A practical mental map of AI

Start with a capability you can recognize

AI systems infer from inputs to produce outputs such as a prediction, recommendation or piece of content. Some predict demand; some recognize patterns in documents; others generate text, images or code. Their degree of autonomy varies. OECD: the AI system definition explained provides the underlying definition.

A large language model, or LLM, is a model trained on large amounts of text and other data that can generate language in response to context. A fluent answer is a candidate to examine. It can contain errors or miss information that matters in your operation.

Four questions make the landscape manageable

• What follows explicit rules? A workflow can route an approved form using known conditions. This is useful automation even when no AI is involved.
• What needs interpretation or prediction? A model may classify a request, estimate demand or extract information from varied documents. Check its performance on the cases that matter.
• What could be generated? Generative AI can help draft an explanation, propose alternatives or sketch a design. Give it a purpose, useful context and a way to evaluate the result.
• What next action may the system choose? An agent uses a model to select steps and tools toward a goal. That choice introduces questions about permissions, supervision and recovery.

These are ways to examine a system, not mutually exclusive product categories or a ladder every organization must climb. A defined workflow can contain AI assistance; an agent can call an ordinary rules-based tool. The workflow/agent distinction used here follows Anthropic: workflows and agents.

Connect the capability to the operation

Imagine onboarding a customer. Drafting a document summary could help an employee. It will not, by itself, settle which team owns approval or remove a weekly decision queue. Ask who uses the output, what decision follows and what changes for the customer. The Operating Blueprint helps you see those connections.

Work with AI and keep your judgment active

• State the outcome and the people affected. Give the system relevant facts, constraints, examples and the format you need.
• Separate what you know from what you assume. Ask for alternatives and for the information that would change the recommendation.
• Review the result against reliable sources and the people who know the work. Refine the brief as you learn. Use only information appropriate for the tool and your organization.

You can use AI to explore an unclear problem. Start with a provisional question and improve it through evidence. Neither a polished prompt nor a confident answer establishes that an idea will work.

A question to practise

Choose one repeated frustration in your operation. Describe the better experience you want, identify the work and decisions behind it, and ask which capability might help. Then work through the customer-onboarding decision.