AI as the Tester's Assistant

The newest tool in the QA toolbox can read a user story and draft test cases in seconds. Large language models — the AI behind modern chat assistants — are already part of daily QA work on many teams, and the professional skill is neither worship nor refusal. It's knowing exactly what to delegate and what to check. This page draws that line, and it opens with the one fact everything else depends on — so read the next paragraph twice.

What an LLM Is, Honestly

A large language model (LLM) is a program trained on enormous amounts of text that predicts fluent continuations — given some words, it produces the words that plausibly follow. That's genuinely useful for language tasks, and it comes with a property you must internalize before using it for anything: there is no built-in guarantee of truth. An LLM produces confident, fluent, well-formed output whether or not that output is correct — the failure mode has a name, hallucination: stating something false with total confidence, in perfect grammar. Fluency and truth are unrelated. Tattoo that sentence; the rest of this page — and much of your professional relationship with these tools — is a corollary of it.

The right mental model: a brilliant, tireless intern with no fear of being wrong. Drafts anything instantly, phrases everything beautifully, works around the clock — and occasionally invents a citation with a straight face. You'd never fire this intern. You'd also never ship their work unread.

Where It Genuinely Helps

Real, daily uses — each a task where a fluent draft saves time and your judgment does the rest. First-draft test cases from a story: paste the gift-card story, get fourteen candidate cases in a minute — then you apply Chapter 4. Bug report polish: feed it your rough notes and the FW-312 anatomy, get a clean, structured report to review and correct. Test data generation: "a hundred realistic traveler names and edge-case dates" — Chapter 5's fixtures, drafted (and note: fictional data, never real). Summarizing a long ticket thread before triage. In every case the pattern is identical: AI produces the draft at speed; you supply the judgment. That division is the whole skill.

The Verification Rule

One law governs all of it: AI output is a draft from a confident stranger — review it with exactly the skills this book taught. And here's the beautiful part, the reason your thirteen chapters matter more in the AI era, not less: the Chapter 4 techniques are precisely how you audit an AI's test cases. The LLM reliably nails the happy path and reliably misses boundaries, invalid partitions, and state transitions — a claim you can test yourself in five minutes with any story. So your audit isn't vague "double-checking"; it's a checklist you already own: which partitions did it skip? which boundaries? which forbidden state transitions? which combination in the decision table? The AI drafts; your test design finds the holes. Un-audited AI test cases are noise generated at high speed — the Chapter 9 refrain, reincarnated: a bad test drafted by AI is a bad test, faster.

AI draftsspeed · fourteen cases in a minute

→

You auditCh 4 lenses · find the holes

→

Real test setjudgment · nine kept + three added

What Not to Paste

A hard boundary, because it has legal teeth. Never paste into an external AI service: real user data (Chapter 5's never-real-data rule extends here with force — an outside AI is the most external place your data could go, and it may be used for training), credentials or secrets, or unreleased features under NDA. Teams increasingly have written AI policies governing which tools are approved and what may be shared — a thing that exists, varies by company, and that you check before pasting, not after. When in doubt, the rule is the same as everywhere in this book: fictional and public in, sensitive nothing in.

Fernway's Gift Cards, Drafted

The gift-card story — Chapter 2's cliffhanger, finally built — needs test cases, and you try the LLM. It drafts fourteen in a minute: fast, fluent, professional-looking. Then you audit with Chapter 4, and the pattern is exactly as promised. All fourteen are happy-path variations — no expiry boundary (what about a gift card one day expired?), no partial-redemption states (card covers part of the booking — then what?), and one case tests a rule that doesn't exist in Fernway, hallucinated from other products' gift cards the model had seen. Your audit turns fourteen drafts into nine real cases (dropping the duplicate and the hallucinated one, fixing the vague expectations) plus three the AI never imagined (the expiry boundary, the partial-redemption state, the gift-card-plus-promo interaction — a decision table the model didn't think to build). Fourteen in, twelve out — and the twelve are better than either you-alone or AI-alone would have produced. That division of labor is the lesson, and it's why the honest 2026 answer to "does AI replace testers?" is: it replaces typing, not judgment.

One honest dating note: the tools here move fast — AI features inside test-management products (case generation, self-healing locators — next-but-one topic weighs that one), better models quarterly. The categories-over-products discipline from Chapter 9 applies: learn the draft-then-audit pattern, and every new AI tool slots into it.