Most business people have a slightly uncomfortable relationship with statistics. You know the numbers matter. You are not entirely sure which method your analyst used, whether it was the right one, or what would have made the answer different.
The good news: for the overwhelming majority of business questions, eight methods cover it. AI tools can now run all eight for you. What AI cannot do is decide which question you are actually asking — and that is the part that determines whether the answer is useful or dangerous.
Here is the map, in plain language.
Which method answers which question
1. Descriptive statistics — "what happened?"
Averages, medians, percentages, growth rates. The foundation of every management report.
The trap everyone falls into: using the average when you should use the median. Say ten customers spend RM100 and one spends RM50,000. The average spend is about RM4,627 — a number describing nobody. The median is RM100, which describes your actual typical customer.
Any time your data has a few very large values — revenue, property prices, salaries, transaction sizes — the average is misleading and the median is honest. Ask for both. If they differ a lot, the distribution is skewed and the average is hiding something.
Business use: monthly reporting, KPI dashboards, board packs.
What descriptive statistics hides
2. Segmentation — "who are they?"
Splitting customers, products or branches into meaningful groups so you can treat them differently.
The most practical version in business is RFM analysis: rank every customer on how Recently they bought, how Frequently they buy, and how much Monetary value they bring. Three columns you almost certainly already have in your POS or invoicing system, and they reveal who is loyal, who is drifting away, and who was never worth chasing.
Business use: targeted campaigns, retention programmes, loyalty tiers, deciding who your sales team should call first. We go deep on this in the RFM article.
What segmentation gives you
3. Correlation — "do these move together?"
A measure of whether two things rise and fall together. Ice cream sales and drowning incidents correlate strongly — both rise in hot weather. Neither causes the other.
The trap: treating correlation as proof that one thing causes another, then spending budget on that assumption. This is the single most expensive statistical mistake in business, which is why it gets its own article.
Business use: finding candidate drivers worth investigating — never as final proof.
What correlation looks like — and does not prove
4. Regression — "how much does X affect Y?"
Correlation tells you two things move together. Regression tries to quantify by how much, while holding other factors constant.
"Every RM1,000 of additional ad spend is associated with 14 more leads, controlling for seasonality and promotion periods." That is a regression output, and it is far more actionable than a correlation.
The trap: the phrase "controlling for" only covers variables you actually included. If something important was left out — a competitor's price cut, a viral post, a public holiday — the number is wrong in ways the model cannot tell you about.
Business use: pricing decisions, marketing-mix analysis, understanding cost drivers.
What regression adds over correlation
5. Time series and forecasting — "what happens next?"
Analysing data ordered over time, and projecting it forward. Distinct from other methods because the sequence itself carries information: last month influences this month.
Any business time series breaks into three parts — trend (the long-run direction), seasonality (repeating patterns like Ramadan, year-end, payday cycles), and noise (everything else). Separating them is most of the work.
The trap: quoting a forecast as a single number. A forecast without a range is a guess wearing a suit. "Sales next month will be RM480,000" is far less useful than "RM440,000 to RM520,000, with 80% confidence."
Business use: inventory planning, cash-flow projection, staffing, budgeting. Covered fully in the forecasting article.
What a time series actually contains
6. Significance testing — "is this real or is it noise?"
You changed the checkout page and conversion went from 3.1% to 3.4%. Real improvement, or normal week-to-week fluctuation?
Significance testing estimates how likely you would see a difference this large if nothing had actually changed. It is the discipline that stops teams celebrating randomness.
The trap: the American Statistical Association's 2016 statement on p-values is unusually direct about the misreadings here: a p-value does not tell you the probability that your hypothesis is true, and "statistically significant" does not mean "commercially important." A 0.05% lift can be statistically significant with enough traffic and still be worth nothing.
Business use: A/B tests, pricing experiments, campaign evaluation. See the A/B testing article.
What significance testing decides
7. Cohort analysis — "does this get better or worse over time?"
Group customers by when they joined, then track each group separately as it ages.
This is the method that reveals problems a headline number conceals. Total revenue can climb every month while every individual cohort retains worse than the one before — growth masking a deteriorating product. You only see it when you split by cohort.
Business use: retention and churn analysis, subscription businesses, measuring whether product changes actually improved anything.
What cohort analysis reveals
8. Anomaly detection — "what looks wrong here?"
Finding the transactions, days or accounts that do not fit the pattern.
This is where AI genuinely outperforms manual review, simply on volume. A person can eyeball a few hundred rows; a model can flag the unusual ones across millions.
The trap: anomalies are not automatically problems. A flagged transaction is a candidate for review, not a verdict. Every anomaly system needs a human decision step, and a tolerance for false positives that your team can actually sustain.
Business use: fraud screening, error detection in month-end close, data-quality checks, operational monitoring.
What anomaly detection surfaces
Where AI genuinely helps — and where it does not
AI has made all eight of these methods dramatically more accessible. Describe your data and your question in plain English and a capable model will run the analysis, produce the chart, and explain the result. That is a real change: the technical barrier that kept these methods inside specialist teams has largely dropped.
What has not changed:
- Choosing the right question. AI will faithfully answer the question you asked, including when it is the wrong one.
- Knowing whether the data can support the answer. Three months of history cannot tell you about annual seasonality, no matter how sophisticated the method.
- Business judgement about what matters. A statistically significant result on a metric nobody acts on is an expensive way to produce a slide.
- Noticing what is missing. Models analyse the data in front of them. They do not know about the branch that stopped reporting, or the promotion nobody logged.
A sensible starting point
If your organisation does little formal analysis today, start with descriptive statistics done properly (median alongside average, segmented by the dimensions that matter) and RFM segmentation. Those two produce more decisions-changed-per-hour than anything else on this list.
Add forecasting when planning accuracy starts costing you money. Add significance testing when you begin running experiments. Add the rest as the questions arise.
Our AI Analytics programme teaches these methods hands-on using AI tools, for Malaysian business teams rather than statisticians — HRD Corp SBL-KHAS claimable for eligible employers.
