The three layers compared
| Dimension | Artificial intelligence | Machine learning | Deep learning |
|---|---|---|---|
| Definition | Systems performing tasks that usually need human judgement. | Systems that improve from data rather than explicit rules. | Machine learning using multi-layer neural networks. |
| Typical tasks | Rules, search, planning, expert systems. | Forecasting, scoring, classification, segmentation. | Language, vision, speech, large-scale sequence work. |
| Data requirement | Domain rules instead of training data. | Labelled or structured history at a usable grain. | Large volumes, often unstructured, with careful splitting. |
| Skills | Process and domain expertise. | Statistics, feature engineering, software discipline. | Model architecture, distributed compute, evaluation at scale. |
| Evaluation | Policy compliance and edge-case review. | Out-of-sample error against a baseline, by segment. | Distribution shift, input sensitivity and cost of evaluation. |
| Operating cost | Maintenance of rules as the business changes. | Retraining and monitoring on a defined cadence. | Compute, monitoring and specialised support. |
| Where it fits | Stable processes with clear rules. | Tabular decisions with available history. | Unstructured input where features cannot capture the pattern. |
The nesting, stated plainly
Artificial intelligence describes systems that perform tasks associated with human judgement, including rule-based ones that never learn. Machine learning is the part that improves from examples, from a simple regression to a gradient-boosted model. Deep learning is the part of machine learning that uses multi-layer neural networks, which is what most current language and image systems are built on.
- Rules and search can be intelligent without learning
- Learning from data is what makes a system machine learning
- Neural networks with many layers are what makes it deep learning
What each layer needs in practice
A rules-based system needs domain knowledge and maintenance. A classical model needs labelled or structured history, careful feature work and honest evaluation. A deep model needs substantially more data, more compute and more specialised skills, and it still needs the same evaluation discipline. Choosing a layer without checking those requirements is the most common cause of stalled projects.
Where deep learning earns its place
Deep learning is the right answer when the input is unstructured or the pattern is too complex to express as features: text, images, audio, and sequence problems at scale. For tabular business data with a few hundred informative features, a well-built classical model is often competitive, considerably cheaper to operate and far easier to explain to the people who act on it.
For tabular data, gradient-boosted trees such as XGBoost or LightGBM are a strong default baseline: they handle mixed feature types, need little preprocessing and train quickly. Published comparisons on tabular benchmarks have found tree-based models competitive with, and often ahead of, deep networks on this kind of data.
The layer changes how you evaluate
Evaluation differs by layer as much as the method does. A rules-based decision is checked against policy and edge cases; a classical model is judged on out-of-sample error against a baseline; a deep model adds distribution shift, prompt or input sensitivity, and often a much larger evaluation cost. The higher the layer, the more you are testing the data pipeline as much as the algorithm.
- Baseline first in every layer
- Error reported by segment, not as a single average
- Cost of evaluation included in the project estimate
Illustrative example · hypothetical organisation
Worked example: three requests, three layers
Consider three hypothetical requests that all arrive labelled as an AI project.
Routing invoices above a value threshold to a second approver is a rule. No learning is needed, and a rule is easy to audit.
Predicting which invoices will be paid late, from customer history, amount and payment terms, is machine learning on tabular data. A gradient-boosted model measured against a simple baseline is the starting point.
Reading scanned supplier invoices and extracting their fields is deep learning, because the input is unstructured. Most teams use a pre-trained document model or service here and spend their effort on evaluation and exception handling.
What this means when you buy or hire
A vendor or candidate using the three terms interchangeably is usually describing capability they have not scoped. Ask which layer the task needs, what data it assumes, how the result will be evaluated and who operates it afterwards. Those four questions separate a specific proposal from a general claim about AI.
A short decision sequence
Start with the decision and its cost of being wrong, then check whether explicit rules would work, then whether a simple model on existing data beats the current baseline, and only then whether deep learning is needed for unstructured input at scale. Each step is cheaper than the next, and most business problems are resolved by the second.
Questions and trade-offs
Question
Which layer do we need for our problem?
Start from the decision, not the technique. If explicit rules cover it, that is the cheapest system to own. If it depends on patterns in structured history, a classical model is usually enough. Reach for deep learning when the input is unstructured or the pattern is genuinely too complex to express as features — and budget for the data, compute and evaluation it needs.
Question
Is deep learning always more accurate?
No. On tabular business data with a moderate number of features, well-built classical models are often comparable and considerably cheaper to operate and explain. Deep learning pulls ahead where the input is language, images, audio or very large unstructured sequences — and it brings higher data, compute and monitoring requirements with it.
Question
Where do large language models fit?
Large language models are deep learning models trained on large volumes of text, so they sit in the innermost layer and bring its requirements: evaluation on your own tasks, attention to cost and latency, and controls on what data reaches them. For most business tasks they are used through an API, not trained in-house.
