Elicitation produces the raw material every later artifact inherits — and AI can synthesize ten hours of workshops in minutes. The catch: it will attribute needs to people who never stated them, fluently. This module builds the citation-disciplined synthesis workflow, the prep and mining accelerators, and the adversarial pass that finds what you didn't ask.
Arjun is a Business Analyst at a Gurugram consulting firm, six years in, running discovery for a warranty-claims transformation at an appliance manufacturer. Over two weeks he collects nine hours of recordings: three workshops, four interviews, one shadowing session. On Thursday night he loads the transcripts into Claude and asks for a thematic synthesis of stakeholder needs. Twenty minutes later he has a beautiful findings pack — themes, sub-themes, stated needs per stakeholder group, even a tension analysis. He circulates it Friday morning.
Monday, the service-operations head replies to all: "I never said we need real-time claim tracking for customers. I said customers keep calling because our letters are unclear. Please correct the record before this reaches the steering committee."
Arjun re-reads the transcript. She is right. She had described the call-volume problem; a different stakeholder — the digital-channels manager — had floated real-time tracking as a solution; and the model had merged the two into "Service Operations requires real-time customer claim visibility." It even attached her name. It reads perfectly. It is false. And it is the most dangerous kind of false: a plausible need, attributed to a credible owner, that would have sailed into a requirements baseline and possibly into a budget — while the actual finding, unclear letters drive call volume, quietly disappeared.
This failure mode has a name in this track: hallucinated attribution. It is the signature risk of AI-assisted elicitation, it survives casual review because everything sounds right, and the entire workflow below is designed so it cannot survive your review.
In synthesis, the model's failure is rarely inventing wild content — it is recombining real content across speakers, merging a problem from one person with a solution from another and assigning the hybrid to whoever sounds senior. Only citation-level traceability catches it.
The fix for Arjun's Monday is not "review harder" — nine hours of transcript defeats diligence. The fix is a synthesis contract that makes verification cheap:
Every claim above is one click from its transcript line. That is what "synthesized" has to mean.
The panel's last row is why this workflow earns its keep: honest synthesis surfaces the informal WhatsApp approval path that no official document mentions — the kind of finding that changes a design, and that a stenographer's summary would have smoothed over.
Two habits from Module 1 get their full form here.
Meeting capture to structured elicitation notes. Run your meeting AI (Teams or Zoom AI, a Fireflies-class notetaker — whichever is sanctioned, and always with recording consent, which client policy may require explicitly). Then run the transcript through a fixed extraction template: stated needs, constraints, assumptions, decisions, open questions, follow-ups with owners. You review the extraction against your memory of the meeting before it circulates — you were there; the review takes five minutes and catches the misreadings.
Elicitation prep at scale. Before sessions, generate stakeholder-specific question guides from role context — the CFO gets benefits-and-baseline questions, the ops lead gets exception-and-volume questions, the architect gets constraint-and-integration questions. And build straw-man models: deliberately imperfect first-cut process sketches or data models to put on the wall, because people who cannot articulate their process from a blank page will instantly tell you everything wrong with your draft of it. The straw man is a provocation artifact; its errors are its function.
Tell the room the straw man is a deliberately rough draft built to be corrected. A straw man presented as a finished proposal doesn't provoke correction — it provokes either silent agreement or a credibility problem, and both are elicitation failures.
Half of discovery evidence is not in anyone's head — it is in documents and ticket queues, and long-context models have made mining them a one-day task instead of a three-week one.
Give the model the org context — structure, project scope, named roles — and have it draft the RACI, the power-interest grid, and an engagement cadence per stakeholder. The draft will be 70% right and confidently wrong about the 30% that matters: it cannot know that the "consulted" enterprise architect actually blocks anything he wasn't accountable for, or that the quiet regional ops manager is the real decision influencer. You correct it from lived knowledge — and, as with the straw man, the corrections are the analysis. A drafted grid you correct in twenty minutes beats a blank template you procrastinate on for a week, and the act of correcting forces you to make your stakeholder knowledge explicit.
The most valuable prompt in elicitation is not generative but adversarial. After each discovery round, hand the model your elicitation notes and draft findings and ask it to attack the coverage:
| Coverage dimension | The adversarial question | Typical catch |
|---|---|---|
| Stakeholder groups | Who is affected by this process but appears in no session? | Downstream finance team, auditors, the customer themselves |
| Exception paths | Which failure and edge cases did nobody describe? | Returns, reversals, partial approvals, dead-letter queues |
| NFR classes | Which of performance, security, audit, accessibility, retention went unmentioned? | Almost always audit and retention |
| Lifecycle states | What happens at setup, migration, month-end, decommission? | Month-end close behavior; data migration rules |
| Volumes and peaks | Do we have numbers, or adjectives? | "High volume" with no measured baseline |
Feed it checklists as context — BABOK technique lists, Volere NFR categories — and the pass gets sharper. The output is not requirements; it is your next elicitation plan. This is the highest-value low-risk critique use in the whole discipline: the model needs no ground truth to notice that no session mentioned month-end.
Coverage critique is safe precisely because its failure mode is a wasted question, not a wrong requirement. When the model suggests a gap that isn't one, you lose thirty seconds. When it finds a real one, you save a change request.
Everything above produces artifacts; the last pattern makes them queryable. Ground a Q&A assistant in your project's own corpus — prior BRDs, decision logs, meeting notes, findings packs — via Confluence with Rovo, Notion AI Q&A, or an API-plus-vector-store setup if your team has one. Two non-negotiable behaviors, which you test before trusting: answers must carry citations to the source document, and the assistant must refuse when evidence is thin rather than improvise. Ask it three questions you know the answer to and one you know is not in the corpus; if the fourth gets a confident answer, your knowledge base is a liability. Done right, this is the onboarding accelerant for every analyst who joins mid-program — "what did we decide about duplicate claims and why" becomes a thirty-second lookup with a citation instead of an afternoon of tribal archaeology.
AI synthesis without citations is rumor at scale — the model will merge speakers, convert proposals into needs, and attribute the result to whoever sounds senior. Demand speaker-and-timestamp citations, keep proposals and conflicts labeled, spot-verify before circulation, and spend the saved hours on the adversarial question no model can answer for you: what didn't we ask?