The situation
A home-delivery retailer has a governance review coming, and somebody has been asked to produce an inventory of everything the business currently calls AI. Five systems are in flight.
The first allocates delivery slots. It is a rules engine with about two hundred written conditions covering vehicle capacity, postcode, and cut-off times. A vendor sold it as AI-powered. The second predicts which customers will cancel in the next ninety days. It is an XGBoost model trained on tabular account data in Amazon SageMaker AI: thirty-odd columns of tenure, spend, complaint counts and delivery failures. The third scores the photograph a driver takes at the doorstep and flags the ones too dark or too blurred to prove delivery. It is a convolutional neural network trained on a few hundred thousand labelled images. The fourth condenses long support conversations into a paragraph a team leader can read, using a model on Amazon Bedrock. The fifth reads a failed-delivery note, works out what happened, calls internal tools to check stock and van capacity, and books a redelivery without a human touching it. It runs on Amazon Bedrock AgentCore.
The label each one gets is not paperwork. It routes the system to a reviewer, sets what data the team has to produce, and decides what kind of explanation the business can give a customer who asks why a decision went the way it did. Getting the five labels right is the first task, and the words involved overlap enough that plenty of inventories get it wrong.
What actually matters
Start with how the terms nest, because most of the confusion comes from treating them as five competing options rather than four sets inside one another plus a behaviour bolted on the end. Artificial intelligence (AI) is the outer set: any system that performs tasks normally associated with human intelligence, whether it learns anything or not. Machine learning (ML) is the subset of AI where the system works out its behaviour from data instead of being told the behaviour by a programmer. That is the dividing line worth memorising. A person writing conditions by hand produces software that can be very sophisticated and still learns nothing. An ML system is shown examples, derives a mapping from input to output, and applies that mapping to inputs nobody has seen.
Deep learning is the subset of ML that uses neural networks with many layers. What changes at this level is where the features come from. Classic ML wants somebody to decide, in advance, which columns matter: hours since last delivery, complaints per quarter, spend trend. Deep learning learns those intermediate representations from raw input, layer by layer, which is why it took over the problem domains where nobody could write the features down. Computer vision, the field concerned with getting meaning out of images and video, was one. Natural language processing (NLP), the field concerned with getting meaning out of text and speech, was another. The cost is data volume and opacity: many layers of learned weights need far more examples, and the resulting behaviour is much harder to attribute to any one input.
Generative AI (GenAI) is a subset of deep learning whose models produce new content rather than a label or a number. A classifier answers “which of these categories”; a generative model answers with text, an image, audio or code that did not exist before. A large language model (LLM) is the text-and-code variety, trained on very large text corpora to predict what comes next, and the next-token mechanism underneath it explains most of its strengths and all of its hallucinations. Agentic AI is the step past that: a generative model given tools it can call, memory that survives between steps, and the autonomy to decide a sequence of actions rather than answering once. The model at the centre of an agent is the same kind of model as the one doing summarising. What differs is that its output reaches external systems and changes them.
Four pieces of vocabulary run through all of this and get used loosely. An algorithm is the procedure, the recipe: gradient boosting, k-means, backpropagation. A model is the artefact you get by running that algorithm over data, the learned weights and structure you then deploy. Training and inferencing are the two phases of a model’s life: training is the expensive, offline process of learning from examples, and inferencing is what happens every time the trained model is asked for an answer. Fit describes how well the learned mapping generalises. A model that has memorised its training data and stumbles on anything new is overfitting; one that never learned enough structure to be useful on either is underfitting. And bias, in this setting, means a systematic skew that pushes results consistently in one direction, usually because the training data was not representative of the people the system will be used on. Fairness is the related question of whether the outcomes are equitable across the groups affected. The two are not the same word: bias is a property you can measure in data or in a model, fairness is a judgement about consequences.
What we’ll filter on
- Does it learn its behaviour from data, or follow rules a person wrote down?
- Are the features hand-engineered by a human, or learned by the model from raw input?
- Is the output a label or a number, or is it new content?
- Does it act autonomously against external systems, or only return an answer?
- How much labelled data of your own does it need before it works?
The landscape
Five categories cover everything in this inventory, and they line up as a staircase rather than a menu.
Rules-based automation
Written conditions, decision tables, constraint solvers, an if-then engine with a few hundred branches. No training data, no learned parameters, entirely deterministic. It is often called AI in marketing copy, and it belongs outside the set: nothing about it learns. Its saving grace is that it explains itself perfectly. Every decision traces to a rule with an author and a date.
Classic machine learning
Regression, classification and clustering over structured data, usually tabular. Gradient-boosted trees, linear models, random forests. A human chooses the features, an algorithm learns weights or splits from labelled examples, and the trained model outputs a number or a category. Supervised approaches here need a labelled dataset you have to build and maintain yourself, which is nearly always the expensive part. Explainability is comparatively good: feature importances and per-prediction attributions are routine.
Deep learning
Multi-layer neural networks, learning features rather than being handed them. This is what makes computer vision and modern NLP work at all, and the statistical methods that ran language processing before it are a useful reminder that deep learning is not the only way to do these jobs, only the way that scaled. Deep learning needs a lot of labelled data and a lot of compute, and it gives back much weaker explanations than a tree ensemble does.
Generative AI
Deep learning models that produce content. LLMs for text and code, diffusion models for images, and the broader field of multimodal and reasoning models beyond both. The practical difference from everything above is that you consume a pre-trained foundation model instead of building one, so the labelled-dataset burden mostly disappears and is replaced by prompting, retrieval and evaluation work. Output is open-ended, so correctness is judged rather than scored against a known answer.
Agentic AI
A generative model plus tools, memory and autonomy. The model is given a goal, decides which tools to call in which order, holds state across the steps, and stops when it judges the goal met. Everything true of generative AI is still true here, with one addition that dominates the review: the system takes actions in the world. A wrong summary is a bad paragraph. A wrong tool call books a van.
Evaluation
Side by side
| Category | Learns from data | Features are learned | Output is new content | Acts on external systems | Needs your own labelled data |
|---|---|---|---|---|---|
| Rules-based automation | ✗ | ✗ | ✗ | ✗ | ✗ |
| Classic machine learning | ✓ | ✗ | ✗ | ✗ | ✓ |
| Deep learning | ✓ | ✓ | ✗ | ✗ | ✓ |
| Generative AI | ✓ | ✓ | ✓ | ✗ | ✗ |
| Agentic AI | ✓ | ✓ | ✓ | ✓ | ✗ |
Read the first four columns as a staircase: each row adds one property to the row above and keeps everything already there. Agentic AI does not replace generative AI, it wraps it, in the same way deep learning does not replace machine learning. The fifth column breaks the pattern deliberately, because the labelled-data burden peaks in the middle. Rules need none since nothing is learned; classic ML and deep learning need a dataset you build and label; generative and agentic systems inherit a model somebody else trained, so your effort moves to prompts, retrieval, guardrails and evaluation instead.
The solution
The slot allocator is not AI. Two hundred hand-written conditions are ordinary software, and calling it AI in the inventory would pull it into a review designed for systems whose behaviour nobody wrote down. Its explanation of any decision is the rule that fired, which is stronger than anything the other four can offer. Leave it in the inventory as a system the review considered and excluded, with the reason recorded.
The churn predictor is classic supervised machine learning: an algorithm learning a classifier from labelled tabular examples, producing a model that does inferencing on new accounts. Its review needs the training dataset documented, because that is where bias enters. If cancellations in the training window came disproportionately from one region or one delivery round, the model will carry that skew forward, and fairness questions about who gets a retention offer follow directly. Fit is a live concern too: thirty features and a few hundred thousand rows is comfortably enough capacity to overfit, and the review should ask for validation numbers rather than training numbers.
The doorstep-photo checker is deep learning applied to computer vision. It sits inside machine learning, so everything said about the churn model still applies, plus two things that come with the layers. Its features are learned rather than chosen, so nobody can hand the reviewer a list of the properties it keys on. And the labelled dataset behind it is orders of magnitude larger, which makes it correspondingly harder to audit for what it under-represents. Photographs taken in poor light on dark-painted doors are the sort of blind spot that only shows up when somebody goes looking.
The support-conversation summariser is generative AI, an LLM doing natural language processing on inbound text. Its review looks nothing like the first two. There is no training dataset of yours to inspect, so questions about the training corpus go to the model provider. Your side of the work is prompt design, evaluation against real conversations, and guardrails on what the summary may contain. Customer names and payment details arriving in the prompt are the exposure worth naming.
The redelivery workflow is agentic AI. Everything said about the summariser applies to the model at its centre, and then the review adds the part unique to agents. It calls tools, so each tool needs a permission boundary and a reviewer who has asked what the worst call it could make would cost. It holds memory across steps, so state from one customer must not leak into another’s session. And it takes actions autonomously, so somebody has to decide which actions need a human in the loop, what the limits are on the ones that do not, and how a wrong booking gets reversed. That is a different review from the other four, which is why the label was worth arguing about.
Worked example
The one the vendor called AI
The slot allocator arrived with AI-powered on the invoice, and the first draft of the inventory copied that across. Run it through the filters and the first one settles it: the behaviour came from a person writing conditions, not from data. Nothing is trained, so there is no model, no training and inferencing split, and no fit to worry about. Its real risk is a rule someone wrote in 2019 that nobody has revisited. Labelling it plainly sends that to a code and change-control review, rather than to a data review it has no use for.
The one somebody wanted to call an agent
The summariser was proposed as agentic AI, on the grounds that it uses a foundation model and runs without a human pressing a button. Filter four decides it. The summariser is handed text and returns text; it calls nothing and changes nothing outside its own response. Automatic is not autonomous. Label it generative AI and its review stays proportionate: content controls and evaluation, without the tool-permission work the redelivery workflow genuinely needs.
What’s worth remembering
- AI is the outer set, machine learning is the part that learns from data, deep learning is the part of machine learning using multi-layer neural networks, and generative AI is the part of deep learning that produces new content.
- The line between AI and ordinary software is whether behaviour was learned from data or written down by a person, and a rules engine sits on the software side no matter what the vendor calls it.
- Deep learning learns its own features instead of being given them, which is what made computer vision and modern natural language processing work, at the price of far more data and far weaker explanations.
- Agentic AI is a generative model plus tools, memory and autonomy, so it is judged on the actions it can take rather than only on the text it returns.
- An algorithm is the procedure and a model is the artefact it produces; training and inferencing are that model’s two phases, and fit describes how well what it learned holds up on data it has never seen.
- Bias is a measurable skew in data or in a model, fairness is the judgement about whether the resulting outcomes are equitable, and only the systems that learn from your data raise the first as your problem.