Machine learning (ML) is a branch of artificial intelligence in which software learns patterns from examples, instead of following rules a programmer wrote, and uses them to make predictions or decisions about new cases. Tom Mitchell's classic textbook calls it the study of algorithms that improve automatically through experience.1
The practical shift is in who writes the logic. With rules, a person decides that a message containing "refund" goes to billing. With machine learning, the system studies thousands of past messages and the queues they ended up in, and works out its own way to route the next one. That handles the variety of real customer language far better, but it moves the hard work to data: the model can only be as good as the examples it learned from.
Almost every AI system a company buys today is machine learning underneath, from fraud scores to large language models.
Routing a message with rules and with machine learning
An insurer receives a message from a customer. The goal is to send it to the right team on the first try.
Rules
Keywords- MatchFinds "premium" in the message. Rule: premium questions go to policy sales.
- RouteSends the message to the sales queue.
- RerouteA sales agent reads it, sees a new claim, and forwards it to claims the next day.
Right words, wrong team. The rule saw a keyword, not the request.
Machine learning
Learned- Learned fromA classifier trained on past messages labeled with the team that finally resolved them.
- PredictMotor claim, first notice of loss. Secondary topic: premium impact.
- RouteSends it to motor claims, with the premium question flagged for the reply.
Routed on meaning. The model recognized a pattern it had seen many times.
Illustrative example. A real classifier also returns a confidence score, and low-confidence messages go to a person.
How machine learning works
A machine learning model is a mathematical function with adjustable parameters. Training shows it many examples, measures how wrong its outputs are, and adjusts the parameters to be less wrong. After training, the parameters are fixed and the model is used on new cases, a step called inference.
Supervised learning
Learns from labeled examples: messages with their correct team, transactions marked as fraud or not. Most business ML is this kind.
Unsupervised learning
Finds structure in unlabeled data, such as clusters of similar complaints nobody had named yet.
Reinforcement learning
Learns by trial and reward. It is used to tune language models toward answers people prefer.
Self-supervised learning
Creates its own labels from raw data, for example by hiding the next word and predicting it. This is how LLMs are pretrained.
Where machine learning sits in AI
The terms nest inside each other. Each level below is a subset of the one above it.
Artificial intelligence
Any system that performs tasks associated with intelligenceIncludes hand-written rule systems and decision trees as well as learned models.
Machine learningThis page
Systems that learn from dataFraud scores, churn prediction, demand forecasts, recommendation, intent classification.
Deep learning
ML with many-layered neural networksSpeech recognition, image recognition, and the models behind modern language understanding.
Generative AI and LLMs
Deep learning models that create contentGenerative AI: text, images, and code. The engine inside chat assistants and AI agents.
Machine learning and neighboring approaches
| Approach | Where the logic comes from | Good for | Weak at |
|---|---|---|---|
| Rules and RPA | People write it | Stable, well-defined processes | Variety and messy language |
| Classic machine learning | Learned from labeled data | Scoring, ranking, forecasting on structured data | Open-ended language and writing |
| Deep learning | Learned from large data, layered | Speech, images, language understanding | Explaining its decisions |
| Generative AI | Pretrained on broad data, then prompted | Writing, summarizing, conversation | Exact numbers and guaranteed consistency |
Machine learning behind customer-facing AI
Inside a modern AI agent, classic ML models often do the scoring and ranking while a language model does the talking.
Intent and priority
Classify each message by intent, language, and urgency so the agent or the right team picks it up first.
message → classify → routePredict who will leave
A telecom scores customers by churn risk, and an agent contacts the high-risk ones with a relevant offer.
usage data → churn score → outreachRank the catalog
Learned ranking decides which products appear first for a query, based on what similar shoppers bought.
query → rank → showFlag risky payments
Fraud models score each transaction in real time; the agent explains a blocked payment and verifies the customer.
transaction → risk score → hold or passMatch and reconcile
Learned matching pairs invoices with orders and receipts when references are missing or misspelled.
invoice → match → post or flagScore inbound leads
Predicts which inquiries are likely to buy, so the agent qualifies those first and books them with sales.
inquiry → score → qualifyMachine learning risks and controls
Bad or biased data
ControlA model learns whatever its data contains, including old mistakes and skews against customer groups. Audit labels and test results by segment before launch.
Drift
ControlCustomer behavior, products, and fraud patterns change, and accuracy falls without warning. Monitor live accuracy and retrain on recent data.5
Hidden maintenance cost
ControlResearchers at Google described the large ongoing maintenance costs of real-world ML systems as technical debt.4 Budget for monitoring, data pipelines, and retraining, not only the first model.
Opaque decisions
ControlSome models cannot explain a single prediction well. Where a decision affects a customer's money or access, log the inputs and keep a person able to review and overturn it.
Common questions
What is the difference between AI and machine learning?
AI is the broad goal of building systems that perform tasks associated with intelligence. Machine learning is the most common way to get there today: learning the behavior from data instead of programming it by hand.
Is ChatGPT machine learning?
Yes. It runs on large language models, which are deep learning models trained with self-supervised learning on text and then tuned with human feedback.
What is the difference between machine learning and deep learning?
Deep learning is a subset of machine learning that uses neural networks with many layers. It needs more data and compute, and it handles speech, images, and language far better than earlier methods.
How much data does a machine learning model need?
It depends on the task. Narrow tasks with a few clear categories need far less data than open-ended ones. Starting from a pretrained language model usually reduces the number of labeled examples needed.
References
- Mitchell, T. Machine Learning. McGraw Hill, 1997.
- LeCun, Y., Bengio, Y. and Hinton, G. Deep learning. Nature 521 (2015): 436-444.
- Zinkevich, M. Rules of Machine Learning: Best Practices for ML Engineering. Google for Developers.
- Sculley, D. et al. Hidden Technical Debt in Machine Learning Systems. NeurIPS 2015.
- Lu, J. et al. Learning under Concept Drift: A Review. 2020.
- European Commission. AI Act, Article 3: Definitions. Regulation (EU) 2024/1689.