Skip to main content

Plain English

Chatbot, automation or AI agent: which one do you actually need?

Every vendor now sells 'AI agents', and much of what ships is a chatbot with a new name or an automation with an AI step bolted on. The three are genuinely different tools. They cost different amounts, they break in different ways, and choosing the wrong one is the most common way we see small businesses waste money on AI.

Rather just ask someone? Talk to us

In short

A chatbot talks. An automation follows fixed steps when something happens. An agent is given a goal, tools and permission to decide the steps itself. Most small businesses need automation with an AI step for reading or drafting. Agents earn their keep only when the work needs judgement every time, and they need guardrails to match.

What is an AI agent, without the marketing?
Software that's handed a goal ('deal with this enquiry'), a set of tools it may use (read the inbox, check the calendar, draft a reply, create a record), and the freedom to work out which steps to take. That freedom is the point and the risk. A fixed automation does the same thing every time and fails loudly when the input is unexpected. An agent adapts to the unexpected, which means it can also adapt in a direction you didn't want. Good agent design is mostly deciding what it may never do.

What actually matters

  1. Start from the shape of the work

    If the task has the same steps every time, it's an automation. If it's a conversation with a customer, it's a chatbot, and probably not needed. If it's 'look at this, work out what it is, and do the right thing', it's an agent. Most 'we need an agent' requests are automations with one messy step, and the messy step is where a small AI call does the reading.

  2. Automations are cheaper because they can't surprise you

    A workflow that moves a form submission into your CRM and drafts a reply costs a fraction of an agent and behaves identically at 3am on a public holiday. Predictability is a feature. Reach for it first, and only give the software discretion where the task genuinely can't be reduced to steps.

  3. Chatbots are customer-facing, so their mistakes are public

    A bot on your website that gets a price wrong has told a customer something false in your name. That's why we talk most clients out of them. The exceptions are real: a large catalogue, a help centre people actually search, or a volume of repeat questions a human can't keep up with.

  4. Agents need a fence, not just a brief

    The design work on an agent is the list of things it may not do: never send to a customer without approval, never change a price, never touch a record outside its scope, always log what it did and why. An agent without that list isn't autonomous, it's unsupervised. Ask any vendor for theirs.

  5. Judge by the exceptions, not the demo

    Every demo shows the happy path. Ask what happens when the email is in the wrong language, the attachment is a photo of an invoice, or the customer changes their mind halfway. Automations should stop and flag. Agents should escalate. If the answer is 'it just handles it', keep asking.

  6. Expect to combine them

    A realistic build for a small business is one automation that catches the trigger, an AI step that reads the messy part, a small agent that decides between three outcomes, and a person who approves anything that leaves the building. Naming it one thing is marketing. Building it as four small things is engineering.

Side by side

The differences that decide cost and risk.

CriteriaChatbotAutomationAI agent
What it doesAnswers questions in a conversationRuns fixed steps when something happensWorks out the steps to reach a goal
Handles messy inputSomewhatPoorly, unless an AI step reads itYes, that's the point
Makes decisionsNoOnly branches you definedYes, within limits you set
Who sees its mistakesYour customersUsually your teamDepends entirely on the guardrails
Typical first-project costLow to midLowestHighest
Runs unattendedYesYesOnly inside the fence
Right forHigh-volume repeat questionsAnything with a patternWork that needs judgement every time

When this isn't the answer

If you're reading this to decide whether to buy an agent, the answer is probably an automation. We build agents, and we still start almost every client on the simpler thing, because it proves the value in weeks and costs less to run. Come back for the agent once the automation has shown you where the judgement calls actually are.

Start with a 30-minute call

Tell us the task that eats the most time, or just that you're not sure where to begin. We'll tell you on the call whether it's worth automating, and what we'd do first.

Email

jayson@pixelapps.com.au

Location

Macedon Ranges, Victoria

Serving clients across Australia

Only if a quick call would be easier than email.

What are you trying to do?(optional)
Rough budget(optional)

A range is enough. It helps us recommend the right approach, not the biggest one.

We reply within 24 hours. No obligation, and we never share your details.

Common questions

On its own, no. It's a model you talk to. It becomes part of an agent when it's connected to tools, given a goal and allowed to act. The same model can sit inside a chatbot, an automation step or an agent; the wrapping decides which one you've got.

Yes, and this is the most useful combination for small businesses: a fixed workflow with one step that asks a model to read an email, extract the details or draft a reply. You get the predictability of automation with the reading ability of AI.

When the input varies too much to branch on, the right action depends on context, and the volume is high enough that a person triaging it is a real cost. Inbox triage, document processing across many formats, and multi-step research are the usual cases.

Most small business sites don't. A visible phone number and a form that works convert better. If you have hundreds of the same question a month, or a catalogue people struggle to search, a well-built one can pay off. We wrote a whole guide talking people out of it.

For automations, n8n and Make where they fit, custom code where they don't. For AI steps and agents, Claude and other current models, chosen per task. You own whatever we build and it runs on accounts you control.