AI fundamentals

A useful way to think about a model

A model is a learned pattern used to produce an output from an input. Training changes that pattern using examples; inference uses the resulting model. A retailer training on last year's sales and then forecasting next week's demand is performing two different activities, with different costs and responsibilities.

AI is the broad category. Machine learning learns patterns from examples rather than relying only on explicit rules. Generative AI produces content such as text or images. A useful forecast does not require a conversational interface, and a fluent answer is not evidence that a prediction is reliable.

Follow the data

A table of invoice totals is structured data. Emails, photographs, and recorded calls are unstructured data. Both can contain missing information, inconsistent labels, duplicates, and sensitive records. More data helps only when it represents the decision you need to make.

Imagine a delivery model trained only on urban routes. Before applying it to rural deliveries, ask whether the examples represent those journeys and how errors will be measured. Do not assume strong training performance transfers to a different population.

Use a shared vocabulary

Ask teams to define terms in a project glossary. ISO/IEC 23053 provides a framework for describing systems using machine learning; ISO/IEC 42001 concerns an AI management system. Neither is a guarantee that a particular model is accurate. Your job is to clarify the business claim and the evidence required to support it.

Start with the shape of the work

If an expense below a fixed threshold must go to a named approver, rules may solve the problem. If the task is estimating demand from messy historical patterns, predictive ML may fit. If the task is drafting a response, generative AI may help. Choose based on the output and error tolerance, not the newest technology.

An agent can act

A chatbot that drafts an itinerary and an agent that books tickets have different risk profiles. An agent can use tools and carry out steps toward a goal. Set boundaries around the actions it may take, the data it may access, and the approvals it must obtain. Multiple agents communicating does not remove the need for a clear accountable owner.

For a travel agent, searching flights could be automatic while purchasing a nonrefundable ticket requires human confirmation. A useful business requirement specifies the permission boundary and a fallback when a tool fails.

Keep managing after launch

Customer behavior and input data change. A churn model that worked last year may deteriorate this year. Define performance checks, escalation thresholds, and ownership before rollout. Maintain an accessible register of tools that are approved, being evaluated, or blocked so employees can make safe choices without guessing.

Give the model a real brief

A strong prompt states the task, relevant context, constraints, and output format. Instead of “summarize this,” ask for a decision memo with three options, cited evidence, and unresolved questions. Examples can communicate what good output looks like. Test prompts against varied cases, including ambiguous inputs.

Context is a limited working space

Models process text as tokens, and a context window limits the information available in a request. A large document may exceed that limit or bury the relevant detail. Retrieve the relevant portions, split the task thoughtfully, and check that critical qualifications survive any summary.

Choose the adaptation for the need

Retrieval-augmented generation (RAG) supplies relevant external information at answer time. It is a useful starting point for a policy assistant whose source material changes frequently. Access permissions and document freshness still matter, and retrieval does not guarantee a correct answer.

Fine-tuning changes a model using training examples. It can help with a repeated style or task behavior. It is not a substitute for a current, permission-aware knowledge source. A team may use both approaches, but should first evaluate whether better prompts and retrieval solve the business problem.

For a staff policy assistant, require links to approved policy documents and a way to escalate unsupported answers. Evaluate factual support, not just fluent wording.