AI & LLM

LLM-Assisted Requirements Analysis

Where a language model genuinely helps a business analyst — drafting acceptance criteria, spotting gaps in workshop notes — and where it quietly makes things worse.

Prima Puspita Prima Puspita CPO & Co-Founder · Business Analyst Lead 2 min read 1,530 views
LLM-Assisted Requirements Analysis

I have used language models on real analysis work for two years now. They are excellent at turning messy workshop notes into structured drafts, and dangerous when you let them invent requirements nobody asked for.

Where the model earns its keep

Summarising a two-hour workshop transcript into decisions, open questions and action items is a solved problem. So is rewriting a vague user story into Given/When/Then scenarios that a tester can execute.

Key points

  • Feed the model the glossary and constraints before asking for anything.
  • Treat generated criteria as drafts that need a named reviewer.
  • Ask for questions back, not just answers.
  • Keep the source notes next to the generated summary for traceability.

Where it needs a leash

Ask for edge cases and you will get twenty, half of which do not apply to the domain. The model does not know that the core banking system rejects transactions after the end-of-day batch unless you tell it, so the prompt must carry the domain constraints.

LLM-Assisted Requirements Analysis
Illustration for “LLM-Assisted Requirements Analysis”

Example in code

const prompt = [
  `Glossary:\n${glossary}`,
  `Constraints:\n${constraints.map((c) => `- ${c}`).join('\n')}`,
  `Workshop notes:\n${notes}`,
  'Return decisions, open questions and acceptance criteria as JSON.',
].join('\n\n');

const draft = await client.complete({ prompt, temperature: 0.2 });
review.push({ draft, reviewer: null, source: notes.id });

A working setup

Every prompt starts with the glossary, the system context and the constraint list. Outputs go into a review column, never straight into the backlog, and each accepted item records who verified it.

A business analyst who can read code and a developer who can read a P&L are the same superpower.

Takeaways

Use the model to reduce the blank-page problem, not the thinking. The analyst still owns the requirement and the conversation with the people who need it.