Every owner who looks at automation now hears the same suggestion: use an AI agent. Sometimes that is right. Often it is not, and the cheaper option has existed for years. A rule that says "when an invoice arrives in this mailbox, copy it to this folder and notify accounting" does not need artificial intelligence. It needs ten minutes in a workflow tool.
This article helps you decide between the two before you spend money. It is written for owners and managers of small and mid-sized companies, not for engineers.
What the two things actually are
Plain automation follows rules you wrote. If this happens, do that. Tools such as Zapier, Make, n8n or the automation features inside your accounting or CRM software work this way. The behaviour is the same every time, and when something goes wrong you can read the rule and see why.
An AI agent is given a goal and some tools, and decides for itself which steps to take. It can read an unstructured email, work out what the customer wants and choose between several actions. That flexibility is the point, and also the cost: the same input can produce slightly different results, and the reasoning is harder to inspect.
The people who build these systems say the same thing. Anthropic's engineering guidance recommends finding the simplest solution possible and adding complexity only when it clearly improves the result, noting that agentic systems trade speed and cost for flexibility (Building effective agents). For a business buyer, this means a simple workflow is a perfectly respectable answer.
Five questions to ask about any task
1. Can you write the steps down?
Take the task and try to describe it as a list: first this, then that, and if X, do Y. If you can finish the list in a page and the exceptions are few, you have a rule-based process. If you keep writing "it depends" or "it needs a person to look at it", you have found the part where judgment is involved. Writing the process down is useful either way, and describing a process before automating it shows how.
2. Is the input tidy or messy?
Rules work well on structured input: a form with fixed fields, a spreadsheet row, an order from your web shop. They struggle with free text, scanned documents in many layouts, or requests written in different languages and styles. Messy input is where an agent starts to pay off.
3. How costly is a wrong result?
A rule makes the same mistake every time, which is easy to find and fix. An agent makes occasional, varied mistakes. If an error means a wrong invoice or a wrong promise to a customer, you want predictability, or a person checking the agent's work. If an error means a slightly clumsy draft that someone reads before sending, an agent is fine.
4. How many times a day does it happen?
Running a rule costs almost nothing. Each agent task uses model capacity that you pay for, usually per amount of text processed. For a task done thousands of times a day, that adds up. For a task done ten times a week, the cost is trivial and the main question is whether the setup is worth the time. Do the arithmetic for your own volume, as described in a 90-day view of automation.
5. Will you need to explain what happened?
Customers, auditors and your own staff sometimes ask why something was done. With a rule, the answer is the rule. With an agent, you need records of what it saw and did, and someone who reads them. If you cannot spare that effort, prefer rules.
A simple way to read the answers
| Your task looks like this | A better fit |
|---|---|
| Fixed steps, structured input, few exceptions | Plain automation |
| Fixed steps, but the input arrives as free text | Agent reads the input, rules do the rest |
| Many exceptions, each needing judgment | Agent drafts, a person decides |
| Rare task, high stakes | Neither: keep a person and a checklist |
The second and third rows are where most small businesses end up. The middle path is common because most real work is partly routine and partly untidy.
Examples from everyday operations
Invoice handling. Invoices arrive by email as PDFs from dozens of suppliers, each laid out differently. Reading the amount, due date and supplier from such a document is a good job for an agent. Once the data is extracted, checking it against the order, posting it and scheduling payment are rules. Using an agent for the reading only keeps the risky steps predictable.
Appointment reminders. A message the day before every booking is pure rules. Putting an agent here adds cost and risk and no benefit.
Customer enquiries. A shared inbox receives questions of every kind. Sorting them into sales, support and billing, and drafting a first reply, suits an agent. Sending the reply without a person looking at it is a separate decision, and for most small firms the answer is "not yet". A staged approach is described in rolling out an agent in shadow mode.
Weekly sales report. Pulling the same figures from the same system into the same table is rules. Writing two sentences of commentary on what changed may be an agent's job, with a person reading it.
Signs you chose the wrong tool
You may have bought an agent when you needed rules if:
- the agent does the same thing nearly every time and you rarely correct it;
- the monthly bill is higher than the time it saves;
- you cannot explain to a colleague why it made a particular decision;
- you keep adding instructions to make it behave like a fixed procedure.
You may need more than rules if:
- the rule list grows every week and still misses cases;
- people spend their time handling the exceptions the rules could not;
- the input changes format often and the workflow breaks each time.
Neither situation is a failure. It is information. The review in looking at a pilot and deciding whether to expand, adjust or retire it covers how to act on it.
How to proceed
- Choose one task and write its steps and exceptions on a single page.
- Mark every step that truly needs judgment about unstructured input. Everything else can be a rule.
- Build the rule-based part first. It is cheap and gives you a baseline of time and errors.
- Add an agent only for the marked steps, with a person reviewing at the start.
- After a month, compare cost per task, error rate and time saved against the baseline.
This order has an advantage: if the agent turns out to be unnecessary, you have lost little, and if it is needed, you already know exactly where.
Conclusion
An AI agent is a tool for untidy work, not a replacement for every automation. If you can write the process as rules, start there. If the input is messy and judgment is needed, let an agent handle that step and keep the rest predictable. The choice that fits your task is the one you can afford, check and explain.
