A facilities management client presented their new predictive maintenance dashboard to us. They said it was AI-powered. The system employed sensors on chillers together with a modern interface to issue alerts prior to equipment failure. We inquired how the software decided that a failure was imminent. After investigation, we discovered that the system depended on a spreadsheet of thresholds that had been created three years earlier. When a vibration reading surpassed a particular value, the system generated a ticket. This process can be useful at times, yet the software does not learn. Despite the sales presentation's claims, it is not artificial intelligence.
Artificial intelligence refers to software that enhances its performance by learning from data rather than by following pre-written instructions. This trait separates it from rules-based engines that are advertised as AI. These systems detect patterns in datasets that are too large for human analysis, yet they find correlations instead of causal relationships.
Artificial intelligence is software that boosts its performance on a task by learning from data. It does not operate according to pre-written instructions. This difference between learning and writing forms the core of its definition. All other categories, technical layers, and public interest derive from this single concept.
Shipshape Data creates data and AI systems. The most common issue we see is not technical, but concerns definitions. Clients frequently get contradictory information from various vendors, with the term "AI" applied to a variety of technologies such as lookup tables, spreadsheet macros, trained models, and chatbots, sometimes all appearing in the same presentation. This guide clarifies what AI means, how machine learning and deep learning fit inside it, the present capabilities and limits of these systems, the primary business application types, and how to tell a genuine AI use case apart from a rules-based system portrayed as more advanced.
What we mean when we use the term "AI"
Artificial intelligence is the broad field that includes any system performing tasks we would normally say need intelligence, such as playing chess, reading an invoice, or recommending a film. That field has a long history. The term was coined in 1956 during a summer workshop at Dartmouth, and for most of its history "AI" referred to hand-written rules: if this, then that, programmed by a person who had deeply considered the domain. In the 1980s, expert systems diagnosed illnesses using this approach. Many fraud-detection systems still operate this way today, and that is acceptable as long as no one labels it as learning.
The bar keeps moving
Professionals in the field often joke that artificial intelligence describes only technology that has not yet been achieved. Optical character recognition was regarded as AI until the technology became reliable, after which it was named OCR. Chess engines were seen as AI until Deep Blue beat Kasparov in 1997. Then, critics described the system as a brute-force search rather than genuine intelligence. This pattern recurs often. When a capability is released and becomes widespread, people reclassify it as ordinary software. This distinction matters because a definition ought to explain how a mechanism works rather than emphasise its perceived novelty.
One term refers to multiple distinct concepts
When people use the term AI, they might be referring to a hand-coded rules engine, a statistical model that finds patterns in historical data, or a large neural network that generates text and images. All three categories are called AI in business contexts. Only the second and third types truly learn. These two types learn via different methods, which explains how the categories are nested inside each other.
How machine learning and deep learning sit inside it
Picture three concentric circles. Artificial intelligence forms the outer circle. It encompasses any technology that carries out tasks that seem intelligent. Machine learning occupies the middle circle. It is a subset of artificial intelligence in which the system learns rules from data rather than following rules written by a person. Deep learning makes up the inner circle. It is a subset of machine learning that employs neural networks with many layers to carry out that learning. All deep learning systems are machine learning, and all machine learning systems are artificial intelligence. The reverse of these statements does not hold, although marketing materials often overlook these distinctions.
Machine learning: rules discovered, not written
A traditional program consists of logic that a person converts into code. A machine learning model works differently. You give it examples of inputs matched with the correct outputs, and the model determines the relationship between them. If you supply ten thousand past loan applications and their outcomes, the model pinpoints which factors predicted default. It weights those factors according to what the data supports rather than on human assumptions. The model discovers a pattern, while a rules engine repeats a pattern supplied by a person.
Most machine learning systems currently employed in business are not high-profile models. They include gradient-boosted trees for credit risk scoring, regression models for forecasting demand, and classifiers for sorting support tickets. These tools are reliable and often suitable for their tasks. They need less data and computing power than more complex models, and they reveal which specific features influenced a result.
Deep learning: the same idea, with more layers
Deep learning is machine learning constructed from neural networks with many stacked layers, each transforming what preceded it. What matters here is what those extra layers provide: the ability to learn directly from raw, messy, unstructured input such as an image, a sentence, or a sound wave, by converting it into the type of numerical representation described in our guide to embeddings. No person initially decides which features matter. That is genuinely a new capability. That is also why deep learning requires far more data and far more computing power than a gradient-boosted tree scoring the same loan applications, and why it is often the wrong tool for a problem that a simpler model would solve just as well, at a fraction of the cost and with an answer you can explain to a regulator.
The capabilities of these systems
Modern artificial intelligence works well at two specific tasks. It finds patterns within datasets that are too large for human memory to handle. It also processes human language and images with a proficiency level that resembles human capability.
Pattern recognition at volume
A model can compare a new transaction to millions of past examples and spot an anomaly within milliseconds. It carries out this task continuously without fatigue. It can detect a subtle correlation between three sensor readings and a failure that occurs eighteen days later, a relationship an engineer would not usually check manually. This is the reliable aspect of the field. It is statistics performed at a scale that surpasses human capability. The process involves no judgement, and this scale alone is valuable.
Language, suddenly, generally
The newer capability involves large language models handling open-ended text: drafting, summarising, answering questions about a document, and holding something akin to a conversation. Five years ago this required a narrow model built for a single job. Today one model performs a passable job at most of them, which is the biggest shift the field has seen in a decade, and it is why "AI" now means something different in a boardroom than it did in 2019.
Where these systems still fail
These processes do not amount to understanding. This difference is more significant than marketing materials suggest.
No sense of cause, only correlation
A model trained on ice cream sales and drowning incidents will observe that both metrics increase at the same time. It cannot identify that summer weather causes both trends. It is unable to ask why these events occur and can only determine "what usually goes with what". If the model encounters a novel situation that was not included in its training data, such as a new regulation, a unique product line, or a market shock, it will still generate a response. This response is a statistical guess presented as a fact. The model lacks an internal mechanism to recognise that the underlying conditions have changed.
Confidence is not an indicator of accuracy
A model cannot identify its own lack of knowledge. It will provide incorrect information using the same tone it uses for correct information.
This situation creates a practical danger. When a person is unsure, they hedge, ask a colleague, or flag the case for review. However, when a model is asked a question it cannot answer well, it still produces fluent and confident output. This occurs because the model was trained specifically to produce fluency. Large language models will invent a plausible fact instead of saying "I don't know". The invented fact appears just as convincing as a true one. This behaviour is a result of how the technology functions rather than a bug that will be fixed in a future release.
Data hunger and the messy middle
These systems depend entirely on the quality of their training data. Most businesses realise during a project that their data is incomplete, duplicated, or stored in conflicting systems. A large percentage of AI projects fail at this stage. The primary obstacle is usually the underlying data work rather than the model development.
This sequence occurs frequently. A team selects a use case, purchases a licence, and hires a specialist. They then find that customer records are stored in four different systems with inconsistent spellings, or that sensor data is deleted every ninety days instead of being logged. These are data infrastructure issues rather than artificial intelligence problems. They typically become apparent in the middle of a project that has already been announced to the organisation.
The primary tasks for which businesses use AI
Most enterprise AI projects fall into a few specific categories when the marketing language is removed. It is useful to identify these categories by their actual functions rather than by the terms used by vendors.
Prediction and forecasting
Common examples include demand for the next quarter, the remaining useful life of a machine, and the probability of customer churn. This category is the most established and reliable. In these cases, a simple and well-understood model typically performs better than a complex one.
Classification and tagging
Categorising incoming items-determining if a transaction is fraudulent, assigning a support ticket to the correct department, or identifying defects in an image-is a common task. This process is quick and lends itself to automation. Occasional errors are acceptable, provided a person reviews difficult cases.
Generation
Creating new text, images, code, or summaries instead of organising existing information is the most recent and rapidly developing area. This is where "it sounds right" and "it is right" can differ most significantly, because these models are specifically designed to generate fluent prose regardless of the content's accuracy.
Retrieval and conversation
Systems that let a person ask a question in plain language and get an answer pulled from the business's own documents, rather than from the model's general training, a pattern our guide to retrieval-augmented generation covers in more depth. Genuinely useful, and only as trustworthy as the underlying retrieval, which is why a poorly indexed document store produces a confidently wrong answer just as easily as a well-indexed one produces a right one.
Optimisation and automation
Routing, scheduling, pricing, and workflow decisions are made continuously without requiring a person to check a dashboard. This category is often the least visible, but in our experience, it is one of the most valuable. It eliminates the manual step completely; it does not simply advise a human who then has to take action. A truck route recalculated each morning by a person using a whiteboard does not scale effectively beyond a certain fleet size. The same decision made by a model that has learned from six months of delivery times does scale, and this process happens without anyone noticing.
Most actual deployments combine two or three of these categories rather than selecting just one. A support system might perform three distinct tasks in a single interaction by classifying a ticket, retrieving a policy document, and generating a draft reply. You must identify which component is performing each task when an error occurs. The method for fixing a classification error differs from the method for fixing a retrieval error.
The rules engine wearing an AI badge
We will return to the maintenance dashboard. We cannot provide a specific number of products marketed as "AI-powered" that do not use machine learning. We have not conducted a formal count, and we do not consider other existing estimates to be reliable. However, we encounter conventional software labelled as AI frequently. Because of this, we now ask vendors about their training data during every initial call before we discuss pricing.
Rules engines are useful tools. If-then logic is transparent, inexpensive to build, and easy to audit. This approach works well for problems with clear, stable rules that a person can document. Problems arise when these engines are marketed inaccurately. A rules engine cannot improve as conditions change, and it will not detect new failure patterns that were not programmed into it; it performs exactly as instructed. Calling this process "learning" creates expectations that the product cannot fulfil.
How to tell the difference before you commit budget
Asking a few direct questions clarifies the situation more quickly than reviewing vendor presentations.
Enquire about the data used for training
A legitimate machine learning product can describe the type and volume of data used to train it. If the provider states "it follows the logic our engineers configured", the product is based on rules. This may be appropriate for your requirements if the price and marketing reflect that reality.
Ask what occurs if the pattern changes
A model retrains, or is built to be retrained, as data changes. A rules engine requires a person to open the configuration and manually adjust a threshold. Both methods are valid, but only the model learns.
Request to view examples of errors
Every trained model generates errors since its outputs rely on probability. If a vendor is unable to give an example of a system failure and describe how it was detected, they are probably hiding information or have not performed enough testing with real-world data. This transparency becomes more critical when the system affects major outcomes such as loan approvals, medical diagnoses, or employment decisions. You ought to demand this information before signing a contract instead of waiting until after the purchase.
Where to start
Pinpoint the problem before choosing a technology. Record the particular decision you aim to improve and set the criteria for a correct outcome. Check whether you have sufficient historical examples of successful decisions that a model can use as training data. When the logic is simple, consistent, and well understood, a standard rules engine will outperform a model. Data science provides the most value when a genuine pattern exists in the data that has not yet been discovered.
In any case, the foundation underneath matters more than the label on top: clean, connected, well-understood data first, model second. If you are trying to determine which category your problem belongs to, or you have been sold something as AI and are unsure it merits the name, talk to us.
Do you need to find out how this impacts your organisation? Our AI strategy service assists you in identifying a profitable first use case and the order for later projects.
Frequently asked questions
What are some examples of artificial intelligence?
Practical examples comprise gradient-boosted trees for credit risk scoring, regression models for demand forecasting, and classifiers for sorting support tickets. These models remain reliable even when they are not prominent. Large language models that draft, summarise, and answer questions represent more visible recent examples. Rules engines, like the maintenance dashboard referenced at the start of this guide, are often labelled AI in business contexts despite not learning from data.
What is artificial general intelligence?
This guide defines a system that can perform any intellectual task a human can. Such a system would reason and adapt to new situations instead of focusing on a single task. Current AI detects correlations rather than causes and produces confident guesses when faced with unfamiliar information. This guide emphasises the distinction between fluent output and genuine understanding.
What is AI chat and how reliable is it?
AI chat is classified in this guide as retrieval and conversation. These systems let users ask questions in plain language and get answers drawn from a company's specific documents instead of the model's general training data. The technology is useful, yet its reliability hinges on the quality of the retrieval process. A poorly indexed document store will produce incorrect answers with the same confidence level that a well-indexed store produces correct ones. The confidence of a response does not guarantee that the information is accurate.
How does AI generate images and other content?
This guide places that under "generation": creating new text, images, code, or summaries, not sorting existing items, and labels it the newest, fastest-moving category of AI work. It is also the category where "it sounds right" and "it is right" diverge most dangerously, because fluent, polished output is precisely what these systems are built to produce regardless of whether the underlying content is accurate. That makes review and verification more important here than in most other areas of AI.
How can you tell if something described as AI is genuinely AI?
Enquire about the training data. A genuine machine learning product can describe the general data it used to learn. A system that "follows the logic our engineers configured" is a rules engine. This kind of system may meet your requirements, yet it does not learn. You should also enquire how the system reacts when patterns change. A real model retrains as the environment evolves, whereas a rules engine needs a person to update thresholds manually.