What is AI and how does it work?

That distinction is the whole thing. Everything else is detail.

Most explanations of AI either stop at the analogy or leap straight to matrix multiplication. This one sits in between, because the middle is where the useful decisions get made. If you are choosing whether to put an AI system into your business, you do not need the maths. You do need to know what the machine is actually doing when it answers, and where that process breaks.

How does artificial intelligence work, step by step?

Every practical AI system in commercial use today runs the same four-stage loop.

1. Training data is collected. Text, images, call recordings, transaction records — whatever the system is meant to learn from.
2. A model architecture is chosen. This is the shape of the network: how many layers, how they connect, how information flows through them.
3. The model is trained. It makes a prediction, the prediction is compared against the right answer, and the model's internal numbers are nudged to reduce the error. Repeat, billions of times.
4. The model is used for inference. Training stops. The frozen model now takes new inputs and produces outputs based on the patterns it settled on.

The nudging in stage three is called gradient descent, and it is the mechanism behind essentially all modern AI. There is no reasoning step being installed. There is a very large number of numerical parameters being adjusted until the outputs stop being wrong so often.

This is why AI systems fail in a specific way. They do not fail like software, with an error message. They fail like a confident colleague who has read a great deal and remembers it imperfectly.

Machine learning vs deep learning: what is the difference?

These terms get used interchangeably in marketing material and they are not the same thing.

Machine learning is the broad category: any system that improves at a task by processing data rather than by being explicitly programmed. A spam filter that learns from which emails you mark as junk is machine learning. So is a credit-scoring model built on a spreadsheet of past defaults.

Deep learning is one branch of machine learning that uses neural networks with many layers — "deep" refers to the layer count, nothing more. It is what made the last decade of AI possible, because deep networks can learn useful features from raw data without a human specifying what to look for.

The practical difference is data hunger and interpretability. A classical machine learning model might work on a few thousand rows and let you see which variables drove a decision. A deep learning model typically needs orders of magnitude more data and gives you very little insight into its reasoning.

For most UK SMEs this matters commercially. If you need to explain a decision to a regulator, a customer or an insurer, a simpler model you can interrogate often beats a deep one you cannot.

What are neural networks actually doing?

A neural network is a stack of very simple mathematical operations arranged in layers.

Each artificial neuron receives numbers, multiplies each by a weight, adds them up, and passes the result through a function that decides how strongly to fire. Stack enough of these and the network can represent extremely complex relationships between input and output.

The weights are the learned part. A trained model is, physically, a large file of numbers. Nothing else. There is no stored database of answers and no lookup table — the "knowledge" is distributed across those weights in a form no one can read directly.

This is worth sitting with, because it explains two things that confuse people. It explains why a model can be confidently wrong: the weights encode statistical tendencies, not verified facts. And it explains why models cannot reliably cite their sources unless a system has been deliberately built around them to fetch and quote real documents.

The transformer, and why 2017 was the turning point

The architecture behind every current large language model comes from a paper called Attention Is All You Need, published by researchers at Google and collaborators in June 2017.

Its contribution was the attention mechanism: a way for a model to weigh the relevance of every part of its input against every other part, in parallel, rather than reading strictly left to right. That parallelism is what made training on internet-scale text economically possible.

Almost everything sold as "AI" in a business context in 2026 is a transformer, a wrapper around a transformer, or a classical model doing a job a transformer would be wasteful for.

What is a large language model?

A large language model is a transformer trained on vast quantities of text to predict the next fragment of text given everything before it. Every capability it appears to have — answering, summarising, translating, writing code — is a consequence of doing that one prediction task extremely well.

That is not a reduction. It is the load-bearing fact about the technology.

When a model answers a question, it is not retrieving an answer. It is generating the sequence of words that its training makes most probable in context. When the probable words happen to be true, you get a correct answer. When they do not, you get a fluent, plausible, wrong one — the failure mode usually called hallucination.

One point often missed: frontier labs no longer publish parameter counts for their flagship models. OpenAI's last reliably disclosed figure was GPT-3's 175 billion, in 2020. Any specific number quoted for a current model is leaked, inferred or invented. Model size has stopped being a meaningful public metric, and anyone quoting one as a selling point is not describing something they can verify.

Narrow AI vs generative AI

The useful split for a business is not "narrow versus general". General AI does not exist. The split that matters is between systems that classify and systems that generate.

Most failed automation projects we audit fail here. A generative model gets deployed on a job that needed a predictive one, or a deterministic rule. It produces something impressive in the demo and unreliable at volume, because the task had a right answer and the tool was never built to guarantee one.

Generative AI examples that hold up in practice

Cutting through the demos, four categories currently earn their cost in UK small and mid-sized businesses.

• Conversation over your own material. A model connected to your documentation, answering customer or staff questions with citations. The technique is retrieval-augmented generation: fetch the relevant real passages first, then let the model phrase the answer from them. This is the single most reliable pattern, because it removes the model's need to recall facts.
• Voice agents on inbound calls. Speech-to-text, a language model deciding what to do, text-to-speech back. Genuinely effective on structured, repetitive calls — bookings, triage, out-of-hours capture. See how this compares with a human option in our AI receptionist vs virtual assistant guide.
• Unstructured-to-structured extraction. Pulling fields out of invoices, CVs, emails and forms. Unglamorous and the most consistently profitable use we see — a pattern that suits recruitment agencies and accountants especially well.
• First-draft generation with human editing. Content, proposals, responses. Value comes from the editing step, not the generation step.

What is not on that list matters as much. Custom model training and in-house model hosting are almost never justified for an SME in 2026. The economics do not work below a scale most businesses never reach, and the capability gap against a hosted frontier model is not closing in your favour. Choosing the right platform to build on matters too — see our comparison of Zapier vs n8n for how that decision plays out in practice.

Where AI actually breaks in a business

Four failure modes, in the order we encounter them.

It confidently invents things. Any system giving customers factual answers needs grounding in real retrieved documents, and a defined behaviour for "I don't know".

It inherits whatever was in the data. A model trained on past decisions reproduces the patterns in those decisions, including the ones you would not defend. Under UK GDPR, automated decisions with significant effects on individuals carry specific obligations, and "the model decided" is not a position you can hold.

Nobody agreed what good looks like. Most AI projects have no acceptance test. Without one there is no way to know whether the thing works, and it quietly degrades.

The process underneath was broken. This is the big one. Automating a bad process makes it fail faster and at greater volume. Most businesses do not have an AI problem — they have an operations problem, and AI applied to it just makes the existing mess move quicker. This is exactly the trap we see in trades and dental practices that automate a broken booking process rather than fixing it first.

Start with the audit, not the software

Understanding how AI works is useful mainly because it lets you tell the difference between a tool that fits your problem and a tool being sold to you.

The order that works: map how work actually moves through the business, find where time and revenue leak, then decide whether the fix is a generative model, a predictive one, a straightforward integration, or a change to the process that needs no AI whatsoever. Often it is the last one.

Thirty minutes on a call. We map how work moves through your business, show you where the time is going, and tell you what is worth building — and what is not. You keep the roadmap either way. See our pricing or book a free automation audit to start.

What is AI and how does it work? FAQ

Sources and links

• Attention Is All You Need, Vaswani et al., arXiv:1706.03762, 12 June 2017 — arxiv.org/abs/1706.03762
• UK GDPR, automated decision-making and profiling, ICO guidance — ico.org.uk
• Internal: Zapier vs n8n guide · AI automation cost for UK small business · Free AI automation audit