CyberFreezeDev- Home
- Blog
- How-to guides
- Machine learning for small business: do you need it yet?
Machine learning for small business: do you need it yet?
Most small businesses should try a spreadsheet or a written rule before machine learning. Here is what forecasting, classification and anomaly detection do, the data each needs, and the signs you are not ready yet.
Most small businesses should try a spreadsheet formula or a written rule before machine learning. Machine learning starts to pay when the pattern is too messy for a rule, the same decision comes up hundreds of times a month, and you have a long, consistent record of past cases with known outcomes. Google's published guide to machine learning engineering makes the point in rule one: "Don't be afraid to launch a product without machine learning."
This guide is for owners and operations heads who keep hearing that they should "use machine learning" and want to know whether it applies to them.
What is the difference between a rule and machine learning?
A rule is a pattern you write down: "if a customer has paid late twice, ask for advance payment". A spreadsheet formula is a rule too. Machine learning is a pattern the computer finds for itself from past examples where you already know the outcome, and then applies to new cases.
Rules are cheap, easy to explain and easy to change. Their weakness shows when you need dozens of them and they start to contradict each other. Google's guide covers this as well: a simple heuristic gets you started, but "a complex heuristic is unmaintainable", and once you have data, that is the moment to move to machine learning.
Three kinds of machine learning a small business might use
| Kind | The question it answers | Small business example | Try this first |
|---|---|---|---|
| Forecasting | How many, and when? | Next month's orders per product for a distributor | Same month last year, adjusted for growth |
| Classification | Which kind is this? | Sorting incoming enquiries by intent, or flagging invoices likely to be paid late | Keyword rules, or a customer's payment history |
| Anomaly detection | Is this unusual? | An expense claim, bank line or machine reading that does not fit | Fixed limits and a duplicate check |
Forecasting predicts a quantity over time: demand, cash coming in, staff needed on a Saturday. It beats a simple average when demand depends on several things at once, such as festivals, price changes, weather and promotions, and you have enough history to see each of them more than once.
Classification puts each item into one of your categories: a hot lead or a browser, a complaint or a question, a valid claim or one to check. It needs past examples that someone has already labelled correctly. A support inbox where every ticket was tagged by hand for two years is good material. An inbox where nobody tagged anything is not.
Anomaly detection asks whether a new record looks like the ones before it. It helps when "normal" is hard to define with a single limit: a supplier's invoices that are usually between two amounts except in March, or a motor that draws more current on hot days. If one threshold catches the problems, keep the threshold.
What data do you need before it is worth it?
- Outcomes, not just records. For classification you need the right answer for past cases: which enquiries became orders, which invoices were paid late. Many businesses have the records but never wrote down the outcome.
- Enough history. There is no magic number. The textbook Forecasting: Principles and Practice calls rules of thumb on minimum sample size misleading, and says you usually need substantially more observations than the model has parameters. With short histories, simple methods tend to win. If Diwali is your peak, a model needs several Diwalis to learn it.
- More than a handful of cases. scikit-learn's chart for choosing a method starts by asking whether you have more than 50 samples, and if not, its advice is "get more data". Treat that as the floor, not the target.
- Consistency. The same product codes, categories and customer names over time, in one place rather than spread across Tally, three spreadsheets and WhatsApp chats.
- A measure of better. Decide how you will compare the model with what you do today, on cases it has not seen.
Signs you are not ready yet
- You can write the rule in one sentence and it is right almost every time.
- The decision comes up a few times a week, not hundreds of times a month, so even a perfect model saves little.
- Nobody recorded the outcomes you want to predict.
- The data lives in chats, paper files and spreadsheets that use different names for the same customer.
- Nobody will act differently because of the prediction.
- A wrong answer would be costly and there is no one to review it.
If two or more of these apply, the better project is usually cleaning up how data is recorded, with a rule or a spreadsheet doing the job meanwhile. Six months of tidy records will tell you more than any model trained on messy ones.
How to decide in practice
Write down the decision you want help with, how often it happens, what a wrong answer costs and what data you already hold. Then try the simplest rule on last quarter's cases and count how often it was right. If the rule is good enough, stop there. If it is not, you now have a baseline any model must beat.
We take the same view in our custom AI systems work: when a simple rule or a database query does the job well, we say so. Our AI strategy consulting ranks candidate projects by value, effort and risk, including the ones not worth doing yet, and our guide to working with an AI consultant as a small business covers where to start. If your team wants to learn to test ideas like this themselves, see our AI training workshops.
