Rule-based automation or AI? How to choose for each step of a workflow

Almost every automation conversation now reaches the same question early: should this use AI? The honest answer is usually that one step of it might. Most of the work in a business workflow is still best done by plain rules. The rest needs something that can read.

This post explains the difference in plain terms, gives you a table to sort your own tasks, walks through a workflow that uses both, and lists the limits we put on any AI step before it goes near real data.

Two kinds of work inside one workflow

Rule-based automation is the classic kind. When a form is submitted, add a row. When an invoice is marked paid, send a receipt. When stock drops below a set level, email the supplier. Every step is written in advance and the system does exactly that, every time.

Rules are cheap to run, fast and easy to check. If something goes wrong, you can open the rule and see why. Their weak spot is input. A rule needs the information in a known place and a known format. Give it a messy email and it has nothing to hold on to.

An AI step does a different job. It reads text, speech or an image that arrives in no fixed shape and turns it into something structured: the order number, the product, the date, the type of request. It can also sort messages and draft replies. What it should not do on its own is decide anything that costs money or cannot be undone.

Where rules are enough

If the information already sits in fields, a rule beats AI on cost, speed and reliability. Typical cases:

• Copying data between two systems that both keep structured records, such as an online store and an accounting tool
• Sending reminders on fixed dates, like the day before an appointment or 7 days after an unpaid invoice
• Moving a record to the next stage when its status changes
• Building a weekly report from numbers that already live in a database
• Checking a value against a limit, for example sending orders above a set amount for approval

Adding AI here buys nothing. It adds a cost on every run and a small chance of a wrong reading where there was none before.

Where an AI step earns its place

AI is worth it where a person is currently reading something and typing it again somewhere else. Three signs point that way.

• The input is free text or speech. Emails, WhatsApp messages, phone calls, notes typed in a hurry.
• The same information arrives in a different shape each time. Ten suppliers send ten invoice layouts. A rule would need ten versions and would break on the eleventh.
• Something has to be sorted before anyone can act. Is this message a complaint, a quote request or spam? Which team should see it? Is it urgent?

A fourth case is answering questions from your own information, such as opening hours, published prices or return terms. That works when the assistant is limited to your documents, which we cover in how AI answers only from your knowledge base.

A quick table for sorting your own tasks

Write down the tasks you want automated, then place each one in this table. The input column usually decides it.

The last row matters. Some steps should stay with a person however good the tools get. Automation still saves time there by gathering everything onto one screen.

If you have several candidates and cannot pick one, which workflow to automate first gives a simple scoring method.

The mixed workflow: AI reads, rules decide

Most useful systems combine both. Here is how a quote request email might move through one.

1. An email arrives. A rule picks it up because it came to the sales address.
2. The AI step reads it and pulls out company name, product, quantity, deadline and contact details. It also labels the request: new quote, follow-up or something else.
3. Rules check the result. Are all required fields filled? Is the quantity a number? If anything is missing or unclear, the email goes to a person with a note saying what is missing.
4. If everything is in order, rules create the record in your sales tool and assign it to the right person.
5. A rule sends the sender a short confirmation that the request arrived.
6. Pricing the quote stays with a person. The system has only done the reading and the filing.

Notice where the AI sits. It is one step, boxed in by rules before and after. If it misreads something, the next rule catches the gap instead of passing it on.

Limits to put on any AI step

AI output can be wrong in ways a rule never is. So the step needs fences. These are the ones we use, and they are worth asking for from anyone who builds a system for you.

• It reads and sorts, it does not commit. No payments, refunds, deletions or contract changes triggered by AI alone.
• A clear hand-off to a person. When the reading is unclear, a field is missing or the topic is sensitive, the job goes to a person. In our packages the written scope states when that happens.
• Answers only from your own information. If the answer is not in your documents, the assistant says so instead of guessing.
• Every step is logged. You can see what came in, what the AI read and what the rules did with it.
• A person reviews high-stakes output. Anything legal, medical or financial gets checked before anyone relies on it.

How to test it before you trust it

You do not need technical knowledge to check an AI step. You need real examples and some patience.

1. Collect 30 to 50 real inputs from the last few months, including the messy and odd ones.
2. Write down by hand what the correct reading of each one should be.
3. Run them through the step and compare. Count the misses and look at what they have in common.
4. Check that every unclear case went to a person instead of being guessed.
5. Repeat after any change to the instructions, the documents or the tools around it.

If the misses cluster around one kind of input, that kind may belong with a person or with a stricter rule. That is a normal outcome, and it is cheaper to find before go-live than after.