AI terms, plainly defined.

An AI system that doesn't just answer — it acts. An agentic system can decide what to do next, call tools, write to a database, send a message, and then react to what happened. Contrast with a chatbot, which only produces text.

In practice, an agentic system has three parts: a reasoning model (usually a large language model), a set of tools it's allowed to call (APIs, databases, messaging apps), and a policy that decides when to stop. Faction AI's Faith OS is an agentic platform — eleven specialist agents that operate inside a configurable budget and audit log.

Agentic AI is harder to deploy safely than a chatbot because mistakes can change the world rather than just embarrass you. The mitigation is bounded autonomy: scope, budget, human-in-the-loop on irreversible actions.

A numerical fingerprint of a piece of text (or an image, or audio) that captures its meaning. Two pieces of text with similar meanings get embeddings that are mathematically close to each other; two with different meanings end up far apart.

Embeddings are the engine room of search and retrieval inside AI systems. When you ask Faith OS "what did we agree with the client about scope in March?", she doesn't grep the chat history for the word "scope" — she embeds your question, embeds every past message, and returns the messages whose meaning is closest.

Modern embedding models (OpenAI text-embedding-3-large, Cohere embed-v3, open-weights alternatives) produce vectors of 1,024–3,072 numbers per chunk. Costs are negligible. Quality varies enough that swapping the embedding model often beats prompt engineering.

Tests for AI systems. Like unit tests for software, evals run a fixed set of inputs through your model or prompt and score the outputs against a known-good answer. Without evals you have no way to tell whether yesterday's prompt change made the system better or worse.

Good evals are specific: "does the lead qualification agent correctly route this enquiry from a UK property surveyor about commercial reinstatement cost?" — not "is the answer good?". They run automatically on every change. They fail loudly when broken.

Most production AI failure modes — drift, regression after a model update, prompt rot — are invisible without evals. Any AI partner who can't show you their eval suite is shipping on vibes.

Taking a general-purpose AI model and training it further on your specific data so it gets better at your specific task. The base model already knows English and the world; fine-tuning teaches it your tone of voice, your domain jargon, or a narrow task you do repeatedly.

Fine-tuning used to be the default move for any custom AI use case. It mostly isn't any more. Frontier models with good prompting and retrieval cover ~90% of what fine-tuning used to be needed for, without the cost, the lock-in to a stale model snapshot, or the data-handling complexity.

Where fine-tuning still earns its keep: narrow classifiers running at high volume (cheaper than prompting a frontier model every time), small open-weights models you need to run on-prem, and replicating a very specific writing voice.

The discipline of getting your website cited by AI search engines — ChatGPT, Perplexity, Claude, Google AI Overviews — rather than (or as well as) ranking in traditional Google results. Same instinct as SEO; different signals.

AI engines don't just rank pages, they cite specific passages. Optimising for that means: machine-readable structured data (JSON-LD), clear authorship and dates, short citable chunks (140–160 words), strong external authority signals, and a robots.txt that explicitly welcomes AI crawlers.

GEO doesn't replace SEO — Google still drives most search traffic and AI engines learn partly from what ranks. The two stack. Most sites that win at GEO also rank better in classical search, because the same hygiene helps both.

The category of AI behind ChatGPT, Claude, Gemini, and most modern automation. An LLM is a statistical model trained to predict the next token (roughly: word fragment) given the tokens before it. "Large" just means many billions of parameters.

What makes LLMs useful for business automation isn't the prediction trick itself — it's that scaling that trick produces emergent behaviours: instruction following, reasoning, code generation, translation, summarisation, tool use. The same model can answer a customer email, classify a lead, summarise a contract, and write a report.

LLMs are still wrong sometimes ("hallucinations"), expensive at scale, and constrained by their context window. The art of deploying them is matching the right model to the right task, grounding them in your data via retrieval, and adding evals + guardrails so failures are caught.

An AI architecture where several specialist agents — each with its own scope, tools, and prompt — collaborate on a task instead of one monolithic AI trying to do everything. A dispatcher routes work to the right specialist; specialists hand back results.

The payoff is sharper behaviour. A single agent told to "be good at research, content, finance, and QA" is mediocre at all four. Three specialist agents (research / content / QA) each tuned for their lane outperform the generalist on their lane and stay maintainable.

Multi-agent introduces orchestration cost: who decides what, how they share memory, how you debug a failure that crossed three agents. Faction AI's Faith OS is a worked example — eleven specialist agents coordinated by a dispatcher ("Faith" herself) with shared persistent memory and a unified audit log.

Designing the instructions you give an AI model to make it produce the output you want, reliably. Not a job title — a craft skill that every team deploying AI needs at least one person to be good at.

Good prompts are specific (state the task, the audience, the format, the failure modes to avoid), grounded (include the relevant facts in the prompt rather than hoping the model knows them), and tested (you don't ship a prompt you haven't run on 20 representative inputs).

Bad prompts are vague ("write me something good about this"), assume the model can see what you can see ("as we discussed"), or rely on the model knowing private facts. Prompt engineering matters less than it did two years ago — frontier models are more forgiving — but it still separates AI systems that work in production from ones that almost work in a demo.

A pattern for grounding an LLM in your private data without retraining it. Instead of fine-tuning a model on your knowledge base, you store your data in a searchable form (usually as embeddings in a vector database), retrieve the relevant chunks at query time, and feed those chunks to the model alongside the user's question.

RAG is how most production AI assistants work. "What's our refund policy?" → embed the question → find the 4 most relevant paragraphs in the policy doc → ask the model to answer using only those paragraphs. The model stays general-purpose; the answers stay grounded; the knowledge base stays editable.

The hard parts of RAG aren't the embeddings — they're chunking (how you split your source documents), retrieval quality (whether the right chunks come back), and grounding (whether the model actually uses the retrieved context instead of making something up).

Connecting business systems so a sequence of steps runs without a human pushing each button. The fundamentals predate AI by twenty years — Zapier, Make, n8n, Power Automate, custom scripts. Adding AI just makes the pipes smarter at the decision points.

Classical automation: "when a form is submitted, create a Trello card, send a Slack message, add the contact to Mailchimp." AI-enhanced automation: the same sequence, but the LLM in the middle reads the form free-text, decides which department it's for, drafts a personalised reply, and only routes ambiguous cases to a human.

For most UK SMEs, workflow automation is where AI pays for itself fastest. Invoicing, onboarding, lead intake, weekly reporting, content scheduling — five hours a week saved across five processes is a person-day a week back.