A machine learning model is what you get when you train an algorithm on data: a saved set of learned parameters that takes new input and returns a prediction, classification or grouping — without anyone writing explicit rules for each case. The algorithm is the method; the model is the trained artefact that method produces.
Most explanations stop at the taxonomy. This one covers the taxonomy and the part that determines whether any of it matters commercially — what happens after the model works in a notebook. Organisational AI adoption reached 88% of surveyed organisations in 2025, up from 78% a year earlier, yet the same research found scaled deployment still sitting in the single digits across nearly every business function (Stanford HAI, 2026 AI Index Report, Ch. 4). The gap between "we have a model" and "it runs reliably" is where most ML budgets are actually spent.
Key Takeaways
- A model is a trained artefact; an algorithm is the training method. Linear regression is an algorithm. The fitted coefficients you deploy are the model.
- There are four learning types — supervised, unsupervised, semi-supervised and reinforcement — and roughly ten model families that matter in practice.
- Tabular business data rarely needs deep learning. Gradient-boosted trees (XGBoost, LightGBM) remain the default winner on structured data; neural networks earn their cost on images, audio, video and language.
- Adoption is near-universal; deployment is not. 88% of organisations report using AI in at least one business function, but AI agent deployment remains in single digits across most functions (Stanford HAI, 2026).
- Accuracy decays after launch. Data drift means a model that scored 94% at deployment will not score 94% a year later unless it is monitored and retrained.
- Trust is still a live constraint. DORA's 2025 study of nearly 5,000 technology professionals found 30% report little or no trust in AI-generated code, even as 90% use AI at work (DORA, 2025).
What Is a Machine Learning Model?
A machine learning model is a file — often surprisingly small — containing the parameters an algorithm learned from training data, plus the structure needed to apply them. Give it an input it has never seen and it returns an output: a number, a label, a probability, a cluster assignment, or generated text.
The defining characteristic is that nobody wrote the rules. A traditional program applies logic a developer specified. A model infers its logic from examples, which is why it can handle problems too messy to enumerate — recognising a product photo, scoring credit risk, predicting which customer will churn.
Model vs algorithm: the distinction that matters
These get used interchangeably and they are not the same thing:
- An algorithm is a procedure — random forest, gradient descent, k-means. It exists before you have any data.
- A model is the output of running that algorithm over a specific dataset. It exists only after training.
The practical consequence: two teams using the identical algorithm on different data produce completely different models with completely different behaviour. The algorithm is rarely your competitive advantage. The data, and what you do with the model afterwards, usually is.
The Four Types of Machine Learning
Supervised learning
The model trains on labelled examples — inputs paired with known correct outputs — and learns to predict the label for new inputs. This covers the large majority of commercial ML.
Two sub-types: classification predicts a category (spam / not spam, will churn / won't churn), and regression predicts a continuous number (next month's demand, a property valuation).
Requires labelled data, which is usually the expensive part.
Unsupervised learning
The model receives unlabelled data and finds structure in it without being told what to look for — clustering customers into segments, reducing hundreds of variables to a manageable few, flagging transactions that don't resemble anything else.
Useful when you don't know what you're looking for. Harder to evaluate, because there's no correct answer to score against.
Semi-supervised learning
A small labelled dataset combined with a large unlabelled one. The practical middle ground when labelling everything is unaffordable — common in medical imaging and document classification, where expert labelling is the bottleneck.
Reinforcement learning
An agent learns by acting in an environment and receiving rewards or penalties, optimising for cumulative reward over time. It powers robotics, game-playing systems and some recommendation and pricing engines, and it underpins the alignment step in modern language models.
Powerful, data-hungry, and operationally the most demanding of the four.
The Main Machine Learning Model Families
| Model family | What it does | Best suited to | Main limitation |
|---|---|---|---|
| Linear / logistic regression | Fits a straight-line relationship between inputs and output | Baselines, credit scoring, anything needing an auditable explanation | Misses non-linear patterns |
| Decision trees | Splits data by feature thresholds into branching rules | Rule-like decisions a human must review | Overfits badly on its own |
| Random forest | Averages many decorrelated trees | Tabular data, robust general-purpose accuracy | Harder to interpret than one tree |
| Gradient boosting (XGBoost, LightGBM, CatBoost) | Builds trees sequentially, each correcting the last | Structured/tabular business data — usually the accuracy winner | Needs tuning; can overfit small data |
| Support vector machines | Finds the widest separating boundary between classes | Small, high-dimensional datasets like text classification | Scales poorly to large datasets |
| k-nearest neighbours | Classifies by majority vote of nearest examples | Recommendation, simple similarity matching | Slow at inference; degrades with many features |
| Naive Bayes | Applies Bayes' theorem assuming feature independence | Spam filtering, fast text classification | The independence assumption is usually false |
| k-means clustering | Groups data into k clusters by proximity | Customer segmentation, anomaly detection | You must choose k in advance |
| Neural networks / deep learning | Layered weighted transformations learning hierarchical features | Images, audio, video, complex non-linear signals | Data-hungry, expensive, opaque |
| Transformers | Attention-based sequence models | Language, translation, code, generative AI | Very expensive to train; costly to serve |
The heavier families — deep learning and transformers — are also where the infrastructure bill lands, because serving them is a permanent, often GPU-backed cost rather than a one-off. How that serving layer is engineered is cloud and DevOps work, not data science.
How to Choose a Machine Learning Model
Four questions settle it faster than any benchmark table:
1. Is your data tabular or unstructured? Rows and columns → start with gradient boosting. Images, audio, video or free text → neural networks or a pre-trained transformer.
2. Do you need to explain individual decisions? In lending, insurance and hiring you often must justify a specific outcome to a regulator or a customer. That rules out opaque models regardless of accuracy, and pushes you toward logistic regression or a shallow tree.
3. How much labelled data do you have? Under a few thousand labelled rows, simple models beat complex ones almost every time. Deep learning needs volume to justify itself.
4. What is your latency and cost budget? A model answering inside a checkout flow has a few hundred milliseconds. A model producing an overnight batch forecast has hours. This constraint eliminates more candidate architectures than accuracy ever does.
Always train a boring baseline first. Logistic regression takes an afternoon and gives you a number to beat. Teams that skip it routinely spend three months on a neural network that beats nothing.
What It Actually Takes to Run a Model in Production
This is the half of the subject that most explainers omit, and it's where the money goes.
The adoption-to-deployment gap
Stanford's 2026 AI Index found organisational AI adoption climbed to 88% of surveyed organisations in 2025 from 78% in 2024 — but that adoption is broad rather than deep. AI agent deployment remained in the single digits across nearly all business functions, with scaled use concentrated in the technology sector: 24% in software engineering, 22% in IT, 21% in service operations (Stanford HAI, Ch. 4). The report notes its survey data is self-reported and "directional rather than comprehensive" — a caveat worth carrying.
A trained model is perhaps 20% of the work. The rest is everything that keeps it answering correctly at 3am.
What production demands that a notebook doesn't
- A serving layer — a versioned API with authentication, rate limiting and a rollback path
- A feature pipeline — the exact same transformations at inference as at training, or predictions silently degrade
- Reproducibility — versioned data, code and model artefacts, so any prediction can be traced to what produced it
- Monitoring — not just uptime, but input distributions and prediction distributions
- A retraining path — automated, tested, and cheaper than a person doing it manually
- Fallback behaviour — what the product does when the model is unavailable or returns low confidence
Most of that list is engineering, not modelling — the same cloud and DevOps discipline that keeps any production system reliable. It's also why teams shipping their first AI feature into a SaaS product spend far longer on the plumbing than on the model.
Model drift: why accuracy decays
A model learns the world as it was during training. The world moves. Data drift is when input distributions shift — new customer demographics, a changed product mix. Concept drift is when the underlying relationship changes, as when a fraud pattern that was rare becomes common.
Neither throws an error. The model keeps returning confident predictions that are quietly wrong, which is why an unmonitored model is more dangerous than an obviously broken one. Monitoring input distributions against the training baseline is what surfaces it.
The infrastructure bill
Training is a spike; inference is a permanent line item. A model answering a million requests a day runs continuously, and GPU-backed inference is materially more expensive than the CPU workloads most teams have budgeted for.
The context in India is a market where end-user public cloud spending is forecast to grow 28.1% to $17.5 billion in 2026, with infrastructure-as-a-service the fastest-growing segment at 40% — driven substantially by AI-ready infrastructure demand (Gartner, June 2026). Serving costs are rising, not falling, and they belong in the business case before the model is built — the levers to keep them down are the same ones in our AWS cost optimisation guide.
Why Machine Learning Models Fail in Production
In rough order of how often they're the actual cause:
- The problem wasn't a prediction problem. A rule, a lookup table or a better report would have solved it.
- Training-serving skew. Features computed differently in production than in training — the most common silent failure in ML systems.
- Data leakage. Something correlated with the answer crept into training. Offline accuracy is spectacular; live accuracy collapses.
- No owner after launch. The data scientist moved to the next project and nobody watches the dashboards.
- Nobody defined success in business terms. "94% accurate" answers nothing if the cost of a false positive differs from a false negative.
- Trust was never established. DORA's 2025 research found 30% of technology professionals report little or no trust in AI-generated code, even with 90% using AI daily (DORA, 2025). A model people override is a model that delivers nothing.
Where Codevibe fits
We don't sell machine learning as a service — but every model that reaches production runs on infrastructure someone has to build and keep affordable: a serving layer, a feature pipeline, monitoring, and a retraining path. That's cloud and DevOps and web and application development work, and it's most of the real bill. If you have a model that works in a notebook and need it to run reliably — and cheaply — in production, tell us what you're building.
Frequently Asked Questions
What is a machine learning model in simple terms?
A machine learning model is the result of training an algorithm on data. It's a stored set of learned parameters that takes new input and produces an output — a prediction, a category, a score or generated content — without a developer having written explicit rules for each case.
What is the difference between a machine learning model and an algorithm?
The algorithm is the method used to learn; the model is what that method produces. Random forest is an algorithm. The specific trained forest you deploy, carrying parameters learned from your data, is the model. One algorithm can produce unlimited different models depending on the data it's trained on.
What are the four main types of machine learning?
Supervised learning (trains on labelled data to predict labels), unsupervised learning (finds structure in unlabelled data), semi-supervised learning (combines a small labelled set with a large unlabelled one), and reinforcement learning (learns through rewards and penalties from acting in an environment).
Which machine learning model is best for tabular business data?
Gradient-boosted tree models — XGBoost, LightGBM or CatBoost — are usually the strongest performers on structured, tabular data and generally beat neural networks on it. Start with logistic regression as a baseline, then move to gradient boosting.
How long does a machine learning model stay accurate?
There's no fixed lifespan. Accuracy decays through data drift (input distributions change) and concept drift (the underlying relationship changes). Models in fast-moving domains such as fraud or pricing may need retraining monthly; stable physical-process models can run for years. Because drift produces no error message, monitoring input and prediction distributions is the only reliable way to detect it.
Do I need deep learning for my business problem?
Usually not. Deep learning earns its cost on unstructured data — images, audio, video, language. For tabular business data with fewer than roughly 100,000 rows, gradient-boosted trees are typically more accurate, far cheaper to train and serve, and considerably easier to explain.
Sources
- The 2026 AI Index Report, Chapter 4: Economy — Stanford HAI — organisational AI adoption 88% in 2025, up from 78% in 2024; AI agent deployment in single digits across most business functions; scaled agent use 24% software engineering / 22% IT / 21% service operations. Underlying survey: McKinsey & Company, 2025. The AI Index notes results are self-reported and "directional rather than comprehensive."
- Announcing the 2025 DORA Report: State of AI-Assisted Software Development — Google Cloud — nearly 5,000 respondents; 90% use AI at work; more than 80% report increased productivity; 30% report little or no trust in AI-generated code.
- Gartner Forecasts End-User Public Cloud Spending in India to Surpass $17 Billion in 2026 — Gartner, June 2026 — 28.1% growth to $17.5bn; IaaS the fastest-growing segment at 40%.
Note on model recommendations: guidance on model families reflects established practice rather than a single cited benchmark. Relative performance always depends on your dataset, and the only reliable comparison is one run on your own data with a held-out test set.
Last reviewed: 4 September 2026.