Machine Learning
Definition
Machine Learning
Machine learning is the branch of artificial intelligence in which software learns patterns from data rather than following rules written out by hand. The model is trained, not programmed — and that one difference changes how the work is built, tested and governed.

The commercial consequence is easy to miss. You are not buying a finished rule set. You are buying a process that needs data, evaluation, and a plan for the day accuracy starts to slip.
Adoption is now mainstream rather than experimental. Stanford’s AI Index reports that 78% of organisations used AI in 2024, up from 55% the year before.
Machine learning sits under artificial intelligence as one approach among several. It is the approach that now does most of the commercial work.
Key takeaways
- Machine learning learns rules from examples instead of being given them.
- A trained model is a perishable asset: accuracy decays as the world moves.
- Three families cover almost all business use — supervised, unsupervised, reinforcement.
- Most buyers need the surrounding discipline more than they need the algorithm.
How it works
A model is fitted to examples, scored on data it has never seen, then adjusted until error stops falling. Training produces the model. Inference is the model answering a live question, thousands of times a second.
The three families differ in what the training data carries.
| Family | What the data holds | Typical business use |
|---|---|---|
| Supervised | Labelled examples with known answers | Churn scoring, fraud flags, document classification |
| Unsupervised | Unlabelled records only | Customer segmentation, anomaly detection |
| Reinforcement | Rewards from trial outcomes | Routing, pricing, resource allocation |
Labelled data is the bottleneck in nearly every supervised project. Labelling is slow, repetitive, judgement-heavy work — which is exactly why so much of it is contracted out.
Before any of that, someone has to decide what the model gets to see. Choosing and shaping those inputs is called feature engineering, and it moves accuracy further than swapping one algorithm for another.
A feature has to be available at the moment the prediction is made. Including something only known afterwards produces a model that tests beautifully and fails on its first live record.
Accuracy is measured against a holdout set the model never saw during training. A model that scores well on its own training data and poorly on fresh data has memorised rather than learned.

Once live, the model degrades. Customer behaviour shifts, products change, and the patterns the model learned stop matching the world. Catching that drift and retraining is machine learning operations, a separate discipline from building the model.
Governance has caught up with the technology. The United States National Institute of Standards and Technology published the AI Risk Management Framework on January 26, 2023 as a voluntary structure for managing AI risk.
That framework is what an AI risk assessment is usually built against before a model touches a customer decision.
Examples
Machine learning earns its place where the rules are real but nobody can write them down. These four patterns cover most of what outsourced delivery teams are actually asked to build and run.
A retail bank scores card transactions for fraud in the moment of authorisation. The signal is a combination of amount, merchant, location and timing that no analyst could specify as a rule, but which thousands of labelled past frauds make learnable.
A Philippine contact centre predicts which accounts will cancel in the next 30 days and routes them to a retention queue. The model is ordinary; the value is that the list arrives before the customer calls to leave.
A healthcare provider sorts incoming clinical documents by type before a coder sees them. Classification is unglamorous and it is where most offshore machine learning work sits, because volume is high and the ground truth already exists in the archive.
An e-commerce retailer forecasts demand per item per warehouse. Seasonality, promotions and weather interact in ways a planner cannot hold in their head, and the model is judged on stockouts avoided rather than on accuracy in the abstract.
What those four share is a decision made thousands of times a day. Machine learning rarely pays off on a decision taken once a quarter, however important that decision is.
Buyers who want the capability without the hiring usually arrive at machine learning outsourcing, where a provider supplies the data preparation, the model and the monitoring as one contract.
Related terms
Machine learning touches a parent field, a job, an operating discipline and a commercial model. These are the entries worth reading next, each bounded so none of them repeats what this page covers.
- Artificial Intelligence: the parent field, of which machine learning is one approach.
- Machine Learning Operations: the practice of keeping a trained model working after launch.
- Machine Learning Engineer: the role that builds, ships and monitors models in production.
- Machine Learning Outsourcing: contracting the whole capability to an external provider.
- Data Science: the wider discipline that produces the methods machine learning applies.
- AI Risk Assessment: the review a model passes before it affects a customer.
FAQ
Is machine learning the same as artificial intelligence?
No. Artificial intelligence is the broad field; machine learning is one approach within it, defined by learning from data rather than from hand-written rules.
How much data does a model actually need?
It depends on how varied the problem is, not on a fixed number. A narrow classification task can work on a few thousand good labelled examples, while an open-ended one may need millions.
Why do models get worse over time?
Because the world they learned from keeps moving. Customers, products and fraud tactics all change, so a model trained on last year’s patterns slowly stops matching this year’s reality.
Can machine learning work be outsourced?
Yes, and labelling, model development and monitoring are routinely contracted separately or together. The buyer keeps the business question; the provider supplies the capability.
Who owns the model once it is built?
Whoever the contract says — and it is worth settling before work starts. Ownership of the trained model, the training data and any derived features are three separate questions, and providers answer them differently.
What is the most common first project?
Classification of documents or contacts, because the labels usually already exist in the archive.
Explore Outsource Accelerator for the outsourcing side of building and running models.







Independent




