AI in Action: Key Branches and Applications
Traditional ML in Business: The Quiet Workhorse
8 min read
The AI you don't see on the news
When people talk about exciting AI developments, they tend to mean robots, chatbots, or image generators. But a substantial portion of AI's actual economic value comes from something far less glamorous: traditional machine learning applied to business operations.
These models don't write poetry or drive cars. They answer questions like: will this customer cancel their subscription next month? Is this credit card transaction fraudulent? How much inventory should we hold in this warehouse next Tuesday? Which customers are most likely to respond to this promotion?
They don't make headlines. But they run quietly in the background of nearly every major industry, making millions of decisions per day.
What "traditional ML" means
In contrast to deep learning — which uses multi-layered neural networks to learn features automatically from raw unstructured data — traditional machine learning typically uses structured, tabular data and explicit feature engineering. The models include:
- Logistic regression — classification from linear combinations of features
- Decision trees and random forests — branching decision rules learned from data
- Gradient boosting (XGBoost, LightGBM) — state-of-the-art for most tabular data problems
- Support vector machines — finding the optimal boundary between classes
- k-nearest neighbors — classifying based on the closest examples in the training set
For many real-world business problems — where data is tabular, features are well-defined, and interpretability matters — these approaches outperform deep learning while being faster to train, cheaper to run, and easier to explain to decision-makers.
Demand forecasting
Every retailer, manufacturer, and logistics company faces the same core problem: how much of which product do I need, where, and when?
Get it wrong in one direction and you're left with unsold inventory clogging up warehouse space. Get it wrong in the other direction and you're out of stock during a peak demand period, turning away customers.
Demand forecasting models learn from historical sales data — alongside signals like seasonality, promotional calendars, local events, and weather — to predict future demand at the product-location-week level.
Walmart was an early pioneer, building demand forecasting systems in the 1990s that helped it maintain industry-leading inventory efficiency. Amazon's supply chain relies heavily on ML-based demand forecasting to pre-position inventory in fulfillment centers closest to where orders are predicted to originate.
The models behind these systems are typically gradient-boosted trees — not deep neural networks. Tabular data with explicit features (day of week, proximity to a holiday, trailing four-week sales trend) is exactly the terrain where these models dominate.
| Feature type | Example |
|---|---|
| Temporal | Day of week, week of year, days until next holiday |
| Historical | Last 4 weeks of sales, same week last year |
| Promotional | Active promotion? What discount? |
| External | Weather forecast, local events |
| Product | Category, price tier, shelf position |
Customer churn prediction
A customer who is about to cancel their subscription often leaves behavioral signals well before they actually leave. They log in less frequently. They open fewer emails. Their usage of key features drops.
Churn prediction models learn to detect these signals. Trained on historical data — the behaviors of customers who did and didn't churn — a model assigns each current customer a probability score: how likely are they to leave in the next 30 days?
This score drives action. A high churn-risk customer might receive a targeted retention offer or a call from a customer success manager. A low-risk customer gets none of that — reaching out unnecessarily wastes resources and can feel intrusive.
The economic case is straightforward: acquiring a new customer costs five to seven times more than retaining an existing one. An accurate churn model that improves retention by even 2% can be worth tens of millions of dollars annually for a large subscriber business.
The models here are usually logistic regression, random forests, or gradient boosting — not deep learning. The features are tabular, and the model needs to produce an interpretable probability score that a business analyst can act on.
Credit scoring and loan underwriting
When a bank decides whether to approve a loan, it's making a prediction: will this person repay this debt?
The traditional approach was the credit score — a single number calculated from a fixed formula based on payment history, credit utilization, and account age. FICO scores, introduced in 1989, are this approach codified.
Modern ML-based underwriting goes further. Rather than a fixed formula, a machine learning model can incorporate thousands of variables — including non-traditional signals like rental payment history and employment stability — and learn complex non-linear interactions between them.
Fintech lenders have built business models around this: using ML models that incorporate factors FICO ignores to better predict creditworthiness, extending credit to people who would have been rejected under traditional scoring while maintaining or improving default rates.
The models here must be interpretable and auditable. Regulations like the Equal Credit Opportunity Act require lenders to provide applicants with specific reasons for adverse decisions. A deep neural network that outputs a probability with no explanation is not compliant. Logistic regression and gradient boosting — where feature importance can be extracted and explained — remain dominant for exactly this reason.
Fraud detection
Credit card fraud detection is one of the oldest and highest-stakes applications of ML. Visa and Mastercard process billions of transactions per year; a small fraction are fraudulent. The challenge: flag the fraudulent ones without generating so many false positives that legitimate customers find their cards blocked at checkout.
A model learns what "normal" looks like for each cardholder — their typical spending amounts, merchant categories, geographic locations, and times of day. A transaction that deviates sharply from that pattern triggers a high fraud score.
The class imbalance is extreme: perhaps one fraudulent transaction per several thousand legitimate ones. Specialized techniques — oversampling minority classes, anomaly detection, and threshold tuning — are needed to catch fraud reliably without crying wolf constantly.
Traditional ML excels here: gradient boosted trees, logistic regression, and isolation forests (an algorithm specifically designed for anomaly detection) are widely deployed. The features are structured, the model needs to run in milliseconds, and the predictions need to be explainable when a cardholder disputes a declined charge.
Predictive maintenance
A factory floor is full of equipment that will eventually fail. The traditional approach: maintain on a fixed schedule regardless of actual machine condition, or wait until something breaks — which usually happens at the worst possible time.
Predictive maintenance takes sensor data from industrial equipment — vibration, temperature, current draw, acoustic emissions — and trains models to detect the early signatures of impending failure. If a bearing is beginning to wear, it produces a slightly different vibration pattern weeks before it seizes. A model trained on patterns preceding past failures can flag equipment for maintenance before it causes a production stoppage.
Unplanned downtime in manufacturing is estimated to cost industrial companies $50 billion per year globally. Even a modest reduction in unplanned failures pays for the entire ML system many times over.
Why traditional ML often beats deep learning for business
It's tempting to assume that deep learning — being newer and more powerful — is always the right choice. For most business problems involving structured data, it isn't.
| Factor | Traditional ML | Deep Learning |
|---|---|---|
| Data type | Tabular, structured | Unstructured (images, text, audio) |
| Data volume needed | Thousands to hundreds of thousands | Often millions |
| Training cost | Low — runs on a laptop | High — requires GPUs |
| Inference speed | Milliseconds | Slower unless optimized |
| Interpretability | Feature importances, decision paths | Largely opaque |
| Regulatory fit | Good — explainable decisions | Poor for regulated decisions |
| Best business fit | Fraud, churn, forecasting, scoring | Vision, language, generation |
The practical rule of thumb: if your data lives in rows and columns, start with gradient boosting before reaching for deep learning. The simpler model will often be more accurate, cheaper to run, and far easier to explain to the stakeholder who has to approve deployment.
Key takeaway
Traditional machine learning — decision trees, random forests, gradient boosting, logistic regression — runs the vast majority of AI that drives measurable business value today. Demand forecasting, churn prediction, credit scoring, fraud detection, and predictive maintenance are all domains where structured tabular data and explicit features make traditional ML faster, cheaper, and more interpretable than deep learning alternatives. Deep learning is transformative for unstructured data; traditional ML is transformative for the structured data that underlies most business operations.
What's next?
We've covered AI that classifies and predicts. But the newest and fastest-growing branch of AI doesn't just analyze the world — it creates new content: images, text, code, video, and music. That's generative AI, and it's the subject of the final lesson in this section.