Do you need an AI agent, or just a workflow?

Most business tasks that people want to hand to an "AI agent" do not need an agent at all. They need a workflow: a fixed set of steps an AI runs through the same way every time. Agents earn their keep only when a task genuinely cannot be mapped out in advance, and most business tasks can be. So before you pay for the impressive-sounding option, here is the plain difference between the two and the one question that tells you which you actually need.

What is the difference between an AI agent and a workflow?

The words get thrown around loosely, so it helps to borrow the clearest definitions going. In its December 2024 guidance on building effective agents, Anthropic (the company behind Claude) draws the line like this. A workflow is a system where the AI and its tools are "orchestrated through predefined code paths". An agent is a system where the AI "dynamically directs its own processes and tool usage", deciding for itself how to get the job done.

In plainer terms: a workflow follows a set route you laid out. An agent works out the route as it goes. That difference sounds small and it changes everything about cost, reliability, and how much you can trust the thing.

Why a workflow is usually the right answer

A workflow does the same knowable thing every time, so it is cheaper to build, faster to run, and far easier to trust. You can look at it and know what it will do, because it does not improvise. For the bulk of what a small business wants automated, following the quote through the system, sorting the enquiry, chasing the unpaid invoice, that predictability is exactly what you want.

An agent, by contrast, is exploring and deciding on the fly. That flexibility is real and sometimes essential, but it costs more, runs slower, and can go wrong in ways that are harder to predict and harder to spot. Anthropic's own advice to builders is blunt about this: "find the simplest solution possible, and only increase complexity when needed", and sometimes that "means not building agentic systems at all". When the company that builds the models tells you not to reach for the fancy option by default, that is worth hearing.

The one test: can you draw the decision tree?

Here is the test we use at Handiwork, and it is the whole article in one line.

If you can draw the decision tree for a task on a whiteboard, every branch, every "if this then that", you need a workflow, not an agent. If the task genuinely cannot be drawn because it changes depending on what turns up, that is where an agent belongs.

Most business tasks can be drawn. Think about accepting a quote, onboarding a client, triaging an email. You already know the branches, because you do them the same way every week. That knowable-ness is the sign a workflow will handle it, more reliably and for less money than an agent would.

The tasks that resist the whiteboard are the ones where the right next move honestly depends on what you find. Open-ended research, or work that has to weigh messy trade-offs with no fixed rule, that is agent territory, and there it is genuinely worth it.

Is an agent ever the right call?

Yes, absolutely, and this is not an argument against agents. We build them at Handiwork, and when a task actually calls for one they are brilliant. The point is not "agents bad, workflows good". It is that reaching for an agent by default, because it is the word every vendor is using this year, usually gets you something more expensive and less reliable than the job needed.

The honest question is not "do you want an agent". It is "is this decision the same every time, or does it genuinely change depending on what happens". The answer tells you which tool fits, and nine times out of ten for a small business it points at a workflow. When it points at an agent, that is exactly when an agent is worth building.

What this means before you buy

If someone is selling you an "AI agent", ask them to draw the decision tree for the task first. If they can draw it, they should be building you a workflow, and it should cost less and break less often because of it. If they genuinely cannot draw it, an agent may be the right call, and now you both know why.

This connects to a wider point about building AI that lasts. The same discipline that favours a workflow over an agent also favours automation that adapts when your process changes rather than a brittle script, and it is a big reason good workflows can be built in weeks, not months. If you want to see the kinds of jobs this applies to, our use cases are the practical next read.

Ready to find out where you stand?

If you are not sure whether the thing you want to automate needs an agent or a workflow, the free AI Readiness Check at handiworkco.com.au/ai-readiness is a good place to start. It takes about five minutes and there is no pitch at the end of it.