AI Foundations · Module 2 · Lesson 2 of 3

Classification and analysis

Sorting items into categories and making sense of a body of information — two AI task types that produce a judgement, and why that judgement still needs a named human check before anyone acts on it.

Lesson · 15–20 minutes · Text-first

By the end, you can

Explain what makes classification and analysis tasks different from drafting and summarizing (AF-1, AF-2).
Describe the specific risk classification and analysis add: a confident-looking judgement standing in for a checked decision (AF-1).
Match an example classification task and an example analysis task to the review step each needs before its output is acted on (AF-2).

Before you start

This is Module 2, Lesson 2 of the AI Foundations course. It builds on Lesson 1's drafting and summarizing distinction and on Module 1's vocabulary — pattern, model, hallucination. This lesson does not require coding or technical setup.

Classification: sorting into categories

**Classification** means asking an AI tool to sort something — a message, a document, a request — into one of a set of predefined categories. Anthropic's own guide to a common production use of this puts it directly: "ticket routing is a type of classification task. Claude analyzes the content of a support ticket and classifies it into predefined categories based on the issue type, urgency, required expertise, or other relevant factors." The output of a classification task is a label: "billing," "technical," "urgent," "not urgent."

Analysis: making sense of a larger body of information

**Analysis** means asking an AI tool to examine a body of information — a document, a set of figures, a collection of files — and produce a conclusion, pattern or explanation, rather than just a shorter version of the same text or a single category label. Anthropic's own capability listing for Claude describes this in general terms: the ability to "process and analyze text and visual content from PDF documents." Where summarizing (Lesson 1) compresses a document, analysis draws a conclusion from it — for example, "this contract's termination clause is unusually strict compared to the other two."

Why classification and analysis carry a different risk than drafting and summarizing

Drafting and summarizing (Lesson 1) both produce text that a person still has to read before doing anything with it. Classification and analysis are different: they produce a **judgement** — a category label or a conclusion — and a judgement is often designed to be acted on directly, sometimes by a system rather than a person. A misclassified support ticket can be routed to the wrong queue automatically. A confidently stated "sales are declining in this category" can walk straight into a business decision without anyone reading the underlying figures first.

This is Module 1 Lesson 2's fluency point again, in a new shape. The US National Institute of Standards and Technology frames an AI system as something that "can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions" — not a system that has verified those outputs itself. The Open Worldwide Application Security Project's definition of hallucination, "content that seems accurate but is fabricated," applies just as much to a wrongly assigned category or a confidently stated but unchecked trend as it does to an invented fact in a paragraph. A classification or analysis output can look exactly as authoritative when it is wrong as when it is right — the review step here is not "does this text read well," it is "does this specific judgement, and whatever will happen if we act on it, actually hold up."

A worked example

A small support team uses an AI tool for two tasks. First, it sorts incoming customer emails into "billing," "technical" and "feedback" categories, and each category routes to a different queue automatically. Second, it analyzes a spreadsheet of monthly sales figures and reports that one product category is declining.

For the classification task, the team does not check every single routed email — that would remove the point of automating it. Instead, they spot-check a random sample each week, and they build a specific escalation path: any message the tool marks with low confidence, or any message a customer marks as "still not resolved" after being routed, gets a second look by a person. This targets classification's particular risk — a wrong label routed straight into an automated action — without requiring a person to redo the whole job.

For the analysis task, the "declining category" claim is treated as a starting point, not a finding. Before it becomes a decision — discontinuing a product line, say — someone opens the actual sales figures and checks the claim against them, because the cost of acting on a wrong analysis (an irreversible business decision) is much higher than the cost of one extra check.

Accessibility notes

This lesson is text-first, with no images, audio, video or downloadable artifacts. The practice exercise's model answer sits behind a native disclosure control that is reachable and operable by keyboard and correctly announced by screen readers. The knowledge check uses native radio-button inputs with a visible question and options, and posts its result to a live status region so assistive technology announces the outcome without a page reload.

Practice

Classification or analysis: name the risk and the check

A finance team uses an AI tool for two tasks: (1) classifying incoming expense receipts as 'travel', 'equipment', 'meals' or 'other' for the accounting system; (2) analyzing a quarter's worth of expense data and reporting which department is over budget.

For the receipt classification task, name one way a wrong label could cause a problem if nobody ever checked it.
For the budget analysis task, name one reason the 'which department is over budget' conclusion should be checked before anyone acts on it.
Suggest one lightweight, ongoing check for the classification task that does not require reviewing every single receipt.
Suggest one specific check for the analysis task before its conclusion is used in a budget decision.

Compare with a bounded first version

A wrong receipt label could misstate a department's spending category in the accounting system, which compounds over many receipts and eventually misrepresents actual spending patterns without anyone noticing a single error. The budget analysis conclusion should be checked because 'over budget' is a confident-looking judgement built from the AI's own reading of the data — it is not automatically the same as a verified accounting figure, and a decision based on it (cutting a department's budget) is costly and hard to reverse if wrong. A lightweight ongoing check for classification: spot-check a random sample of labelled receipts each period, and route any low-confidence classification to a person automatically. A specific check for the analysis task: pull the underlying department totals directly from the accounting system and confirm the 'over budget' figure matches, before the conclusion is used in a budget decision.

Knowledge check

Try the idea

Low-stakes practice only. This does not score, block progress or create a learner record.

Sources and limits

This lesson synthesises the sources below into a practical learning model. It is not a security standard, legal advice or a guarantee that any particular agent design is safe.

Ticket routing — Anthropic. Describes ticket routing as a classification task: Claude analyzes a support ticket and classifies it into predefined categories based on issue type, urgency, required expertise or other factors.
Features overview — Anthropic. Describes a Claude capability to process and analyze text and visual content from PDF documents.
Artificial Intelligence Risk Management Framework 1.0 — NIST AI Resource Center. Frames an AI system as an engineered or machine-based system that generates outputs such as predictions, recommendations or decisions — not a self-verifying decision-maker.
LLM09:2025 Misinformation — OWASP Gen AI Security Project. Defines hallucination as an LLM generating content that seems accurate but is fabricated — the basis for this lesson's point about confident-looking but unchecked judgements.