Overview
Conductor Task Creation & Management — v1.0
Maestro's task layer detects the dozens of small jobs behind a perfect trip, routes each to the cheapest competent worker (AI or human), and closes them. v1 ships with human approval on every AI-suggested task; v2 graduates to auto-approval above a calibrated confidence threshold per task type.
Every task moves through a 4-stage flow: Signal (raw event) → Suggestion (system's proposal) → Approval gate (human in v1; auto above threshold in v2) → Canonical task. Every task answers a "why" via one of five origin buckets (Client-driven, World-driven, Trip lifecycle, Relationship, Internal ops) and executes via one of three routing tiers (auto, assist, escalate). The unit economics depend on this layer working — one TA should run the trip volume that today takes a small ops team.
Start with the executive summary above (2 minutes). For depth: §3 origin taxonomy and §7 lifecycle. For the decisions the alignment meeting needs to make: §8 (15 open questions). Appendices A, B, C, D, E, F are reference material for engineering and design.
Problem Statement
Maestro's business model rests on a simple split: the customer just wants the perfect trip; the business needs that perfect trip to be delivered correctly and cheaply.
Planning a trip isn't the hard part — anyone can do that. The hard part is the long tail of small jobs behind a perfect trip. Each is small. There are dozens per trip. Across a portfolio, thousands.
"Cheaply" means AI does the work where it can; humans do it where AI can't. Both kinds of work need to live in one system so the team never loses track of what's happening, who (or what) is doing it, and whether it's done.
If we get this right, a single travel advisor can run the trip volume that today takes a small ops team. If we get it wrong, the AI either drowns the team in noise or misses the moments that matter. Both kill the unit economics.
Reference Patterns
External products with shipped patterns we can borrow from. Linked here so the alignment meeting has a shared visual vocabulary.
Conversational creation ("set up a daily X"), proactive firing with desktop notification, dedicated Scheduled section in sidebar, view/edit/disable from same UI. A widely-cited pattern for proactive task surfacing — useful reference for World-driven and Trip lifecycle origins, and for proactive closure suggestions.
Context-aware reply suggestions with source transparency (which past conversations / knowledge base entries informed the suggestion). Useful reference for in-chat AI suggestions — particularly the source-citation pattern, which builds operator trust.
Inline reply variations generated as you read; multiple variations to pick from, send with one click. Useful reference for the lightweight UX of an inline suggestion that doesn't break flow.
Three Jobs of the Task Layer
See every task that needs to happen, across human-judgment moments and ambient system signals.
Send each task to the cheapest competent worker (AI or human).
Confirm completion confidently, auto-close where the system can prove it's done, manual where it can't.