AI agent or a simple workflow? Decide before you hire

The short answer

Ask one question about each step of the task: could you write the rule down? If every step is “when this happens, do that”, you need a workflow automation. If one step needs judgment, such as reading a messy email, picking the right answer or writing a reply, you need an AI step at that point. You need an agent, which decides its own next steps, only when you cannot say in advance which steps the job will take.

Most small-business tasks land in the middle: a fixed workflow with one or two AI steps inside it. That is cheaper to build, easier to test and easier to hand over than an agent that runs the whole job. Decide which of the three you are buying before you write the brief, because it changes the price, the test and who you hire.

What the words mean when you are buying

Gig titles use “agent”, “AI automation” and “chatbot” loosely. For a buyer, three shapes matter:

Shape	Who decides the steps	Good for	Main risk
Workflow	You, in advance, as rules	Moving data between tools, reports, reminders, publishing on a schedule	Breaks when a connected tool changes
Workflow with an AI step	You decide the steps; the AI does one judgment step	Sorting emails, summarising, drafting, extracting fields from free text	The AI step’s output is wrong or in the wrong shape
Agent	The AI chooses tools and steps toward a goal	Open-ended research, multi-step support cases, tasks with many branches	Does something you did not expect, or does not know when to stop

The chatbot and AI agent job page draws the same line from the other side: a chatbot answers, and an agent also acts.

A one-question test for each step

Write the task out as it runs today, step by step, as what to automate first suggests. Then mark each step:

• R (rule): the same input always gets the same action. “New form entry → add a row to the sheet.”
• J (judgment): a person reads something and decides. “Is this enquiry a sales lead or a complaint?”
• O (open): the person decides what to do next and in what order, depending on what they find.

All R: hire for a workflow. Some J: a workflow with an AI step at each J. Several O steps that cannot be made into J steps with a narrower question: that is where an agent earns its cost.

How I split it in my own jobs

My own unattended setup uses AI in a few places, and the pattern is the same everywhere: the AI does the judgment step; fixed rules do everything around it.

• A nightly drafting job in my content pipeline. An AI writes the drafts. That is the only step where judgment is needed. Everything after it is fixed: a script checks each draft against hard rules (word count, every internal link resolving, the draft kept unpublished, no forbidden names). Then a second, read-only AI checks that numbers and first-person claims trace back to files. Finally a script assigns publishing dates in order. The publishing job at 00 is plain rules: build, audit, push, and undo the change if the build fails.
• The comment-reply job. It checks for new comments every 20 minutes and replies when a comment matches a keyword. No AI at all. It needed no judgment, so it has none, and when it went wrong the cause was a setting written in plain sight. (It looked at too few recent posts; the webhook explainer tells that story.)
• The social content job. An AI writes the content of each item overnight. A fixed renderer turns it into images and refuses to produce a broken card: text that overflows, or a setting it does not recognise, stops the render with an error instead of quietly falling back.

That last point is the one to copy. The step after an AI step should check the AI’s output, not trust it. The day the content format changed and the fixed steps were not updated with it, nothing got published for a stretch. That was a wiring problem, not an AI problem, but it only surfaced because the fixed steps rejected what did not fit.

The costs are different too. One overnight AI job hit a monthly spending limit on the account it ran on. It produced its output and ended with an error. A rules-only workflow has no such line on the bill.

What changes when you choose an agent

If your test says agent, or AI steps, expect these differences in the quote and in the brief:

1. A test set, not a demo. Twenty to thirty real cases with the answer you expect, agreed before the build. The chatbot job brief asks for exactly this.
2. A never-do list and a way out. What it must refuse, and who takes over when it is unsure.
3. Limits on actions. Which tools it may use, what it may change, and a maximum number of steps or retries per run.
4. A readable record of decisions. A log line per run showing what it decided and why, so a wrong result can be traced.
5. A usage estimate. The AI provider bills per use, on your account. Ask for the expected monthly usage at your volume.
6. A check after the AI step. Whatever consumes the AI’s output should validate its shape and stop loudly when it is wrong.

For a plain workflow, most of that list shrinks to the acceptance checklist for unattended automations: scheduled runs, a forced failure, a duplicate test and credentials on your accounts.

Who to hire for which

• All rules: a workflow automation freelancer. The tool choice (n8n, Make, Zapier) matters less than the brief; n8n vs Make vs code covers that decision.
• A workflow with AI steps: either job fits. Look for sellers who mention AI steps inside workflows, such as Abdullah Ramzan, whose gig focuses on AI agents inside n8n workflows.
• A real agent that answers and acts for customers: the chatbot and agent job, with the stricter brief and acceptance.

A brief paragraph you can copy

Shape. Steps 1–3 and 5–6 are fixed rules (listed below). Step 4 needs judgment: [describe it]. Please use an AI step only there, with its output checked before the next step runs. If you think an agent is needed instead, tell me which step cannot be written as a rule and why. Limits. Maximum [N] AI calls per run; any output that fails the check stops the run and alerts me. Test. Acceptance on [20] real cases for the AI step, plus three unattended scheduled runs.

The last sentence of the first paragraph is the useful one. A good seller will either agree or explain, in plain terms, which step really needs an agent. Either answer tells you what you are paying for.

Wrap-up

An agent is not a better workflow. It is a different trade: flexibility in exchange for predictability. Mark each step as rule, judgment or open. Put AI only where judgment is needed, and check its output before anything acts on it. Most tasks are then a workflow with one smart step. That is quicker to build and cheaper to run, and you can test it before you pay. The pricing guide shows how AI steps move the quote.