A business problem needs machine learning when the right answer depends on patterns that are too many or too subtle to write down as rules, and when you have enough past examples to learn those patterns from. If you can describe the decision in a short list of if-then rules, write the rules. If the input is free text, speech or open questions, a large language model is often the faster route.
Most of the work is choosing the right one of these three. Here is how we decide.
When are plain rules enough?
Rules win when the logic is known, stable and needs to be explained. Examples: send a reminder 24 hours before an appointment, flag invoices over a set amount, route enquiries from one region to one person. Rules are cheap to run, easy to test and never surprise anyone. Reaching for AI here adds cost and risk for no gain.
When does classic machine learning fit?
Classic machine learning learns from labelled examples in structured data. It is a good fit when you have a clear question, a history of answers and a way to check results. Three typical cases:
- Classification. Is this email spam? Will this lead convert? A model trained on past emails or past leads learns which signals matter. Spam filtering is the textbook example, because thousands of labelled messages exist and the cost of a mistake is easy to measure.
- Recommendation. Which product or article should this person see next? Models learn from what similar people chose before.
- Prediction of a number. What will this house sell for? How many calls will we get on Monday? Regression models learn the relationship between inputs and the number you care about.
These models are small, fast and cheap to run once trained. The hard part is the data: it has to be clean, labelled and representative of the cases the model will see.
When is a large language model the better tool?
Large language models are trained on huge amounts of text and can follow instructions without task-specific training. They fit problems where the input is language and the rules would be endless:
- Reading an enquiry and pulling out the name, need and urgency.
- Answering questions from your own documents, with sources.
- Holding a natural conversation on the phone or in chat.
They are slower and more expensive per request than a small model, and their output needs checking. We wrap them in narrow instructions, give them tools for facts, and keep a person in the loop where mistakes are costly.
What should you have before starting?
Whatever the method, four things decide whether the project works:
- A clear question, such as "which leads should we call first", not "use AI on our data".
- Data you can access, with permission to use it.
- A way to measure success that someone on the business side agrees with.
- A plan for mistakes, because every model gets some cases wrong.
Can you combine them?
Yes, and the best systems usually do. A language model reads an incoming email and extracts the fields. A small classifier scores the lead. Plain rules decide who gets it and when. Each part does the job it is good at, and each can be tested on its own.
If you are not sure which of these your problem needs, that is the first thing we work out on a call. Often the honest answer is a few rules and a better form.
AI and machine learning