AI terms, in plain English

Software that uses a language model to decide the next step in a task and then takes it by calling a tool: look up a record, draft a reply, update a CRM, book a meeting. A chatbot answers questions. An agent goes and does the thing, within the permissions it has been given.

AI agent development

A model trained on large amounts of text to predict and generate language. GPT, Claude, Gemini and Llama are examples. On its own an LLM only produces text; connected to your data and tools it can answer, summarise, classify and act.

LLM & ChatGPT integration

A pattern where the system first searches your own content for the passages relevant to a question, then gives them to the model with the question. The model answers from those passages and can cite them. It's the standard way to make an assistant know your products, policies or documents without retraining a model.

Numeric representations of text where similar meanings sit close together. They power semantic search: a question about "cancelling my plan" finds the help article titled "How to end your subscription" even with no shared words.

A database built to store embeddings and find the nearest ones quickly. Pinecone, Qdrant and Weaviate are dedicated options; pgvector adds the same ability to PostgreSQL, which is often all a project needs.

Training an existing model further on your own examples so it picks up a style, format or narrow skill. It changes how a model behaves. It's a poor way to teach a model facts that change, which is what RAG is for.

A model feature where the LLM replies with a structured request to run one of a set of functions you define, with typed arguments, instead of plain text. Your code runs the function, checks permissions and returns the result. It's how agents take actions safely.

An open standard, introduced by Anthropic in 2024, for connecting AI applications to tools and data. An MCP server exposes a system's actions and content once, and any MCP-compatible client (Claude, ChatGPT, IDEs, your own agents) can use them.

A test set of real inputs with known good outcomes, run against an AI system before every change. Evals catch the case where a better prompt for one question quietly breaks ten others. They are the closest thing AI work has to unit tests.

The checks around a model that keep it inside agreed limits: input filtering, permission checks on every tool, output validation, PII redaction and rules for what it may say. Good guardrails live in code on the server, where the model can't talk its way past them.

When a model states something false with confidence, such as an invented policy, price or citation. It is reduced by grounding answers in retrieved sources, asking the model to say when it doesn't know, and testing with questions designed to tempt it.

A design where a person approves, edits or rejects an AI step before it takes effect. It's the safe way to launch agents that touch money, customers or anything irreversible. The approval steps are removed one by one as the logs show they're no longer needed.

An attack where text the model reads, such as an email, web page or document, contains instructions meant to hijack it. The defence is architectural: treat everything the model reads as data, limit what its tools can do, and require approval for sensitive actions.

The amount of text a model can consider at once, measured in tokens. Bigger windows let you pass whole documents, but cost and response time grow with every token, so retrieval that picks the right passages still matters.

A model whose weights are published so you can run it on your own servers, such as Llama, Qwen, Mistral or DeepSeek. It's the option when data must never leave your cloud account, at the price of running and scaling the model yourself.

Automating a multi-step business process, like lead routing or invoice intake, where some steps need judgement a fixed rule can't make. Classic tools such as n8n, Make or Zapier handle the predictable steps, and a language model handles the reading, classifying and drafting.

AI workflow automation

An AI agent that talks on the phone or in an app using realtime speech models. Modern voice agents respond in under a second, handle interruptions and can book, qualify or route a call before handing over to a person.

Chatbots & voice agents

An engineer who works inside the client's team, in their Slack, standups and repository, instead of behind a ticket queue. They learn the workflow first-hand, scope what's achievable, build it and stay accountable once it's live.

What is a forward deployed engineer?