Machine learning (ML) is a branch of artificial intelligence in which software learns patterns from historical data and uses them to make predictions, classify things or spot unusual activity, instead of following rules a programmer wrote by hand. A spam filter that improves as it sees more email, a forecast that learns from past sales and a fraud check that flags odd transactions are common examples. Most products sold as AI, including generative AI, are built on machine learning.
At a glance
- ML systems learn from examples rather than explicit rules, so their quality depends heavily on the data they learn from.
- Common business uses are forecasting, classification, anomaly detection, recommendations and scoring.
- A trained model produces probabilities, not certainties; it will be wrong some of the time.
- Models need ongoing monitoring and retraining as conditions change.
- Many buyers get ML inside existing software rather than building their own models.
What problem it solves
Some business questions are too complex or change too often for hand-written rules. Which customers are likely to leave? Which invoices look like fraud? Which server is about to fail? Writing a rule for every case is impractical, and rules that worked last year quietly stop working.
Machine learning addresses this by finding the patterns in past data and applying them to new cases. It can weigh many factors at once, update as new data arrives and process volumes no team could review by hand. For a mid-market company, the practical result is usually better forecasts, earlier warnings and less manual sorting, delivered through tools the business already uses, such as analytics and business intelligence platforms.
How it works
Data. You start with historical examples relevant to the question: past transactions marked as fraud or not, past sales by week, past tickets by category. Cleaning and preparing this data is often the largest part of the work.
Training. An algorithm studies the examples and adjusts a model until it predicts the known outcomes well. Broad families of methods include:
- Supervised learning, where examples come with the right answer (this email was spam).
- Unsupervised learning, where the model finds groupings or outliers without labeled answers.
- Reinforcement learning, where a system learns by trial, error and reward.
- Deep learning, which uses large neural networks and underpins most language, speech and image models.
Testing. The model is checked against data it hasn’t seen to estimate how often it will be right, and where it tends to fail.
Deployment and monitoring. The model is put into an application, where it scores new cases. Its accuracy is tracked over time, and it is retrained when performance drifts.
When it matters for buyers
- When a vendor says its product “uses AI”. Ask what the model actually predicts, what data it learned from and how accuracy is measured.
- When planning analytics or AI projects. ML projects succeed or fail on data quality, so data governance usually comes first. Our analytics and business intelligence overview covers the platforms involved.
- When decisions affect people. Models used in hiring, lending, pricing or similar decisions can carry bias from their training data, and some jurisdictions regulate these uses; check with counsel.
- When budgeting. Training and running models, especially large ones, can consume significant compute, and retraining is a recurring cost.
- When choosing build vs buy. Many ML features come built into CRM, security, IT operations and finance tools; custom models make sense where your data and problem are distinctive.
Questions to ask vendors
- What exactly does the model predict or detect, and what action does it drive?
- What data was it trained on, and is it trained on our data, pooled customer data or both?
- How is accuracy measured, and what are the false positive and false negative rates in practice?
- How do you detect drift, and how often is the model retrained?
- Can we see why the model made a given decision?
- Who owns models trained on our data, and what happens to them if we leave?
How it differs from artificial intelligence (AI)
Artificial intelligence (AI) is the broad field of making software perform tasks that normally need human judgment. Machine learning is the dominant way of achieving that: learning from data instead of being explicitly programmed. Older AI approaches, such as expert systems built from hand-written rules, are AI but not machine learning. Generative AI and large language models (LLMs) are a newer kind of machine learning focused on producing content rather than predicting a value, and natural language processing (NLP) applies ML to human language.