On this page
Machine learning is the part of artificial intelligence that learns patterns from historical data and applies them to new cases. Unlike traditional software, which follows rules a developer wrote, a machine learning system derives its own rules from examples, and keeps improving as it sees more of them.
That distinction matters commercially. It means machine learning is well suited to problems where the rules are hard to write down but examples are plentiful, and poorly suited to problems where the rules are already clear.
How machine learning actually works
At a practical level, every machine learning project follows the same shape:
- Assemble labelled examples. Past invoices marked paid or unpaid. Past enquiries tagged by outcome. Past scans marked normal or abnormal.
- Train a model. The algorithm adjusts internal parameters until its predictions on those examples are as accurate as possible.
- Validate on data it has never seen. This is the honest test. Performance on training data always looks better than reality.
- Deploy and monitor. Real-world data drifts. A model that was accurate in January can quietly degrade by June.
The three main families you will encounter:
| Type | What it learns from | Typical business use |
|---|---|---|
| Supervised learning | Labelled historical examples | Demand forecasting, churn prediction, document classification |
| Unsupervised learning | Unlabelled data | Customer segmentation, anomaly and fraud detection |
| Reinforcement learning | Trial, error, and reward | Pricing, routing, resource scheduling |
Where machine learning adds real value
Across UK SME and charity engagements, the returns cluster in five areas:
Forecasting. Demand, cash flow, donation income, staffing requirements. Machine learning handles seasonality and multiple interacting variables better than a spreadsheet trend line.
Prioritisation. Which leads to call first, which cases to escalate, which equipment to service. The model does not make the decision; it orders the queue so human attention goes where it pays.
Classification at volume. Routing thousands of emails, extracting fields from invoices, tagging support tickets. Dull, high-volume, rule-resistant work is machine learning’s home ground.
Anomaly detection. Spotting the transaction, expense claim, or sensor reading that does not fit the pattern. Effective precisely because nobody has to define “unusual” in advance.
Personalisation. Matching content, products, or interventions to individuals, with the caveat that this is the use case regulators scrutinise most closely.
Where it does not
Be equally clear about the poor fits:
- Small data. A few hundred examples rarely supports a reliable model. Rules-based automation is usually the better answer.
- Stable, well-understood rules. If a policy document already specifies the decision, encode the policy. It will be cheaper, faster, and explainable.
- One-off decisions. Machine learning earns its cost through repetition.
- Decisions requiring justification you cannot give. If you cannot explain to an affected person why a decision was made, you have a compliance problem regardless of accuracy.
The governance you need before you start
Machine learning applied to personal data engages UK GDPR. Three obligations do most of the work:
- Lawful basis and purpose limitation. Data collected for service delivery cannot automatically be repurposed as training data. Check before you train.
- Article 22: automated decision-making. Where a decision produces legal or similarly significant effects, individuals have rights to human intervention and to contest the outcome. Design the appeal route in from the start.
- Fairness testing. Historical data encodes historical bias. Test outcomes across protected characteristics before launch and at intervals afterwards, and record what you found.
Add to that a documented model register, a named owner for each model in production, and a monitoring plan for drift. This is not heavy bureaucracy (for most SMEs it is four or five pages), but its absence is what turns a technical success into an organisational liability.
A realistic first project
The pattern that works:
- Choose a repetitive, high-volume decision with a measurable baseline.
- Confirm you hold at least a year of clean, labelled history.
- Run a four-to-six week pilot with the model advising rather than deciding.
- Compare model recommendations against human decisions and quantify the delta.
- Only then discuss automation, and only where the governance case is as strong as the commercial one.
Most failed machine learning projects we review failed at step two. The data was messier, thinner, or more inconsistent than anyone had checked. An honest data audit is the cheapest risk reduction available.
If you are weighing up a first machine learning project, our AI readiness assessment covers data quality, governance, and commercial fit in a single structured review.