What is a Task?
Origin taxonomy — five buckets
A task exists because something needs to happen. There are five buckets of "why" — every task we ever create maps to one or more of them. Origin answers why does this task exist? Creation surface (defined in §2, detailed in §7.1) answers where was it created or detected? A single origin can be reached from multiple surfaces.
3.1
Client-driven
The client tells us — explicitly or implicitly — that they need or want something. Client-driven signals split into two sub-buckets because confidence and routing implications differ.
3.1a Explicit — client stated something
The client directly said what they need or want, in a message, voice call, or stated in-app request.
Direct booking request (transactional): "We need a 4-night stay at the Aman Tokyo for our anniversary, last week of October — suite with city view." → open a search task and map results onto trip components.
Mid-trip emergency (urgent): client lands in Mumbai and the pre-arranged driver isn't there → re-dispatch transport + send an apologetic gesture from the TA.
Future-trip seed (low priority): "We're thinking about Patagonia for our 25th anniversary in 2027." → create a soft trip + tickle file for outreach in late 2026.
Typical owner: TA for relationship-critical responses; ops for transactional execution.
Default routing: Assist — AI drafts response, human approves.
Default routing: Assist — AI drafts response, human approves.
3.1b Behavioural / implicit — inferred from context
The client didn't directly ask for anything, but their behaviour (in chat or in the app) signals preference, intent, or need.
Vague taste signal (from message content): on a past trip the client said "the Lutetia was way too modern for us" → save preference: avoid contemporary properties.
Health / dietary signal (from message content): client mentions in passing "I've been doing keto since January" → update client profile dietary; brief all upcoming restaurant bookings.
App behavioural signal: client taps 4 Aman properties in same region during an active planning window → consider a planning nudge. Or: client saves 3 deals to Maldives over 2 weeks → consider a discovery push.
Key rule — behavioural signals don't become tasks by default.
They feed the taste graph / client profile as background context. A Suggestion fires only when the signal meets a threshold + trip-context condition: enough signals in a tight window, plus an active or imminent trip the signal informs.
| Behaviour | Outcome |
|---|---|
| Single tap on a Maldives property | Client profile update only, no task |
| Liked 4 properties in same region during an active planning window | Suggestion fires — "client seems decided on region X; explore?" |
| Searched "honeymoon Bali" 3 times in a week, no active trip | Client profile update + soft trip seed; no immediate task |
| Repeated dynamic-trip-card opens on an upcoming trip's hotel detail | Suggestion to TA — "client engaging with this booking; offer to walk them through?" |
Typical owner: TA for the inferred preference-to-trip translation; ops for client profile updates.
Default routing: Assist (preference capture) or Suggestion-only (behavioural threshold) — depending on confidence.
Default routing: Assist (preference capture) or Suggestion-only (behavioural threshold) — depending on confidence.
3.2
World-driven (external)
Something outside our system happens that affects an active or upcoming trip.
Weather disruption: Typhoon warning issued for Seoul, client arriving in 48 hrs → alert client, monitor hourly, prepare reroute scenarios.
Schedule cancellation: The airline cancels the client's outbound the day before departure → rebook on a partner carrier, alert client, coordinate fresh ground transfers.
Vendor-side change: Four Seasons George V notifies that the suite category the client booked is closed for renovation → inform client and offer alternatives within their preference profile.
Award space opens: First-class award appears on the client's saved LAX-SYD August dates → book within minutes before it disappears, then notify.
Compliance advisory: State Dept upgrades a destination to Level 3 a week before departure → brief client on safety considerations, offer no-penalty cancellation, document the conversation for audit.
Typical owner: AI for detection and templated alerts; human for client-facing communication and re-booking decisions.
3.3
Trip lifecycle
The trip itself, as it moves through our pipeline or its calendar, generates work.
Pre-arrival window (T-10 days): Trip enters its final pre-departure week → spawn pre-arrival checklist (re-confirm transfers, brief restaurants on dietary notes, send weather + packing notes).
Day-of-departure send-off (T-4 hours): Trip is hours from wheels-up → send "have a wonderful trip" message with confirmation links and the 24/7 emergency contact.
Day-1 check-in (T+24 hours after arrival): First night complete → automated "How's the room? Anything we should fix?" — catches issues early before they fester.
Post-trip review (T+2 days): Trip flips to complete → personalised follow-up email, feedback request, and capture any new preferences observed during the trip.
Pipeline stall detection: Trip sits in
Booking stage > 5 days with no component changes → flag to ops manager, "what's blocking?"
Typical owner: ops for procedural work; AI for templated communications and scheduled checks.
3.4
Relationship
Pure relationship-layer moments, sourced from the client profile or accumulated context, surfaced by date or pattern.
Birthday outreach (T-7 days): client's birthday in a week → TA hand-crafts a card or gift, ideally tied to a property or preference they love.
Anniversary trip seed: couple's wedding anniversary visible in the client profile → propose a destination drawn from their history (e.g. "where they honeymooned 5 years ago") ahead of the next planning cycle.
Returning-client gesture (3rd trip): client books their 3rd trip with us → trigger a personalised welcome amenity request at the destination property, charged to the relationship.
Care-context check-in: client mentioned a parent's chemo on a past trip → 6 weeks later, a gentle "thinking of you" message before any commercial outreach lands.
Referral acknowledgement: existing client's friend signs up via referral → thank-you message to the referrer, and a flag to mention the new client's first trip on the next monthly check-in.
Typical owner: TA — human voice irreplaceable here. Automation tolerance is near zero; this is the bucket where a wrong AI message destroys trust.
3.5
Internal operations
Our own workflow signals — surfaced by self-monitoring rather than external events.
Coverage gap: TA going on 2-week leave → reassign their 8 active trips, brief covering TAs, set OOO auto-replies on relevant channels.
Shift handoff prep: APAC shift end → generate end-of-shift summary with open issues + watch list for the EMEA shift starting in 2 hours.
Client profile hygiene gap: trip closed last week with a new preference observed in chat ("hates spicy food") but never logged → reminder to update the client profile before the trip file is considered done.
SLA breach forming: VVIP client's message unanswered > 90 min → escalate to ops manager and the assigned TA in parallel.
Workload imbalance: TA A has 12 active trips, TA B has 5 → recommend the next incoming request to TA B; flag if A is consistently above capacity week-over-week.
Typical owner: ops manager for escalations; high automation tolerance for the rest — this is self-driving territory.