AI Project Kickoff Automation
Problem
A boutique hospitality design studio runs the same manual grind on every won project: read the discovery-call notes, set up a project workspace, and hand-write a kickoff brief and meeting agenda. It's slow, inconsistent, and the output quality depends entirely on who did it — and on how good the notes were.
Approach
Built a two-agent n8n pipeline triggered by a 'deal won' webhook. Claude Haiku 4.5 drafts a structured kickoff brief and a separate kickoff agenda from the raw discovery notes; both gather into a single Notion project workspace. Every LLM response passes a defensive parser — code-fence stripping, validated JSON, Notion block-limit handling, and select-option normalisation. Crucially, the brief drafter judges its own confidence: rich notes produce a Complete brief, while notes too thin to work from are flagged Needs Review with the missing information listed. The system does less when it knows less — routing uncertain cases to a human instead of inventing details.
Outcome
One deal in → a project workspace with a tailored brief and a venue-specific agenda out, in about 10 seconds. Tested against three synthetic deals of decreasing note quality: the two rich ones produce Complete briefs with specific agendas; the vague 10-minute intro call correctly flags Needs Review on every run, listing exactly what's missing instead of guessing. Built to mirror an AI-architect assessment for an Australian hospitality-design group — designed on paper, then shipped as working software.