An AI automation runs a fixed set of steps you defined, and a model does the thinking in one of them. An AI agent receives a goal and works out the steps itself, every time it runs. Plenty of what gets sold as an agent today is a stack of automations underneath, which is the thing worth checking before you pay for either one.
What is an AI automation?
Automations have run quietly behind your business for years. Something happens, software does the next thing, nobody watches it. The AI version changes one thing: one task in the chain now has a bit of intelligence in it.
A subcontractor invoice arrives. Step one, the file lands. Step two, a model reads it and pulls out the amount. Step three, the result is saved to your database. Defined steps, in a fixed order, finishing where you said they finish.
What is an AI agent, and how is it different?
An agent receives a goal rather than a first step. It then works out which steps get it to that goal. Today it might choose three. Tomorrow five, depending on the inputs and the outputs it runs into. After each step it settles on the next one and carries on down that path.
The thinking runs the whole way through. In an automation it sits in one place and the chain carries on without it.
What is a node in an AI workflow?
A node is one step in a workflow: fetch the data, classify it, write it away. Each node does one piece of the puzzle and hands the result to the next one.
The classifying node can be an AI node. Same model, same intelligence, same logic, same reasoning you would get from it anywhere else. Nothing about the model gets weaker for sitting inside a workflow. It gets asked one question, in one place.
An agent is responsible for every node instead of one, the simple ones where it fetches data and passes it along as well as the ones that need judgment. That is also where it goes wrong. The agent adds a node that did not matter, or misses one that did. Getting deterministic behaviour out of an agent is hard, which is why somebody on your side has to understand when an agent needs particular information and what it is good and bad at.
Is that AI agent actually just a set of automations?
A lot of what is being sold as an agent right now is a bunch of automations under the hood. There is a tell, and it has nothing to do with the product page. Ask who defined the logic.
With an automation, the human defines the logic of the entire thing, whatever the steps are. There are no shortcuts available. Every case has to be designed up front and then embedded in whatever software runs your automations.
With an agent, you define roughly the end outcome, plus good prompts on what the steps could be and how to approach different cases when they come up. You do not define all of the detail. The detail you leave out is the detail the agent reasons its way through.
So put the question to the vendor. If a person had to sit down and write every step and every branch before the thing ran, you are buying an automation, whatever the invoice calls it. If what got defined was the outcome and the prompts, it is an agent.
What are the trade-offs between an AI agent and an automation?
| Where they differ | AI automation | AI agent |
|---|---|---|
| Steps | Fixed, defined by you | Chosen by the agent on each run |
| Where the AI sits | One node | Every node |
| Who defines the logic | You define every branch up front | You define the outcome plus prompts |
| Cost to set up | Low, and cheap to run | Higher, and more to configure |
| Strength | Deterministic, easy to debug | Flexible, handles complex work and workflows that change often |
| Weakness | Rigid when a case can go many ways | Unpredictable and harder to debug |
| Best for | A process that runs the same way each time | A job that can go many ways |
Easy to debug is worth more than it sounds. When an automation breaks you open it and see which node caused the issue. An agent that misbehaves gives you a run that went somewhere you did not expect, and finding the point where it turned takes longer.
When should you use an AI agent instead of an automation?
Count the ways the job can go. A recurring subcontractor invoice that follows the same path every month goes one way, so it is automation work, and buying an agent for it means paying for unpredictability you had no use for.
A coding problem goes in an infinite number of directions, so you want an agent on that one. So does a tenant maintenance request. The tenant says the boiler is out. Or that it is out and there is a newborn in the flat. Or that the last contractor never showed and they want a different one. Each of those is a different job, and writing all of them down in advance is the thing you cannot do.
Three checks before you decide:
- Count the ways the job can go. One path every time points at an automation. Twenty points at an agent.
- Try to write every branch down today. If you can fill a page and know the page is complete, an automation will run it.
- Check how often the process changes. Workflows that change frequently are what agents are good at, and every change to an automation is another branch somebody has to design and embed.
An agent costs more to set up and configure, and it is harder to debug. Get it working right and the return can be substantially more than the automation would have given you. Getting it working right is the part that takes the expertise.
Questions people ask about AI agents and automations
Is an AI agent just an automation?
Sometimes, and that is worth checking before you pay. The distinction is whether a person defined every step up front or only the outcome and the prompts. A fixed chain of steps with a model sitting in one of them is an AI automation, however it is marketed.
Are AI agents more expensive than automations?
To set up, yes. An automation is cheap to run and an agent costs more to configure and more to debug. What justifies the spend is a job that goes too many ways for a fixed chain to cover.
Can you use an agent and an automation together?
Most systems already are both, since much of what is sold as one agent is several automations underneath. Split them by node. A step that runs the same way every time does not need an agent deciding anything about it. A step that can go many ways does.
What is an agentic workflow?
The thinking-loop version of a workflow. Rather than a fixed chain, the agent receives the goal and decides which nodes belong in this run, so it might use three today and five tomorrow. How much of that decision you hand over is agent autonomy, and when several agents each take a piece of one job, an orchestrator decides who does what.
