Artificial intelligence (AI) is software that carries out tasks we usually associate with human thinking, such as understanding and writing language, recognising images, finding patterns and making predictions. Most AI in use today is built with machine learning: the system learns from large amounts of example data, and its behaviour comes from those patterns, with no one writing the rules by hand.
Google’s machine learning glossary puts it briefly: AI is a non-human program or model that can solve sophisticated tasks, and machine learning is a sub-field of it.
How AI, machine learning and generative AI relate
The terms are often used as if they meant the same thing. They are layers of one idea:
- Artificial intelligence is the whole field.
- Machine learning is the main way AI is built now: models trained on data.
- Neural networks are one family of machine learning models, loosely inspired by the brain, and the basis of most recent progress.
- Generative AI produces new content, such as text, images or audio. Large language models, which predict text one piece at a time, are the best known example.
- ChatGPT is a product: a chat application built on large language models.
Examples of AI you already use
AI has been part of everyday software for years, often without the label. Your email filters spam with it. Your bank uses it to flag unusual card payments. Online shops and streaming services use it to recommend products and films. Maps predict traffic, phones recognise faces in photos, and translation tools turn a web page into another language in a second. The chat assistants that arrived in recent years are the most visible kind, but not the only one.
How AI features are built into a website
On a business website or web application, AI is usually a feature inside a larger system. Take a builders’ merchant with 8,000 products online and a busy enquiry inbox. Three realistic features:
- Search by meaning. A customer types “something to stop my gutter leaking” and gets sealants, brackets and gutter joints, even though none of the product names contains those words.
- Enquiry sorting. Each message from the contact form is read on arrival, tagged as a trade quote, a delivery question or a complaint, and sent to the right person with a short summary.
- Draft product descriptions. Staff upload a supplier’s spec sheet, and the system drafts a description in the shop’s style for someone to check and publish.
Behind each feature, the pattern is similar:
- A model from an AI provider. Most businesses use models from providers such as OpenAI, Anthropic or Google, called through an API. Few need to train their own.
- Your data. The model is given the information it needs for the task: the product catalogue, the enquiry, the spec sheet. For search by meaning, product texts are converted into numerical representations called embeddings and stored so they can be compared.
- Rules around it. What the feature may and may not do, what happens when the model is unsure, what gets logged, and where a person checks the result.
- Measurement. Did the search find more products people bought? Did enquiries reach the right person faster? AI features are worth keeping when they move a number that matters.
On a Laravel site, much of this is now part of the framework’s own tooling. The Laravel AI SDK provides one interface for providers including OpenAI, Anthropic and Gemini, covering text generation, agents, embeddings, images and audio, and Laravel 13 added native vector search for PostgreSQL databases with pgvector. On other platforms the same features are built through each provider’s API, often a REST API.
What AI does well, and where it goes wrong
AI is good at work that is repetitive, language-heavy and tolerant of an occasional correction: sorting, summarising, drafting, extracting details from documents and finding things by meaning.
It is weaker where every answer has to be right. Language models can state wrong facts confidently, and they reflect the gaps and biases in the data they learnt from. They also cost money each time they run, so a feature that calls a model on every page view can become expensive as traffic grows. A sensible first AI feature is one where a person reviews the output or where an occasional mistake is cheap, such as a draft, a tag or a suggested search result.
AI and personal data
When an AI feature processes personal data, such as names in enquiries or customers’ order histories, UK data protection law applies to it as to any other processing. The ICO’s guidance on AI and data protection covers accountability, lawfulness, fairness, transparency, accuracy, security, data minimisation and individual rights. The ICO notes that the guidance is under review after changes made by the Data (Use and Access) Act.
In practice, three questions come first: which personal data the feature really needs, which provider processes it and under what contract, and how visitors are told that AI is involved. This is a general overview, not legal advice.
How we work with AI
Integrating AI into custom web applications is a core part of our Laravel development work. An AI feature goes into the project scope with the rest of the build, and the final price is confirmed after a short Discovery stage, with a line-item estimate.
FAQ
Can AI be added to an existing website?
Usually, yes. Most AI features connect to an AI provider through an API, so they can be added to a site that already exists, provided its platform lets developers add server-side code. A custom application gives the most freedom; on hosted platforms, features are often limited to what apps and plugins offer.
How much does it cost to add AI to a website?
There are two parts: the development work to build the feature, and the provider’s usage fees, which grow with how often the feature runs and how much text it handles. A feature used by staff a few dozen times a day costs far less to run than a chat assistant answering every visitor. Estimate the usage before choosing the design.
Is AI the same as automation?
No. Automation follows fixed rules: when a form is submitted, send this email. AI makes a judgement from patterns: this message looks like a complaint. Many useful systems combine both, with AI deciding what something is and ordinary automation doing the rest.
Will AI give the same answer every time?
Not necessarily. Generative models can phrase answers differently from one run to the next, and a change of model version by the provider can change results. That is why AI features are tested on a fixed set of real examples before launch and checked again when the model is updated.
