All seven angles of this expedition converge on the same structural finding: AI output quality is bounded by input structure. Attaching EventStorming board photos to an LLM prompt produced 3× more domain concepts than unstructured prose [1]. EARS “shall” statements produce requirements that AI coding agents can parse as executable contracts [2]. Meeting summaries that don’t push action items to a project board are never acted on [3]. The shared failure mode across every task type is treating AI as a shortcut that works on messy inputs. Structure-first discovery methods — EventStorming, Domain Storytelling, EARS — are the multiplier; prompting is only the interface.
What the numbers add up to. Current AI tooling addresses roughly 10–15 hours/week of mechanical BA/PO work: backlog management consumes ~20% of a PO’s workweek, with AI trimming up to 10 hours/week [4]; status reporting runs 3–5 hours/week (roughly 200 hours/year), compressible to under 30 minutes with prompting [5]; teams with integrated data pipelines report 30–50% reduction in report creation time [6]. Realising these gains requires integration — standalone LLM sessions that don’t push artifacts to Jira, Confluence, or ADO capture a fraction of the value [3].
The consistent risk signal. Every child that involved AI-generated artifacts identified the same failure mode: syntactically correct output with wrong business intent. AI generates EARS-formatted requirements that look precise but drop real constraints [7]; it calculates WSJF scores with engineering effort estimates 10–20× too high [8]; it writes “so that” clauses that sound plausible but don’t capture actual stakeholder motivation. The mitigation is consistent: AI drafts, human verifies with domain knowledge. What remains underexplored is who owns that verification step — the validation and tooling angles ran at shallow depth (3 and 7 citations respectively), leaving governance, sign-off, and organisational accountability as the least-documented part of the playbook.
A tension in adoption sequencing. The backlog and communication children recommend starting with raw LLM prompting — zero integration cost, immediate wins. The discovery and spec children show that structured methods (EventStorming, EARS, Domain Storytelling) multiply AI effectiveness across the entire downstream pipeline, but carry a higher adoption barrier. These recommendations pull in opposite directions: quick wins favour prompt-first; systemic ROI favours structure-first. Organisations that start with prompting and never invest in structured discovery will capture perhaps 30% of the available efficiency gain — enough to notice, not enough to change the workflow.
The two-audience problem for specs is the hardest discipline shift in this playbook. Functional specs now serve human stakeholders and AI coding agents simultaneously [7], and optimising for one audience degrades the other. BAs and FAs who write only for human readability will erode their engineering team’s AI-assisted development velocity. This dual-audience habit is also the one with the longest leverage — it propagates value from discovery all the way through implementation.
What the expedition doesn’t answer: given that structured discovery methods multiply AI ROI across the full lifecycle but require organisational buy-in, what is the minimum viable structure a solo BA or PO can introduce unilaterally — before the rest of the organisation is on board — to still capture a meaningful downstream AI multiplier?