AI or Rule-Based Automation? Which Fits Which Task

Key Takeaways

• Classic automation follows fixed rules and gives the same result for the same input, every time.
• You need AI where language has to be understood or written, or where inputs vary too much for fixed rules.
• Most good solutions combine both: rules for the reliable steps, AI for understanding, a person for decisions.
• Calculations, deadlines and prices belong in fixed rules, not in a language model.
• Start with the process, not the tool: once every step is clear, choose a rule, AI or a person for each.

Classic, rule-based automation fits when a task always follows the same rules and the data arrives in a fixed form. AI fits when language has to be read, understood or written, or when the inputs vary too much for fixed rules. The best solutions combine the two: rules for the reliable steps, AI for the step that needs understanding, and a person for the decisions.

What is classic automation?

Classic automation works on the principle of “if this, then that”. Someone sets the rules and the system carries them out: when a form comes in, a contact is created in the CRM. When an invoice has been open for 30 days, a reminder goes out. The tools for this include workflow tools such as Zapier, Make or n8n, macros, features inside your industry or accounting software and, for older systems without an interface, robotic process automation (RPA), where a program imitates clicks and keystrokes. Some of these rules may already be built into your software, just not switched on.

Its great strength is reliability: same input, same result, every time. You can test it, audit it and explain it, and running a rule costs next to nothing. When something goes wrong, you can see exactly which rule fired. The weakness shows as soon as something departs from the pattern. If the format changes, a field is missing or a customer writes in their own words, the rule gets stuck.

What can AI do that fixed rules can't?

AI understands input that has no fixed form. A freely written email, a PDF with an unfamiliar layout, a photo of damage or a voice message: a language model recognises what it's about, pulls out the important details, categorises, summarises and writes drafts. Doing the same with rules would take endless lists of special cases. A pattern that works well is to have the AI return its result in a fixed structure, such as fields for name, address and type of job, so the next rule-based step can pick it up cleanly. The price is that AI works with probabilities. It's usually right but not always, and the same input can give slightly different results. So every AI step needs an answer to the question of what happens when it gets something wrong.

How do you decide which fits?

Go through your process one step at a time and ask these questions about each step:

• Does the input always arrive in the same form, such as a form field or a spreadsheet row? Then a rule is usually enough.
• Is there one clearly correct result that you could write down as a rule? Then the step belongs in a rule.
• Does text have to be understood, categorised or written? Then AI comes into play.
• What does a mistake cost? The more expensive it is, the more checking an AI step needs, or the step stays with a person.
• Do you have to be able to justify every decision in full? Then rely on rules, or let the AI only make suggestions.
• How often does the step happen, and how different are the cases? High volume with little variation points to rules, a lot of variation to AI.

If you can write it down as a rule, make it a rule. If someone has to read and understand it, that's a job for AI.

What does a combination look like?

An example: a painting and decorating firm receives an inquiry by email with three photos attached. Here's how rules, AI and people share the work:

1. Rule: capture the inquiry.The email lands in the inquiries inbox and starts the process. The photos are saved in the right customer folder.
2. AI: understand the request.It reads the text and the photos, picks out the room, the surfaces, what the customer wants and when, and flags what's missing, such as the ceiling height.
3. Rule: create the job and calculate.The details become a job in your software, and the line items are priced from your price list.
4. AI: draft the reply.It drafts a friendly reply asking for the missing measurements, and a first quote in your template.
5. Person: check and send.Someone at the firm checks the draft and the quote, adds what you can only see on site, and sends it.

The calculating is done by fixed rules, the understanding and wording by AI, the deciding by you. If the AI misses something, the process doesn't fall over: the case simply lands with a person. For how this works for quotes in detail, see faster quotes in the trades.

Which mistakes should you avoid?

• Using AI for things a rule handles reliably: totals, VAT, deadlines, date calculations.
• Building rigid rules for free text, such as a hundred keyword filters for sorting email that fail with every new phrasing.
• Having no fallback: when the AI is unsure or a detail is missing, the case has to go to a person.
• Automating an unclear process. What's chaotic today just becomes chaotic faster. Sort out the process first, as described in map the workflow before you automate it.
• Nobody being in charge. Every automation needs an owner who notices when something stops working.

What do running and maintenance cost?

Both have ongoing costs, just in different places. A rule costs almost nothing to run but needs adjusting whenever a form, a field or a connected system changes. Once you have many automations, you need an overview of what runs where and who looks after it.

An AI step costs something every time it runs, because model providers bill by the amount of text processed. Then there's oversight: spot checks, a list of errors and a fresh test when the provider updates the model, since the same instruction can produce slightly different results afterwards. Include these costs when you compare options. Our guide to building a business case for AI shows how to set them out clearly.

Where should you start?

1. Write down the process.Take one process that comes up often and record it step by step, with the systems involved, the handovers and the waiting times.
2. Label every step.Mark each step as a rule, AI or a person.
3. Rules first.Automate the rule-based steps. That's often quick and immediately brings some calm to the process.
4. Add one AI step.Add one step that needs language understanding, such as reading the inquiries, and have the results checked at first.
5. Measure.Compare throughput time and errors with how things were before. Also count the cases that ended up with a person, and why. They show you where the next improvement lies.

A workflow audit helps with taking stock, and our automation and systems integration service with building it. If you'd rather go through your processes with someone, get in touch.