Forecasting
Demand, churn, support load.
What is machine learning? Learning from data
Machine learning refers to methods that let a system learn patterns from data instead of hand-coding every rule. Classification, regression, clustering, and generative models are in scope. Google’s ML Crash Course, ISO/IEC 23053, and the NIST AI RMF frame practices, lifecycles, and risk.
In one sentence
ML learns regularities in data to predict or decide.
Key points
- Supervised, unsupervised, reinforcement: three major families.
- Data and label quality often beat algorithm choice.
- Models need out-of-sample evaluation and production monitoring (drift).
- ML ≠ LLM: many SME wins come from simple, well-targeted models.
Term at a glance
- Machine learning
- ML · Statistical learning
- English term
- Machine learning
- Domain
- Artificial intelligence
- Category
- AI fundamentals
- Level
- Beginner to intermediate
What does “machine learning” mean exactly?
Instead of writing “if amount > X then fraud,” you train on historical examples. The system generalizes to new cases—with measurable error.
ISO/IEC 23053 provides a framework for ML systems; NIST AI RMF stresses risk governance across the lifecycle.
For an SME, starting small (lead scoring, demand forecast, ticket triage) often beats a vague “AI” program.
How do you run a useful ML project?
- 01
Define the decision
What action changes if the prediction is good?
- 02
Build the dataset
Sources, labels, bias, leakage.
- 03
Train and evaluate
Simple baseline, business metrics, frozen test set.
- 04
Deploy and monitor
Latency, drift, human feedback loop.
A concrete machine-learning example
A Montreal delivery company predicts delays from history (weather, zone, volume). High-risk routes are reordered each morning. After three months, off-SLA deliveries drop 12%—no LLM, just a supervised tabular model.
What is machine learning used for?
Scoring
Leads, internal credit, prioritization.
Anomaly detection
Fraud, failures, quality outliers.
Classic vision / NLP
Business OCR, document classification.
Benefits and limits of ML
- Automates repetitive decisions
- Improves with relevant data
- Measurable (metrics, A/B)
- Mature algorithm range
- Depends on data quality
- Can encode bias
- Maintenance (retraining, drift)
- Needs data + domain skills
How does ML differ from fixed business rules?
| Machine learning | Hard-coded business rules | |
|---|---|---|
| Adaptation | Learns complex patterns | Manual if/then updates |
| Transparency | Variable (often gray box) | Fully explicit |
| Data | Needs history | Works without rich history |
| Best fit | Strong statistical signal | Strict regulatory policies |
Why ML matters for Quebec SMEs
Once you have history (sales, tickets, sensors), ML can prioritize and anticipate. The trap is chasing a flashy demo instead of a measurable decision that pays for the project.
Frequently asked questions
Do we need deep learning?
Not always. Regression, forests, gradient boosting cover many tabular cases.
How much data?
Enough to cover important cases and evaluate honestly—volume alone guarantees nothing.
ML and privacy (Law 25)?
Document purposes, minimize attributes, manage explainability when decisions affect people.
Who maintains the model?
Plan data/ML ownership and a retrain schedule from day one.
Related terms
Sources and references
Have data and a decision to improve? We frame a measurable ML case before any algorithm.
Frame an ML project