Does this task need AI, or would a simple rule do?

Start with the decision the task actually makes. If you can write that decision as a small, stable set of conditions, a rule is often enough. If the task has to interpret varied language or messy documents, AI may help prepare an answer. If the consequence of getting it wrong matters, keep a person responsible for the final call.

These choices can be combined. A workflow might use rules to route a form, AI to draft a summary and a person to approve a customer-facing response.

Three ways to handle the same workflow

Imagine a small business receives enquiries through a form. The examples below are illustrative; they are not claims about an existing customer's system.

The rule is easier to test: submit one enquiry for each category and check its destination. The AI summary needs a different check: compare it with the original, especially for quantities, addresses, timing and anything that sounds certain. The final promise needs the person who can actually commit the business.

Use a rule when the inputs are clear

Rules work well for known fields and repeatable decisions: assign a lead by service type, flag a missing postcode or send a reminder after a defined period. They are usually cheaper to run, easier to explain and more predictable than a language model for the same fixed task.

The limitation is that a rule cannot reliably understand every exception hidden in free text. If you keep adding special cases until no one understands the workflow, step back and redesign the form or the process before adding another condition.

Use AI when language is the hard part

AI can help draft, classify or summarise information that people express in many ways. Good uses have a clear source to work from and a practical way to check the result. For example, an assistant could prepare a draft response based on an approved service description and a customer message. It should indicate uncertainty or hand off when the approved information does not answer the question.

Define the boundary before using it: which sources may it read, what may it produce, who reviews it and what happens when it cannot answer. Do not let a convincing paragraph stand in for a verified fact.

Keep a person where judgment and accountability sit

Price commitments, safety decisions, unusual customer complaints and access to sensitive records often need human review. Automation can collect the facts and present them clearly, but someone must own the decision and be able to explain it.

For invoice reminders, the boundary might be: a rule finds unpaid invoices, AI helps draft a courteous reminder, and a person checks the balance, recipient and wording before anything is sent. That is a proposed workflow, not a claim that every business needs AI to chase invoices. A fixed template may do the job if the messages do not need interpretation or varied wording.

A five-minute decision check

For the task you want to change, write down:

1. The input, including where it comes from and how reliable it is.
2. The action or decision the workflow produces.
3. The consequence of a wrong result.
4. How a person would notice and correct a mistake.
5. Whether a simpler form, rule or checklist would solve the problem.

If the answer is still unclear, try the process manually on a small set of real examples before building anything. Explore automation work when the task and its review point are clear enough to discuss.