AI or Automation? Picking the Right Tool for the Task

Two different tools

"Automate it with AI" has become the default suggestion for any repetitive task. Often it is the wrong one.

Conventional automation follows rules you write: when an invoice arrives, if the amount is over this threshold, send it for approval. It does exactly what it is told, every time.

AI — in practice, a language model — handles input that rules cannot describe: an email written in someone's own words, a scanned contract, a support message that might be a complaint or a compliment.

The useful question is not "can AI do this?" It nearly always can. The question is whether it should.

The test

Look at the input, not the task.

• Is the input structured and are the rules clear? Use conventional automation. It is cheaper, faster, and gives the same answer every time.
• Is the input unstructured — free text, documents, images, speech? This is where a model earns its cost.
• Does the task need judgement that even people disagree on? Keep a person in the loop, whatever tool you use.

How they compare

	Conventional automation	AI
Input	Structured: forms, fields, records	Unstructured: text, documents, images
Behaviour	Identical every time	Varies; needs checking
Cost per run	Negligible	Real, and grows with volume
Speed	Instant	Slower
When it is wrong	It breaks loudly	It can be confidently wrong
Changing it	Edit a rule	Adjust prompts and re-test

Tasks that look like AI but are not

• Moving data between two systems that both have APIs
• Sending reminders on a schedule
• Routing a form based on a dropdown the customer already filled in
• Generating a document from a template and a record

Each of these is a workflow. Adding a model makes them slower, more expensive and less predictable.

Tasks where AI is the right tool

• Pulling the supplier, date and total out of invoices that all look different
• Sorting incoming messages by what they are actually about
• Summarising a long thread so someone can act on it
• Answering questions from a body of documents — see RAG explained
• Drafting a first reply for a person to review

The pattern that works best: both

The most dependable systems use a model for the one step that needs it, and rules for everything else.

1. AI reads the messy input and turns it into structured data: supplier, amount, due date, confidence.
2. Rules decide what happens next, exactly as they would for data typed in by hand.
3. A person reviews anything the model was unsure about.

The model does the part only it can do. The parts that must be predictable — approvals, payments, anything with consequences — stay under rules you can read and audit.

Before you start

Ask three things:

1. How many times a month does this happen, and how long does it take a person?
2. What does a mistake cost?
3. Could a simpler tool do it?

If the volume is low, or a mistake is expensive and hard to notice, automation of any kind may not be worth it yet. If a rule will do, use a rule. And when the input really is messy, that is the moment to bring in a model — which is where most of our AI and automation projects begin.