Artificial intelligence: a practical field guide

The AI map

Artificial intelligence is the broad field of systems that perform perception, prediction, language or decision tasks. Machine learning learns patterns from examples rather than relying only on hand-written rules. Deep learning uses multi-layer neural networks and powers much of modern image, speech and language AI.

A model makes predictions or generates content. A platform makes models available through a product. An application combines a model with a task and data. An agent adds tools, state and a controlled loop of actions. Keeping these layers separate prevents most AI terminology confusion.

How AI learns

Training adjusts a model’s numerical parameters using text, images, audio, measurements or labelled examples. Supervised learning uses known answers; self-supervised learning finds patterns in the data itself; reinforcement learning uses feedback or reward. Inference is the later operational stage: a trained model receives a new input and returns an output.

Neural networks learn useful representations, but not a transparent rule book. They can generalise to new cases, yet cannot independently establish truth, permission or relevance to your organisation.

Generative AI and LLMs

Generative AI creates new text, images, audio, video, code or structured data. A large language model (LLM) predicts the next token of text. Most leading LLMs use transformers, whose attention mechanism weighs relationships among parts of the current input. Multimodal models apply similar ideas to text, images and audio.

LLMs generate plausible continuations, not guaranteed facts. A hallucination is a fluent but unsupported answer, invented detail or inappropriate pattern match. Grounding with approved information, tool results and human review reduces this risk but never removes the need for checks.

Core terminology

Platforms and model families

ChatGPT, Claude and Gemini are general chat platforms: strong for drafting, explanation, analysis and multimodal conversation, but their available models, tools, privacy controls and behaviour vary by plan and evolve quickly. Microsoft Copilot and similar work-suite assistants apply AI inside documents, mail, meetings and governed enterprise data. Open-model ecosystems such as Llama, Mistral and Hugging Face enable custom hosting and adaptation, but shift more security, evaluation and operational responsibility to the adopter.

Choose a model by a task evaluation—not a leaderboard. Compare representative inputs, quality, reliability, modalities, tool use, context, latency, cost and data controls. A smaller model is often the better production choice when it meets the requirement.

Agentic AI

An agentic system pursues a goal over multiple steps: read state, select an allowed tool, perform a bounded action, observe the result, then stop, continue or ask a person. It combines a model with instructions, trusted knowledge, state, tools, an orchestrator, permissions, logs and evaluation.

Use a single chat request for drafting; deterministic automation when the process is known; an agentic workflow when inputs vary but the action set is narrow. Start read-only. Require approval before payments, deletion, publication, external messages or production changes. Guard against wrong-tool selection, loops, prompt injection and data exposure with allow-lists, least privilege, limits and audit trails. MCP is a standard way applications can connect models to context and tools; it does not make a connected tool safe automatically.

Practical applications

• Research: retrieve from a curated source set, then verify consequential claims against the originals.
• Writing: draft, translate and adapt tone; a person checks facts, sensitive data and final wording.
• Software engineering: generate tests, explanations or refactors; run tests and review the change.
• Data analysis: translate questions into governed queries; validate filters, units, calculations and source data.
• Support: answer from current policy documents; constrain actions and provide human handoff.
• Operations: triage and prepare recommendations; keep high-impact action behind authorised approval.

Reliable, responsible use

Pair probabilistic models with deterministic controls. Verify critical facts and calculations. Classify information before upload and understand retention, training and access settings. Treat webpages, documents and email as untrusted data: they may contain prompt injection instructions. Use scoped accounts, minimal permissions, confirmation gates and logs for actions.

Evaluate representative users and edge cases for bias, failures and escalation quality. Monitor quality, cost, latency and changing behaviour. Review intellectual-property obligations and environmental cost. AI can accelerate work; accountability stays with the people and organisation deploying it.

Glossary

Alignment: behaviour consistent with intended goals and constraints. Benchmark: a standard test, useful but not task-specific proof. Multimodal: handling more than one data type. Model weights: learned numerical parameters. Structured output: a schema-constrained response, such as JSON, for safer software integration. Synthetic data: artificially generated data that still requires quality checks.

Choosing an approach

1. Define the user, outcome, acceptable error and consequence of failure.
2. Use conventional software when rules are stable and auditability matters.
3. Use a chat model for open-ended language work with human review.
4. Add retrieval when answers need current or private knowledge.
5. Add tools or agents only when an action is needed; begin with narrow, reversible, read-only work.
6. Measure quality, safety, latency and cost on real examples before scaling.