Most business automation needs a script or a workflow, not an AI agent. Use a workflow when the input is stable, the steps are known and a failure is cheap. Add one AI step when a single step needs judgement on messy input, such as reading an order written as free text. Build an agent only when the path itself cannot be known in advance and stopping costs more than the agent does. One question sorts almost every case: if the input changed shape tomorrow, what should happen?
What is the difference between a script, a workflow and an AI agent?
"Automation" hides three different animals. A script executes a fixed sequence: fetch, transform, write, done. A workflow is the same idea drawn as boxes in a tool such as n8n, with triggers, branches and retries that someone who does not program can read and change. An AI agent evaluates: it looks at what came back, decides the next step, retries differently and asks a person when it is stuck. The first two cost less to build and to run. The agent survives change. Between them sits the most useful hybrid of all: a fixed workflow with one AI step inside it.
When is a script or a workflow the right answer?
If the input format is stable, the failure is harmless, and a person checks the output anyway, write a script or build a workflow. A nightly export, an invoice renamer, a report that goes out every Monday. I have written hundreds. Paying agent prices for script problems is how AI budgets die, and the four cases where an agent is the wrong answer are in when not to build an AI agent.
When does a workflow need one AI step?
When one step needs reading or judgement and the rest does not. Supplier emails that write their orders as free text, messages that must be sorted by topic, invoices that arrive as PDFs. Keep the path fixed and give the model a single job: read this, return these fields. Code then checks the fields, and anything the model is unsure about goes to a person. The path stays predictable and testable, and the AI handles only the messy part. Returning fields in a fixed format is covered in structured outputs, sorting email in AI text classification and reading invoices in AI document processing.
When does only an agent survive?
The moment the input is messy and the path itself varies, or the failure is expensive, the calculus flips. Reconciling data that three systems disagree on. Chasing an order across a supplier portal, an email thread and the ERP. Watching a process where "it silently did the wrong thing for three weeks" costs more than the build. A script meets the unexpected and either dies loudly or, worse, carries on quietly wrong. An agent notices, retries with a different approach, and calls a person when it should. What an agent is, step by step, is in what is agentic AI.
What does each option cost?
A script or a workflow on your own server costs close to nothing per run, and n8n can be hosted on your own server (n8n, hosting). An AI step adds one model call per item: by my arithmetic at September 2026 prices, sorting a 600-token email into one category costs about 0.65 US dollars per thousand emails on Claude Haiku 4.5 (Anthropic, pricing). The same email written in Greek needs roughly twice the tokens, and so costs roughly twice as much, as measured in Greek in AI models. An agent makes many calls per task, repeats some of them, and needs tests and monitoring before anyone can trust it, which is where most of its cost sits, as broken down in what an AI agent costs.
What is the test that decides?
One question sorts almost every case: if the input changed shape tomorrow, what should happen? If the honest answer is "it should stop and tell someone", a script or a workflow with good alerts is enough. If it is "it should read the new shape and carry on", add one AI step and check what it returns. If it is "it should figure it out and keep going, because stopping costs us money every hour", that is an agent, and it needs state, evals and cost ceilings, which is engineering, not prompts. How to prove an agent works before trusting it is in LLM evals.
Where does n8n fit?
n8n is a workflow tool I use for the first two cases. Its AI nodes put a model call inside a fixed flow, and it offers an agent node as well (n8n, advanced AI). For a workflow with one AI step it is quick to build and easy for a team to read. When the job is a real agent that must keep state, recover from failures and be tested release by release, I write the orchestration in code, with LangGraph or my own, because tests, versions and reviews are simpler there. The tool matters less than the test above.
Frequently asked questions
Is an AI agent better than a workflow?
No, it is different. An agent handles situations nobody listed in advance, and pays for that in cost, speed and the testing it needs. A workflow does exactly what it was drawn to do. For most business processes, the second is what you want.
Can I start with a workflow and add an agent later?
Yes, and it is usually the best path. Run the workflow, log where it fails, and hand only the failing step to a model or an agent. Your own logs then tell you what the AI has to handle and how often.
Do AI automations work in Greek?
Yes, with care. Test on real Greek input, including messages typed without accents and in Greeklish, and budget for roughly twice the tokens of English. The search side of the same problem is in RAG in Greek.
I build all three, and I will happily tell you your problem is a script: it is cheaper for you and faster for me. When it is an agent, it is the work described under AI agents for business. Describe the process and you will get a straight answer on one call.