1Goal
Automate the entire lead-to-booking workflow through an agent. Keep the human in the loop for one job only: reassurance — giving the customer confidence that we are here.
Four outcomes, in order:
- Seller conversion improves — sellers spend their minutes only on verified, scored, context-rich leads.
- We are relevant to the customer where he is — WhatsApp, the app, the call. Right content, right moment.
- The customer uses the app more and edits the itinerary — editing is engagement, and it is also the strongest intent signal we can capture.
- We learn from every interaction — score the lead consistently, move the stage dynamically, and adjust margin by departure-date urgency.
2Design principles
- Hot leads get picked up, not queued. A high-score lead goes to open pickup — any seller who can take it, takes it, immediately. Assignment logic can wait; the customer cannot. (External research — Jacada — pegs the conversion cost of slow first contact at 8x.) Score-based routing exists precisely so the AI call filters the doubtful without delaying the obvious.
- The human is for reassurance, not administration. Every task the seller does today that isn't building trust or closing gets absorbed by the agent.
- Price competitiveness, not discounting. When the agent senses a deal hunter, the play is a different destination at a lower price point — switching the product, not cutting the price of the same one. Consistent with the incentive redesign: discount is not a converting lever; competitiveness is.
- Discount is psychology, not economics. The data says discount doesn't buy conversion — but the coupon still matters. It creates urgency through expiry, and the customer feels he has won something. The agent deploys coupons deliberately as a closing ritual: right moment, tight expiry, capped value. The win feeling is the product; the percentage is just the packaging.
- Every interaction is training data. App events, call transcripts, WhatsApp replies, itinerary edits — all of it feeds the score, the stage, and the next best action.
3Architecture — three engines and a brain
Lead lands → instant score from product interactions (search depth, destination certainty, dates entered, budget signals, repeat visits) plus source and profile.
- High score → open pickup, immediately. Any available seller grabs it. No AI call in the path, no assignment queue in the way.
- Low score → AI verification call. The agent calls to verify intent: real trip or browsing, dates, pax, budget band, destination certainty. Verified → seller queue with full context. Not verified → nurture-only track. No answer → retry policy, then nurture.
- Context messages with itineraries relevant to his search — before the first call, so he arrives informed.
- The generated itinerary, sent to him — editable in the app.
- Automated follow-ups, paced by his engagement — with deliberate depth: high-consideration closes typically take six-plus touches over ~two weeks, and most sellers give up after two. The agent doesn't fizzle at three.
- Testimonials — matched to his destination and traveller type.
- Destination content — the fall-in-love layer.
- Deal coupon — only on stall triggers, within guardrails.
- Callback asks — inviting him to call the seller back when his signals spike.
- Call setup and scheduling.
- First-call scoring — quality, coverage of discovery questions, next-step clarity.
- Profile enrichment after every touch: first-time traveller, food preferences, city, why this trip, where he's been, budget.
- Callback nudges — 4, 6, or 10 hours, chosen by lead heat and departure proximity.
- Deal-hunter sensing on calls — will he go somewhere else? If yes, the agent proposes an alternative destination at a lower price point.
- Read-up briefs — if the seller doesn't know the alternative destination well, the agent generates the brief before the callback.
Every signal refreshes the lead score; score movement changes the lead stage automatically — no subjective stage calls. Margin moves dynamically by departure date within guardrails. Everything the brain knows surfaces on the seller dashboard.
4The flow, end to end
5Scoring, stages & guardrails
The score
One lead score, continuously refreshed from: product interactions (search depth, itinerary edits, repeat sessions), AI call answers, WhatsApp engagement (replies, clicks, read receipts), call outcomes, and profile completeness. Itinerary edits weigh heaviest — an editing customer is a planning customer.
Stage movement
Stages move on score thresholds, automatically. No seller judgment calls on stage — automating qualification removes the subjectivity the July notes flagged. The seller sees the stage and the why: which signals moved it.
Guardrails — the agent acts alone inside these
An urgency creator, deployed deliberately: timed to decision moments, tight expiry windows to force the date, value capped by market, one active coupon per lead, every issue logged. The customer should feel he won; the margin should barely notice.
Departure-date bands per market. Close departures hold or raise margin — urgency is on our side and he will book early. Distant departures get flexibility. Floors tied to the market margin credit, so seller payouts are never touched. Every move logged.
Frequency caps, quiet hours, engagement-paced sequencing, instant opt-out honoured.
Confusion, complaint, high-value threshold, or explicit ask for a person → seller or TL immediately, agent steps back.
6Call scoring — natural, not rubric
Four flags, three skills, situations scored only when they occur, conversion calibrates everything — and a weekly coaching ritual that actually moves the number.
The only rulebook: four flags per call
Pickyourtrail mentioned with clear differentiation — why the brand, why us.
Pitched on every call — customize your itinerary, get coupon offers on the app.
"Save my number — this is me." The customer saves the seller.
Committed and time-boxed: "Itinerary by 6 tonight, we talk tomorrow at 11 — works?" A warm goodbye with no locked next touch is a weak close — the time-boxed next step is the strongest end-of-call predictor of conversion.
Binary, every call. This is the entire checklist; everything else is judgment.
Three skills, read holistically
The model judges the full transcript with context — repeat customer, existing itinerary, departure date — and writes a short read per skill plus one coaching line. No sub-weights, no criteria lists.
- Probing — why this trip, why a travel agent, the occasion, the decision maker, how well they know the destination.
- Destination authority — engaging the customer on the destination, teaching, clearing queries. Visa included.
- Communication & listening — interruptions, tone, sentiment — plus three measured numbers that are objective, un-gameable and language-agnostic: talk-to-listen ratio (the best sellers sit near 45:55; weak sellers talk 65%+), longest customer monologue (60+ uninterrupted seconds from the customer is a strong positive), and question count in discovery.
Situational overlays — scored only when the situation appears
Close departure → urgency. Competing quote or booked flights → urgency plus rate-match. Final stages → objection handling. A seller is never marked down for not handling an objection on a call that had none.
Urgency must be anchored. A real fact — fare class expiring, hotel inventory, visa processing time — scores; pressure language ("prices go up tomorrow!") doesn't. Fake urgency converts once and burns the repeat relationship — and repeat is our highest-quality lead.
Context first
The call must use what we already know. A repeat customer greeted with "so, what's your holiday?" is a failed call regardless of tone.
The objective function is conversion — and the plumbing exists
Every call traces through IDs to its lead and to whether that lead booked, so calibration runs from day one. Every 4–6 weeks the model recalibrates what "good" looks like against actual conversion — controlled for lead score, so it learns what the seller added, not which leads were hot.
Cold start: before the conversion data accumulates, our best TLs — Chandru and Apoorva's teams — hand-label 50–100 calls as good/bad with one line of why. That seeds "what good looks like at Pickyourtrail" instead of generic sales priors.
Languages
Calls happen in English, Hindi, Tamil, Telugu and Malayalam. Start with the top 3–4. Transcription accuracy is the gate: tone and sentiment judgments are only as good as the transcript underneath — which is exactly why the measured numbers matter; they survive a noisy transcript.
The coaching ritual — the part that moves the number
Scores don't change behaviour; rituals do. Every company that bought call scoring and stopped at dashboards got nothing. Weekly, calendared: the seller self-reviews one scored call first, then the TL reviews one or two per seller against the reads. Dashboard patterns feed the ritual; the ritual moves conversion.
7The seller dashboard
Everything above surfaces in the dashboard we've built:
- Queue — the open-pickup pool for hot leads, plus scored, verified leads with full context: profile, AI call summary, WhatsApp history, itinerary edits.
- Timers — callback nudges with the countdown and the reason (heat, departure proximity).
- Signals — stage, score trajectory, deal-hunter flag, engagement decay.
- Plays — the agent's proposed next best action: alternative destination with price point, read-up brief, coupon status, margin band.
- Call scores — his own call quality, against the discovery checklist.
8V1 pilot & open items
Scope
One market, full flow. Recommendation: Bali — short-haul, high volume, fast booking cycle. The learning loop turns in weeks, not months; Europe's long checkout lag would starve the feedback loop in a pilot.
What we measure
- OCC / SCC / GLOCC for pilot cohort vs control, cut by lead source
- Time-to-first-contact and time-to-first-convert
- AI verification accuracy — agent's verified/not-verified vs eventual seller judgment
- App engagement: itinerary edit rate, sessions per lead
- WhatsApp: reply rate, sequence completion, opt-out rate
- Coupon incidence and coupon-attached conversion vs organic
- Seller minutes per converted lead — the augmentation dividend
Open items
- Score model v0. Which signals, what weights, what thresholds route to seller vs AI call. Start heuristic, learn into it.
- AI call script & voice. Verification questions, retry policy, and the tone bar — it must feel like Pickyourtrail, not a robocall.
- Guardrail numbers. Stall days for coupons, margin bands per departure window, WhatsApp caps. Set before launch, not during.
- Escalation SLA. When the agent hands to a human, how fast does the human appear.
- Itinerary editing scope. What the customer can edit solo vs what triggers seller review.
- Control group design. Same market, same period, agent-off cohort — or we won't be able to attribute anything.
- Pickup fairness. Fastest-finger rewards the idle: the best sellers are busy on calls when the hot lead drops. Watch whether top sellers systematically lose pickups; cap concurrent grabs or heat-weight the pool if they do.