What Is an AI Agent? Agent vs Chatbot vs Workflow, Separated
Three things get called AI agents, and only one of them is.
A chatbot answers. You ask, it responds, the exchange ends. Whatever happens next is your job.
A workflow executes a path you drew in advance. A trigger fires, steps run in the order you connected them, and the sequence is fixed before anything starts.
An agent decides its own next step, in a loop, until a goal is met. You give it an outcome rather than a procedure, and it chooses what to do — then looks at the result and chooses again.
The distinguishing property is who chooses the sequence. In a chatbot, you do, one turn at a time. In a workflow, you did, in advance. In an agent, the software does, at runtime.
The same task, three ways
Take a real job: find every unpaid invoice and chase it.
As a chatbot
You paste in your overdue list and ask for a chaser email. It writes a good one. You copy it, open your mail client, adjust the name, send it. Then you do that eleven more times.
The model did the writing; you did the work. Nothing happens unless you’re sitting there, and the quality of the outcome depends entirely on how many times you’re willing to repeat yourself.
As a workflow
You build it in something like n8n: every morning, query the accounting system for invoices past due, filter to those with no payment recorded, look up the contact, send a templated email, log it.
This is genuinely good. It runs without you, it’s reliable, and it’s auditable — you can see exactly which steps ran.
Its limitation is that it can only do what you drew. Reality deviates constantly:
- An invoice is overdue but the client already emailed to say payment is scheduled for Friday
- Two invoices belong to the same client and should be one email, not two
- One is overdue because you invoiced the wrong entity
- A client replies asking for a breakdown
A workflow handles none of those unless you anticipated each one and built a branch for it. When reality goes off-script, the workflow keeps executing the script.
As an agent
You give it the goal — chase unpaid invoices — plus access to the accounting system, the mailbox and a note of your policy on tone and escalation.
It queries the invoices. It notices two belong to one client and writes a single email. It reads the mailbox, sees the note about Friday, and skips that one. It drafts a breakdown when someone asks for it. When it finds the invoice with the wrong entity, it flags it to you rather than chasing.
Nobody wrote a branch for any of that. It chose each step by looking at what it found.
That’s the capability, and the honest cost of it sits in the same sentence: it decided all of that itself, and you didn’t pre-approve any of it. An agent that can compose a helpful email can compose an inappropriate one, to the wrong person, without a branch anywhere for you to inspect afterwards. Workflows are predictable because they’re rigid; agents are useful because they’re not.
Which one you actually want
| Your situation | Build |
|---|---|
| You need help thinking or drafting | Chatbot |
| The process is stable and you can describe every step | Workflow |
| The process is stable but high-volume | Workflow — cheaper and auditable |
| Reality deviates constantly and judgement is needed | Agent |
| A mistake is expensive or irreversible | Workflow, or an agent with approval gates |
The mistake people make is reaching for an agent because it sounds more advanced. If you can draw the flowchart, build the flowchart — it’s cheaper to run, easier to debug, and it does exactly the same thing every time, which is usually what you want from a business process.
Agents earn their cost when the flowchart would need fifty branches, and you’d still miss cases.
Why only the agent needs a server
Here’s the practical consequence, and it’s the part most definitions leave out.
A chatbot needs nothing from you. It runs in someone else’s application; you visit a website.
A workflow needs somewhere to run, but only when triggered. It can live on a hosted platform and wake on a schedule.
An agent needs a machine that stays on — and not the laptop you’re reading this on — for three specific reasons:
- It runs continuously. The loop has to be alive to receive a message and act on it. An agent that only exists while you have a tab open isn’t an agent.
- It holds state. Conversation history, what it already tried, credentials for the tools it uses. That has to persist between runs.
- It executes tools. Reading files, running commands, driving a browser. Those need an environment — and one you’re willing to give real access to.
That’s why “AI agent” turns into a hosting question so quickly, and why the answer is usually a small server rather than anything exotic. The model runs on the provider’s hardware via an API; your machine runs the loop. 4GB of RAM is the floor, 8GB once browser automation is involved, and no GPU is needed unless you’re running a model locally too.
If you’ve got as far as wanting to run one, where an agent should live compares a VPS against hardware at home, and what is a VPS covers the server itself.
The security shape is different too
Worth stating because it follows directly from the definition rather than being a separate topic.
A workflow does what you drew. An agent does what it decides — and it decides partly based on text it reads, which may include text written by someone else. That’s a genuinely different threat model: a message can be an instruction.
The practical consequence is that agent security is about limiting what it can do rather than about what it’s told. Its own documentation puts the model last in the defence order: assume it can be manipulated, and design so manipulation has limited blast radius. We cover that properly in securing a VPS that runs an AI agent.
How we checked this
The definitions here are conceptual rather than sourced from a vendor, and deliberately so — “agent” is a contested term and every vendor defines it to include their product. The distinguishing property we’ve used, that the agent selects its own sequence at runtime rather than following one specified in advance, is the one that survives across implementations and produces the infrastructure consequences described above.
The hosting figures — 4GB for a text-only agent, 8GB with browser automation, no GPU for API-backed models — are the published floors we use across our AI hosting coverage. The security ordering is from agent software documentation covered in our hardening guide.
What we did not do: we did not build the invoice-chasing example in all three forms. It’s an illustration of the structural difference, not a case study, and the failure modes listed for the workflow version are the ordinary ones anyone who has automated a business process will recognise rather than measured outcomes.
This article has no affiliate links.
FAQ
What is an AI agent?
Software that pursues a goal by choosing its own next step in a loop — deciding, acting, observing the result, and deciding again — rather than following a sequence specified in advance. The defining property is that it selects the sequence at runtime.
What’s the difference between an AI agent and a chatbot?
A chatbot responds to a prompt and stops; you carry out any resulting work. An agent acts on the world through tools and continues until its goal is met, without you driving each step.
Is workflow automation the same as an AI agent?
No. A workflow runs a path you drew before it started and can only handle situations you anticipated. An agent decides its path as it goes, which is why it copes with cases you didn’t foresee — and why its behaviour is less predictable.
When should I use a workflow instead of an agent?
Whenever you can draw the flowchart. Workflows are cheaper, faster, auditable and do the same thing every time. Reach for an agent when the flowchart would need dozens of branches and would still miss cases.
Why does an AI agent need a server?
Because it runs continuously, holds state between actions, and executes tools. A chatbot needs none of those; a workflow needs a runtime only when triggered. An agent has to stay alive to be an agent.
Do I need a GPU for an AI agent?
Not unless you run the model locally. The agent orchestrates on your machine while the model runs on the provider’s hardware via an API, which is why a modest server is enough.
Are AI agents less safe than workflows?
They have a different risk shape. A workflow does what you drew; an agent decides based partly on text it reads, so a message can function as an instruction. The mitigation is limiting what the agent is able to do, rather than trying to make it un-persuadable.
