts, and we'll tell you which one your process actually needs.
Frequently asked questions
Not by itself. A chatbot that only answers questions isn't an agent. Once it looks up records and takes actions in your systems, choosing the steps itself, it's doing agent work and should be planned like an agent.
No. AI automation follows steps you set ahead of time, with AI handling one part like sorting or summarizing. An AI agent picks its own steps toward a goal, which makes it more flexible and more expensive to run.
AI automation. An AI step in a workflow costs about $0.0002 to $0.004 a run, against about $0.16 for a six-step agent task on Claude Sonnet 5.5.
Only if a lot of your requests come in by phone. A chatbot can't pick up a call, so count your calls and chats for a month and put the AI where the volume is.
Yes, and most good ones do. A chatbot or voice agent takes the request, automation handles the fixed steps, and an agent handles the cases that need judgment.
With AI automation for the task you repeat most. It's the cheapest to run and the easiest to check, and it shows you which exceptions are worth an agent later.
Sources
- OpenAI, A practical guide to building agents (opens in a new tab), accessed November 17, 2026.
- Anthropic, Building effective AI agents (opens in a new tab), accessed November 17, 2026.
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (opens in a new tab), June 25, 2025, accessed November 17, 2026.

