Start with the process, add a tool to a step
Find a step AI can speed up, bolt a tool onto it, measure whether that task got faster. Mental model: AI as a feature. Outcome: incremental and additive. Most of what gets called “using AI” is this.
TipTip.id
SatuSatu.com
AI in attractions & OTA
Put intelligence around the booking—not in place of its truth.
Global leaders are using GenAI to improve discovery, listing quality, localisation, and service. Their transaction rails remain conventional. For SatuSatu, the reusable advantage is an AI-assisted operating layer for direct local supply, anchored by human accountability and deterministic booking state.
Evidence index
Four findings that connect the market scan to the workshop decision.
3 of 6
benchmarked leaders show GenAI-level discovery or conversion.R2
0 of 6
expose a material AI capability in the B2B booking API surface.R1
3 to 5
bets must leave the workshop ranked, with explicit rejection or deferral reasons.A1
90 days
is the proof horizon for an owner, test, success metric, and reject or defer trigger.A1
15–20 min
Read for the argument, not feature ideas. The concise `research-codex/` portfolio identifies promising AI interventions. The deeper `research/` red team narrows them with materiality, capacity, exclusivity and safety gates. This pre-read reconciles both — and, as of 2026-07-27, also reconciles with the companion PM pre-read (P1): its framings, SEA benchmark, TipTip coverage and provocations are folded in, every divergence between the two is named in the evidence notes, and a full side-by-side sits in differences.md.
Observed
Derived
Hypothesis
Internal data gap
00 · Two framings to carry in
Adding AI to what you do, versus rethinking what you do because AI exists.
Both framings in this section are adopted from the companion pre-read. Neither is in the research corpus, and neither is about product — they are about how a company decides and how a team is sized. The evidence sections that follow are sharper for having them stated first.P1
Start with the outcome, redesign around AI
Ask what this would look like if AI had existed from day one, then redesign the workflow. Measure whether the underlying outcome improved. Mental model: AI as infrastructure. Outcome: potentially structural.
The organisational finding
BCG’s position, and the reason this section leads: “Becoming AI-first is an organisational challenge, not a technical one.” The companies achieving transformative results are not the ones with the best models — they are the ones that redesign end-to-end how work gets done. Its five no-regrets moves: build a business-led agenda, not an IT-led one · role-model adoption from the top · pinpoint where roles shift, proactively · pick a few high-value initiatives and prove impact rather than sprawling pilots · fund what works properly.B1
The capacity question · and its published caveat
Tasks completed
Speed
Quality
🔴 Outside the frontier
The useful question is not “what can AI do?” but “what becomes possible at 10× capacity?”
This is the strongest external evidence in either pre-read, and it is the only place we have measured numbers on AI and knowledge work. It is also the one piece of evidence with a published warning attached, which is why both halves appear here.J1
+12.2%
More tasks completed with AI, across 18 realistic consulting tasks.
25.1%
Faster execution on the same task set.
>40%
Higher quality, rated by independent human evaluators. Largest gains went to below-median performers — AI raised the floor.
−19pp
On tasks beyond AI’s capability frontier, consultants using AI performed 19 percentage points worse than those who did not.
Dell’Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon & Lakhani — Navigating the Jagged Technological Frontier, HBS Working Paper 24-013, published in Organization Science (2025). 758 BCG consultants (~7% of the firm’s individual-contributor pool), 18 tasks, three arms — no AI, GPT-4, and GPT-4 with a prompt-engineering overview. ✅ All four figures verified against the paper on 2026-07-27. The −19pp result is why this pack keeps a mutation boundary: AI is not uniformly helpful, and knowing where the frontier sits is the actual skill.
10× is not automatic
Three things decide whether a team reaches 10× or stays at 1.1×: process redesign — AI layered on a broken process yields a faster broken process; judgement development — knowing when to question output is a skill that has to be built deliberately; and a culture of experimentation — small iterated tests beat waiting for a perfect implementation plan. As Mollick puts it, giving everyone tool access and expecting transformation to follow is the common mistake: the tools are necessary but not sufficient.M1
01 · Global + SEA benchmarks & reusable JTBD
The industry automates work that can be reviewed before it becomes a promise.
The benchmark set spans destination marketplaces, passes and concierge models, connectivity infrastructure, Indonesian/SEA distribution, and AI disintermediation. Scores are directional because public evidence differs sharply by company.R1
Discovery, conversion, trust
AI is visible in semantic search, ranking, planning, review summaries, and content production. Booking still hands off to owned transaction rails.
Klook · GetYourGuide · Viator / Tripadvisor · Eventbrite (ticketing side — organiser-facing AI; acquired Mar 2026) · Headout (cautionary case, not a leader — claims are unnamed and its API docs went stale)
Distribution before differentiation
The hard competition is broad inventory plus established agent workflow. Public AI evidence is thinner than the commercial distribution evidence — with one exception: Pelago already runs an AI assistant on web and WhatsApp with measured outcomes, and it is the closest comparator to SatuSatu in the set.
Traveloka TPN · Pelago (Singapore Airlines) · GlobalTix · KKday / rezio · TBO · Golden Rama · Panorama
State, connectivity, control
The leverage is to own the catalog record and specification while using specialist rails. OCTO reduces the need for bespoke connectors.
Prioticket / ETG · Bókun · Ventrata · OCTO · Rezdy / Expedition Software (avoid for now — mid-consolidation, no stable surface to integrate)
Adopted from the PM pre-read · where the two markets diverge
Global players lead on sophistication. SEA players lead on contextual relevance.
| Global market priority | SEA market priority |
|---|---|
| Conversational discovery (ChatGPT-style planning) | WhatsApp-native support and booking automation |
| Dynamic pricing optimisation | Payment method diversity — BNPL, e-wallet, cash-on-delivery |
| Agentic commerce (AI buys for you) | Multi-language and dialect content localisation |
| AI-generated editorial content at scale | Price-sensitivity-driven recommendation — best value, not best match |
| Premium itinerary personalisation | Trust signals — review aggregation, verification |
| API distribution to travel agents | Super-app integration (Grab, GoTo ecosystem) |
🔴 The operative insight: a recommendation engine is useless if it offers a $200 experience to someone whose history shows $30 bookings. SEA AI has to understand local price elasticity, not just preference. ⚠️ One qualifier this pack adds to the first row: a WhatsApp presence is table stakes in Indonesia — Antavaya publishes a named WhatsApp concierge, Golden Rama runs a widget. What is not table stakes is AI-native WhatsApp service with measured outcomes — and Pelago already has it at 30-second turnaround, which is why our earlier refutation of “WhatsApp-first SEA players” was drawn too broadly.P1, R1
Benchmark signal · horizontal bar chart
AI is concentrated before checkout, where output can still be reviewed.
Companies scoring ≥3 (GenAI-assisted or higher) among six consistently scored leaders.
Source: `research-codex/02-competitor-matrix.html`, value-chain matrix. The chart uses direct labels, a common scale, and a single accent to foreground the message, following the selection and decluttering principles in the Storytelling with Data chart guide. Scores are not fully commensurable because evidence availability differs by company.
Reusable jobs—not reusable features
Four actors, one shared need: trustworthy progress with less coordination cost.
The job statements below come from the formal JTBD synthesis and preserve its evidence labels.R3
Customer: certainty, not discovery
“Help me get access to a moment I care about — secure my spot easily and let me get in without friction.” 🔴 The decision is usually already made. The job is certainty and ease of access, not discovery — which makes most recommendation work a poor fit here.
Operator: fill it, and de-risk it
“Help me fill and monetize my event — sell tickets, reach the right audience, and de-risk whether it will work before I commit.”
Customer: discovery, curation, trust
“Help me make the most of my limited time here — find things genuinely worth doing that I can trust, without the effort of planning and vetting.” The decision is not yet made — the opposite of TipTip’s customer job.
Operator: demand without a marketing function
“Bring me paying customers I could never reach on my own — without having to run marketing or a booking system myself.”
🔴 The two products sit on opposite sides of the same axis. TipTip’s customer has decided and wants friction removed; SatuSatu’s customer has not decided and wants judgement supplied. An AI investment that serves one may do nothing for the other — which is the strongest argument in this pack against a single shared “AI discovery” bet.P1
| Actor | Reusable JTBD pattern | What “good” means | AI role | Control that must remain |
|---|---|---|---|---|
| Traveller | T1Plan a feasible unfamiliar trip without building an impossible day. | Constraint-aware plan; less anxiety. | Retrieve, compose, draft. | Human approves feasibility. |
| Traveller | T6–7Execute safely and recover when weather, traffic, or operators disrupt reality. | Correct instructions; accountable recovery. | Summarise context and surface options. | Confirmed record + human judgement. |
| B2B buyer | B2Add differentiated Bali supply without adopting another broad platform. | Unique SKU; sellable spread; reliable fulfilment. | Match client need to approved Pool B supply. | Exclusivity and rate proof. |
| B2B buyer | B3–5Quote quickly while knowing confirmation latency, rate, FX, and cancellation terms. | Fast, feasible, margin-safe quote. | Compose draft and explain trade-offs. | Availability state and price rules. |
| Supplier | S1Turn a brochure, social page, or WhatsApp message into a complete listing. | Accurate SKU live faster. | Extract, structure, localise, draft. | Human verification of inclusions and terms. |
| Supplier | S2–3Declare capacity and confirm inside the workflow already used. | Low adoption burden; truthful SLA. | Parse replies and flag ambiguity. | Explicit confirm / decline state. |
| Event-goer TipTip | E1Get access to a specific moment I have already chosen, without friction. | Ticket secured; entry works on the day. | Little to none — retrieval and reminders at most. | 🔴 Entitlement and entry state are deterministic. Never generated. |
| Promoter TipTip | P1Fill and monetise my event, and know before I commit whether it will work. | Sold-out or profitably-sold event; fewer surprises. | Draft listings and campaign briefs; summarise performance; signal demand. | The commitment decision. Forecast informs; a person underwrites. |
| Operations | O1Maintain one trustworthy catalog across direct and aggregator supply. | No duplicate, stale, or off-brand SKU. | Map, compare, flag, summarise. | Canonical record + QA rules. |
| Operations | O2–4Maintain confirmed booking and commercial state while scaling service. | Accurate booking, refund, settlement, and escalation. | Assist exception handling and reconciliation. | Rules, audit trail, human accountability. |
Reusable pattern: let AI compress interpretation and creation; let systems of record and people own commitments.
Four sprint-1 candidates
Adopted from the companion pre-read: where a competitor has proven the mechanism and the infrastructure already exists internally (n8n, Respond.io, Retool, Content Hub), the row is a sprint-1 candidate rather than a roadmap item. Four qualify: (1) AI SKU onboarding — the confirmed internal bottleneck and the best-evidenced external pattern; (2) AI review summaries on product pages; (3) AI CS reply assist via the existing messaging layer; (4) organiser daily performance summary. ⚠️ Each still has to clear its gate in the decision map — and (2) is weakest for us, because review volume is a function of booking volume, which is the number nobody has counted.P1
02 · AI-assisted vs AI-first
“AI-first” is an operating model. In this category, AI is assisted in the pipe, selectively first in the storefront, and not agentic in production.
The Opus proposition is directionally useful but too absolute: not every benchmark has an AI-first storefront, and several surfaces are unverified. The evidence-safe read separates who commits an outcome from where AI sits in the journey.Evidence, A1
Human-owned outcome
AI drafts, retrieves, translates, classifies, or summarises inside an existing workflow. A person or rule approves the consequential output.
Best fit: listing production, itinerary drafts, review feedback, generic support, reconciliation triage.
State-owned truth
Rules and records decide availability, rates, refunds, confirmation, entitlement, and day-of instructions.
This is not “less innovative.” It is the safety and commercial spine.
AI is the primary interface
The system begins from intent, plans across steps, calls tools, adapts to outcomes, and escalates exceptions. Human work shifts to policy and supervision.
Only credible after catalog, availability, evaluation, permission, and transaction state are reliable.
Two dimensions · one shared scale
01 · Control mode
02 · Surface design
03 · Evidence-safe read
Control mode and surface design answer different questions.
Who commits the outcome?
At level 3, a model proposes and a person or deterministic rule verifies. At level 4, the system takes multi-step action toward a goal without a person approving every step.
Where does the user begin?
AI can be a feature inside a conventional funnel, or intent and AI-led orchestration can become the funnel. An AI-first surface can still hand booking to deterministic rails.
The pattern is asymmetric.
Public AI proof clusters in content, discovery, planning, service assistance, and internal work. Availability, pricing, booking, cancellation, and B2B APIs remain rules-led.
| Level | Name | What it means | Benchmark examples |
|---|---|---|---|
| 0 | Absent | No observed AI or automated judgement on the inspected surface. | SatuSatu’s current product surface. |
| 1 | Rules | Thresholds, flags, workflows, CRUD, and state machines—useful automation, but not model judgement. | Availability, pricing, cancellation, partner-rate, and connectivity surfaces across the set. |
| 2 | Classical ML | Trained scoring, ranking, or prediction; method may be undisclosed. | Product-quality scoring, ranking, and earlier review pipelines. |
| 3 | GenAI-assisted | The model proposes; a person or rule verifies; deterministic control commits the state change. | GYG supplier wizard; Klook localisation and K.AI; Viator/Tripadvisor consumer AI. |
| 4 | Agentic | The system plans and takes multi-step action toward a goal without a person approving each step. | No broadly shipped independent attraction-booking proof; Klook is a closed-beta trajectory. |
Source: detailed competitor matrix, agentic-booking evidence review, Klook/GYG teardown, and Viator/KKday/Headout teardown.
Benchmark classification · current public evidence
Only Viator clears the AI-first bar—and only inside a bounded, owned consumer surface.
A maturity score is not the same as an operating-model classification. This test asks whether AI merely improves tasks, or whether intent and AI-led orchestration have become the primary workflow. “Not yet evidenced” prevents rules-led rails and unverified claims from being forced into either group.
| Benchmark company | Current group | Why it belongs there | Boundary / evidence call |
|---|---|---|---|
| Klook | AI-assistedAI-first trajectory | Shipped GenAI improves localisation, content production, review-to-merchant feedback, SEO, engineering, and trip planning inside existing Klook workflows. | The shopping agent is designed to complete booking and post-sale work, but remains closed beta. Until it is broadly shipped, Klook stays AI-assisted.Evidence |
| GetYourGuide | AI-assisted | Its supplier wizard, semantic retrieval, ranking, image optimisation, review summaries, and internal tools compress work that is still reviewed or constrained by the marketplace. | No agentic itinerary, availability, pricing, booking or customer-service surface was verified, and the partner API remains conventional. 🔄 Corrected 2026-07-27: the same Spring 2026 release also shipped a ChatGPT app, so GYG does have an external agent surface — alongside Klook and Viator on distribution-into-the-assistant. Our earlier read of that release as “a conversion layer, not an external agent” was incomplete.Evidence, P1 |
| Viator / Tripadvisor | AI-first · bounded | An explicitly AI-native consumer MVP begins from traveller intent and uses personalisation, real-time recommendations, GeoAware suggestions, proactive offers, planning, and pre-booking chat; it reached half of English-market web traffic. | AI-first applies to the owned discovery and orchestration surface—not the transaction core or partner API, which remain deterministic and conventional.Evidence |
| Pelago | AI-assisted | 🆕 A Yellow.ai generative-AI assistant runs on Pelago’s website and WhatsApp, handling bookings, cancellations, voucher retrieval and live-agent handoff, with automatic escalation into Zendesk. Enquiry turnaround fell from 2–3 hours to 30 seconds and product-enquiry volume reaching agents individually fell 60%. | Assisted, not AI-first: the assistant sits inside a conventional booking funnel and hands exceptions to humans by design. The closest SEA comparator to SatuSatu — curated activities, premium positioning, and a loyalty moat (KrisFlyer) that AI alone cannot replicate.P1 |
| KKday / rezio | Not yet evidenced | The inspected product is a strong operator system of record for products, sessions, orders, payments, redemption, and channels. | No named AI feature was found in the rezio product, pricing page, or OpenAPI specification. Strong deterministic rails are not AI-assisted by themselves.Evidence |
| Headout | Not yet evidencedClaimed | Company language claims AI-powered support and smarter inventory and pricing, while the Dabble acquisition added computer-vision and spatial capability. | No named, dated, measurable shipped feature was verified. Vendor prose and an acquisition do not establish an operating model.Evidence |
| GlobalTix | Not yet evidencedAnnounced | The live five-question chatbot and demand-tier flags are rules-led. AI and predictive analytics appear in funding announcements. | No follow-up AI release or AI, ranking, recommendation, or forecasting endpoint was found in the partner API.Evidence |
| Traveloka TPN | AI-assisted · consumerNot yet · B2B | TPN has shipped B2B inventory and integration workflows across APIs, mini-apps and redirection. 🔄 On the consumer side, Traveloka 5.0 and Amazon Personalize recommendation scaling are now evidenced — closing our own open question #20. | No named attractions AI feature was verified in the B2B surface, and commercial workflow maturity is not evidence of AI-first operation. The consumer and B2B surfaces must not be conflated: the first is assisted, the second is still unscored. ⛔ The circulated “120M users / ~22% conversion” figures are excluded — they trace to a Tier C domain our research refutes elsewhere.Evidence |
| Eventbrite | AI-assisted · organiser side | 🆕 First ticketing platform to ship AI tools (2023): an event-description generator, an AI email-campaign builder, and audience-finding at scale. All of it compresses the organiser’s creation and marketing burden. | Assisted throughout — the organiser still decides. ⚠️ Acquired by Bending Spoons (closed 10 Mar 2026, ~$500M, delisted), followed by significant US workforce cuts. Read it as a cautionary case for TipTip: shipping organiser AI early did not, by itself, secure the company’s independence.P1 |
| TBO Holidays | Not yet evidenced | TBO’s observable advantage is the breadth of its B2B portal, sightseeing inventory, agent network, XML API, and white-label distribution. | No named TBO attractions AI feature was verified; scale and automation should not be relabelled as AI.Evidence |
| Prioticket / ETG | Not yet evidenced | Its benchmark role is neutral infrastructure: maintained distributor rails, public documentation, and OCTO-native connectivity. | No material AI differentiator was needed or verified. It is a deterministic-core benchmark, not an AI operating-model benchmark.Evidence |
| TipTip / SatuSatu | AI-assisted · TipTipNot yet · SatuSatu | TipTip’s production forecasting improves an event-business decision workflow. SatuSatu currently exposes no AI surface; its catalog, Pass, feed, and WhatsApp queues remain human or rules-led. | Transfer the production discipline, people, evaluation, and monitoring—not the forecasting model, because the attractions demand shape differs.Evidence |
Storytelling-with-data view · all twelve benchmarked companies
The AI is in the storefront; the pipe remains deterministic.
Storefront
Supply / B2B pipe
Closed-beta trajectory
Self-reported band — claimed, never named or measured
Not scored on comparable evidence
Highest externally verifiable maturity on the traveller-facing surface versus the supply and B2B connectivity layer. All twelve benchmarked companies are listed, and the encoding distinguishes three evidence states rather than flattening them. A filled marker is a score the research assigns from inspectable artefacts. A hollow marker on a dashed band is a range the company claims in a dated release but never names or measures. A marker in the not-scored gutter left of the scale means no comparable score exists at all — never a zero, because absence of evidence is not evidence of absence. Scope: “pipe” is the highest score across feed ingestion, availability, pricing, B2B rate management and connectivity, so it can read higher than the connectivity API alone — GlobalTix is pipe 1 here and connectivity 0 in the §03 matrix, and both are correct.
Open accessible table view
| Company | Storefront | Supply / B2B pipe | Evidence interpretation |
|---|---|---|---|
| Viator / Tripadvisor | 3 | 1 | AI-native discovery over conventional partner and transaction rails. |
| Klook | 3; beta trajectory to 4 | 1 | Consumer assistance is shipped; the shopping agent remains closed beta. |
| GetYourGuide | 3 | 1 | AI in search, ranking, summaries and supplier creation; conventional partner connectivity. |
| Pelago (Singapore Airlines) | 3 | Not scored | 🆕 Added 2026-07-27 — the closest SEA comparator to SatuSatu, and absent from our benchmark set until now. A Yellow.ai generative-AI travel assistant runs on the website and WhatsApp, handling bookings, cancellations, voucher retrieval and live-agent handoff. Enquiry turnaround 2–3 hours → 30 seconds; 60% fewer product-enquiry questions reaching agents individually. Its own tech blog documents the LLM architecture. No B2B distribution surface was assessed, so the pipe is unscored. |
| GlobalTix | 1 | 1 | Scripted five-question service widget and deterministic commercial rails. |
| Headout | 1–3 (self-reported) | 0–1 | The research scores customer service at 2–3 “self-reported, unnamed” and discovery and in-destination at 1 “direction only” — so a band exists. Its basis is a single dated release (2025-01-15) claiming AI-powered support and smarter inventory and pricing, with no product name and no metric. Public API docs last pushed July 2024. |
| KKday / rezio | Not scored | 1 | No AI in the product, pricing page or 752 KB OpenAPI spec. The consumer app was never inspected, so the storefront is unverified rather than zero. |
| Traveloka TPN | 1–2 (self-reported) | 1 | 🔄 Updated 2026-07-27, closing our own open question #20 (“consumer-side assistants exist but were not verified”). Basis: Traveloka 5.0 (company newsroom, Feb 2025) and Amazon Personalize recommendation scaling (iTnews Asia). ⛔ The widely-quoted “120M users / ~22% conversion lift” pair is excluded: it traces to businessmodelcanvastemplate.com, the same Tier C SEO domain our research names and refutes elsewhere. The B2B partner surface remains unscored — that assessment was separate and stands. |
| TBO Holidays | Not scored | 1 | Same basis: one overall score, no per-stage assessment. Scale and automation are not AI. |
| Prioticket / ETG | Not scored | Not scored | Carries no AI score anywhere in the research. Its benchmark role is neutral infrastructure. |
| Eventbrite | Not scored | Not scored | 🆕 Added 2026-07-27 — the ticketing-side analogue this pack was missing. First ticketing platform to ship AI tools (2023): AI event-description generator, AI email-campaign builder, audience-finding at scale. Its AI is organiser-facing content tooling, so it maps to neither of this chart's two surfaces. ⚠️ Acquired by Bending Spoons, closed 10 Mar 2026 (~$500M, $4.50/share, NYSE delisting), after which a significant part of the US workforce was cut — so the tooling exists but the team that built it largely does not. |
| SatuSatu today | 0 | 0 | No current product AI surface; catalog, Pass, feed and WhatsApp queues are human or rules-led. |
The mutation boundary
AI may propose before a transaction mutates. Deterministic controls or accountable people must authorize state changes with real-world consequences. Drafting an itinerary is assistable; confirming capacity requires explicit evidence; day-of instructions should be rendered from confirmed records; disruption and safety recovery remain human.
| Dimension | AI-assisted | AI-first | TipTip / SatuSatu workshop call |
|---|---|---|---|
| Starting point | An existing task and interface. | Intent becomes the primary interface. | Start from actual creator, event/community, concierge, catalog, supplier, and partner journeys. |
| Accountability | Human signs off or rules constrain output. | System acts within policy; human handles exceptions. | Keep human sign-off anywhere a wrong answer changes a trip, payment, entitlement, or partner promise. |
| Data requirement | Examples, templates, knowledge base, edit feedback. | Reliable tools, state, event history, evals, permissions. | Inventory current TipTip evidence and fill SatuSatu baselines before claiming autonomy. |
| Benchmark proof | Shipped Klook and GetYourGuide are clearly assisted; TipTip is assisted at group level. Viator also contains assisted components. | Bounded proof Viator clears the bar in its owned consumer surface. Klook is a closed-beta trajectory, not yet a shipped AI-first operating model. | Copy the reviewed assistance patterns now; validate journey-level AI-first behavior without surrendering transaction truth. |
| Failure mode | Bad draft consumes review time. | Bad action creates financial, safety, or fulfilment exposure. | Match autonomy to blast radius, not model fluency. |
| Economic logic | Capacity, cycle-time, quality, language coverage. | New workflow or revenue model at scale. | Indonesian labour weakens pure deflection economics; value capacity and quality honestly. |
Market proof ≠ local proof
Vendor-reported deflection consistently exceeds measured production results in the collected evidence. For SatuSatu, silence after a wrong in-destination instruction can look like “resolution” while the traveller is walking toward the wrong location. Measure confirmed resolution, re-contact, instruction error, and human recovery—not containment alone.R4
Reconciling two vocabularies · adopted from the PM pre-read
Filter · question 1
Filter · question 2
Filter · question 3
“Automate or augment” is the same question, one mode short.
The companion pre-read frames the choice as automate vs augment, which is the more natural language for a workshop. It maps cleanly onto the three control modes above, with one addition that matters: a large share of our highest-value near-term work is neither automated nor augmented — it is deterministic, and calling it “automation” invites a model where a rule belongs.P1
| PM pre-read term | Control mode here | Who commits the outcome | Where the two framings differ |
|---|---|---|---|
| Automate | Deterministic | A rule or record, from a known state transition. | 🔴 The important gap. Rate integrity, refund rendering, entitlement, sold-count separation and day-of instructions are not “AI automation” — they are rules. Most of our first-quarter value sits here, and a model would add risk without adding information. |
| Automate | AI-assisted, high-confidence | A model proposes; a gate or sampling check commits. | Same intent. The codex insists the verification step be named, because “automated” usually hides one. |
| Augment | AI-assisted, human-in-loop | A person, every time. | Identical. |
| (no equivalent) | Human judgement | A person, and no model output is offered at all. | Safety-adjacent classes — water and surf conditions, volcano status, scooter advice, medical or allergen guidance — where the right answer is not to generate, reviewed or otherwise. “Augment” understates this. |
Is the task well-defined and repetitive?
If yes, it is a candidate for a rule or high-confidence assistance. If it changes materially every time, or needs context to perform correctly at all, start with a human in the loop.
What does an error cost, and would we see it?
Cheap and catchable → automate. Expensive, or invisible until a customer is affected → keep a person. This is the decisive question, and the one most often skipped.
Is there irreplaceable human judgement?
Pricing strategy, partner relationships, crisis response, creative direction, in-destination recovery. Augment; do not automate. No amount of model quality changes this answer.
Klarna: contested, and the lesson survives either reading
The companion pre-read cites Klarna as a clear case of automating what needed augmenting. The evidence is genuinely split, and both sides are credible. Forbes, Entrepreneur, Fast Company and CX Dive all report a reversal after satisfaction fell on complex cases, with the CEO conceding “we focused too much on efficiency and cost.” But Klarna itself disputes the framing: “Klarna never eliminated human support. We still work with several thousand outsourced agents. The current pilot involves just two new agents… it’s an addition, not a rehire or reversal.” It has since moved to a gig-workforce model — neither replaced nor restored. Two details do settle: the press release is February 2024, not 2023, and the 700 figure was an FTE-equivalent capacity claim, not a headcount fired.R4, P1
🔴 The part that matters for us is not in either version. Klarna could absorb a quality dip because it had a floor of several thousand outsourced agents. SatuSatu, with one concierge queue, has no floor — so the downside here is larger than Klarna’s, whichever account is true.
03 · Competitor benchmark table
The benchmark to copy is a pattern—not a company.
This is an exhaustive register of the named shipped, beta, announced, and observable artifacts found in the research—not a winner ranking. Every artifact links to the cited release, news item, official product, or full local evidence page. “Maturity” follows the research scale: 0 absent, 1 rules, 2 classical ML, 3 GenAI-assisted, 4 agentic. “—” means not scored on comparable evidence.
Read column-wise before company-by-company
Seven value-chain zones reveal two category-wide holes.
Seven companies, not ten: Traveloka TPN, TBO Holidays and Prioticket/ETG are absent because the research assigns them a single overall score and never produced a per-stage assessment — inventing one here would imply a precision the evidence does not have. Highest verifiable score in each zone. “?” means the surface was not inspected, or no public artifact was sufficient to score it — it does not mean zero. An italic band such as 2–3* means the company claims capability in a dated release but never names the feature or publishes a metric: the band is claimed, not demonstrated.
| Company | Supply & catalog | Availability & pricing | B2B partner surface | Connectivity API | Discovery & planning | Service & in-destination | Internal ops |
|---|---|---|---|---|---|---|---|
| GetYourGuide | 3 | 1 | 1 | 1 | 3 | 1 | 3 |
| Klook | 3 | 1 | 3 | 1 | 3 | 3 | 3 |
| Viator / Tripadvisor | 3 | 1 | 1 | 1 | 3 | 3 | 3 |
| KKday / rezio | 1 | 1 | 1 | 1 | ? | ? | ? |
| Headout | 2–3* | 2–3* | ? | 1 | 1 | 2–3* | ? |
| GlobalTix | 0 | 1 | 1 | 0 | ? | 1 | ? |
| SatuSatu today | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
* An italic band is claimed, not demonstrated. Headout is the only company in this state: a dated 2025-01-15 release asserts “AI-powered customer support” and “smarter inventory and pricing systems”, and the research scores those stages 2–3 on that basis — but no feature is named and no metric is published, and the public API repository has not been touched since July 2024. Treat the band as the company’s own claim, not as verified capability.
Klook is the closest named case and remains closed beta. Viator’s AI-first evidence is bounded to its owned consumer surface.
The strongest verifiable score in this zone is 1. Models do not manufacture truthful capacity, commercial terms, or confirmation state.
No natural-language query, semantic-search parameter, MCP, or agent manifest was verified in the inspected partner surfaces.
Source: detailed competitor matrix, Klook/GYG teardown, Viator/KKday/Headout teardown, and connectivity teardown.
| Company | Lens | Strongest shipped / observable evidence — linked | Maturity | B2B / transaction reality | Reusable lesson for TipTip / SatuSatu | Evidence call | Cited references |
|---|---|---|---|---|---|---|---|
| Klook | Global + SEA |
|
3 | No public outbound distributor API found; AI remains in Klook-controlled discovery and checkout. | Compress content and turn demand-side feedback into supplier action. Do not copy a generic planner as the wedge. | Shipped + BetaPublic evidence is mainly official co-marketing and interviews. | |
| GetYourGuide | Global |
|
3 | Partner API is active but conventional. GYG owns the catalog record and spec, then lets connectors implement it. | Make unstructured supply sellable, keep the catalog as system of record, and partner for connectivity. | ShippedStrongest public engineering evidence in the set. | |
| Viator / Tripadvisor | Global |
|
3 | Mature 33-endpoint partner API and certification; no material AI exposed in the partner surface. | Discovery can move into assistants while booking remains a tested, auditable transaction surface. | ShippedConsumer metrics are stronger than partner-AI evidence. | |
| Pelagoby Singapore Airlines | SEA curated experiences |
|
3 | Consumer OTA; no public B2B distribution surface assessed, so the pipe is unscored rather than zero. | 🔴 The closest SEA comparator to SatuSatu, and it was missing from this pack until 2026-07-27. Curated activities, premium brand, human escalation retained — and an AI-native WhatsApp service surface already live. It also falsifies our earlier refutation of “WhatsApp-first SEA players” as a category with “zero named consumer-facing companies.” | ShippedMetrics are vendor-published; architecture is first-party. | |
| Eventbrite | Global ticketing |
|
— | Ticketing platform, not an activities distributor. Its AI is organiser-facing content tooling, so it maps to neither the traveller storefront nor the B2B pipe. | 🆕 The ticketing-side analogue this pack was missing — and the most direct benchmark for TipTip’s organiser tools. Every one of its three shipped features corresponds to a TipTip organiser job. ⚠️ But read the ending: shipping organiser AI first did not secure independence, and the team that built it was largely cut post-acquisition. | Shipped / AcquiredAI direction post-acquisition is now a real unknown, not a rhetorical one. | |
| KKday / rezio | SEA infrastructure |
|
1 | Operator system of record for product, prices, sessions, orders, redemption, and channels. | The valuable asset is first-party operational state, not the SaaS fee or an AI label. | Rules-led | |
| Headout | Global |
|
1–2 | Public API documentation was stale in the research cut. | AI narrative and published docs do not prove a maintained product. | Claimed | |
| GlobalTix | SEA infrastructure |
|
1 | Strong reseller and supplier rails; prepaid credit, nett pricing, price floor, FX markup. Supplier and competitor to D2/D3. | Use for Pool A coverage, not defensibility. The same rail can digitise “exclusive” Pool B operators. | Announced / Rules-led | |
| Traveloka TPN | Indonesia + SEA |
|
— | Travel activities are a named B2B line; API, mini-app, redirection, and announced travel-agent booking engine. | The near-term threat is workflow and pooled supply, not an AI feature. Win a differentiated Bali line item. | Commercial proof | |
| TBO Holidays | Global B2B / SEA |
|
— | Claims 200,000+ sightseeing products and 25,000+ agents; broad wholesale workflow already exists. | Do not try to win the agent’s platform login with 253 SKUs. Test Pool B as a high-value line item. | Commercial proofActivity depth by Indonesia remains unverified. | |
| Prioticket / ETG | Neutral infrastructure |
|
— | Public Distributor API, OCTO-native posture, independent of competing OTAs. | Strong second-aggregator candidate if Bali depth is proven; standards reduce vendor lock-in. | Shipped railsIndonesia depth is the gating unknown. | |
| TipTip / SatuSatu baseline | Group context |
|
0* | SatuSatu has a live catalog, GlobalTix feed, Pass, and human WhatsApp queues; D2/D3 are planned. | Reuse production discipline and people—not the event forecasting model, which solves a different demand shape. | No model transfer*SatuSatu product baseline only. |
Showing all 12 benchmark rows.
04 · AI innovation pattern summary
Eleven patterns, graded by whether TipTip and SatuSatu can actually use them.
This reconciles the Opus synthesis with the evidence tiers and current product constraints. Each verdict is a workshop input—not a pre-approved roadmap decision—and the order follows prerequisite logic rather than novelty.
Supplier-side GenAI catalog production
Shipped · quantifiedGetYourGuide’s 16-step supplier wizard auto-completes eight key steps. The research records approximately 14-minute creation, a failed first experiment caused by UX and trust, and eventual 100% rollout with human approval retained.
Pilot now · buy before build. Process 50 Pool B listings from supplier-owned source material with human QA. Continue only if edit time falls ≥30% and no customer-visible content defect escapes.
Own-content GenAI localisation
ShippedKlook reports more than 80% lower production time across 4,200 destinations. This is Klook-owned content inside a controlled workflow—not supplier onboarding and not transaction state.
Conditional on demand and checkout. Validate target-language demand, currency, and payment readiness together; translation alone does not create value.
Retrieval quality as the AI product
ShippedGetYourGuide documents hybrid lexical and semantic retrieval, production ranking, transformer ranking, cold-start intelligence, review summaries, and AI search. The mechanism compounds only where query and outcome volume exist.
Instrument first. Add query logging, a synonym dictionary, and weekly zero-result review before funding a ranking model or flagship planner.
Demand exhaust → supplier feedback
Shipped pilotKlook’s review-summary pilot turns customer feedback into merchant action. One cited merchant created a product that later represented 24% of that attraction’s revenue; the reusable asset is the learning loop, not the summariser UI.
Strategic target after instrumentation. Label complaints, edits, refunds, incidents, and search gaps now; expose supplier insight only when the corpus produces stable signals.
AI-first storefront as acquisition strategy
AI-first · boundedTripadvisor’s AI-native consumer MVP combines personalisation, recommendations, GeoAware suggestions, offers, planning, and pre-booking chat; it scaled to half of English-market web traffic. The evidence also links it to direct traffic, repeat use, and conversion.
Not a first 90-day bet. Establish demand, query, funnel, and booking baselines. Validate one narrow journey change instead of copying a horizontal planner.
Distribution into the assistant
Shipped channelViator/Tripadvisor and Klook can surface inventory inside AI assistants, but booking and payment still resolve through merchant-owned rails. The assistant becomes a channel; it does not replace catalog, availability, or checkout truth.
Minimal, measurable preparation. Keep Pool B entities and schema machine-readable, preserve source rights, and measure referral quality. Attach no traffic claim until observed.
AI-assisted onboarding conversion
Attribution unresolvedTripadvisor reported AI-assisted sign-up more than doubled conversion, but the passage does not cleanly isolate Viator supplier sign-up from TheFork restaurant sign-up. The mechanism is plausible; the segment is unresolved.
Fold into pattern 01. Compress internal operator onboarding with extraction and drafting. Do not create a separate portal workstream or transfer the conversion figure.
Internal operations and engineering AI
ShippedKlook reports code-assistance and internal GenAI gains; Tripadvisor reports a large increase in one engineering pilot; GetYourGuide publishes both adoption activity and junior-engineer guardrails. These are capacity patterns, not customer propositions.
Adopt with evaluation. Reuse group controls, review discipline, monitoring, and learning. Measure returned cycle time and escaped defects; do not transfer the forecasting model.
Render, never generate
Control patternExecution instructions, refund terms, confirmation, entitlement, and meeting points should be rendered from confirmed records. Generative prose is least appropriate where a fluent error can strand a traveller or create liability.
Fund as a control. AI drafts and summarises; templates render day-of truth; people own disruption, safety, health, culturally sensitive, and ambiguous recovery.
The plumbing inversion
Structural lessonGetYourGuide owns the catalog record and integration specification while partnering for connectivity. KKday/rezio shows how operator workflow and first-party state can be more valuable than the SaaS fee.
Own truth; partner for rails. Structure Pool B data, adopt OCTO where justified, test Prioticket/ETG depth, and contract non-overlapping territory before building dedup.
The Pass is a merchandising wrapper—not an AI moat
Not AI differentiationKlook’s Bali Pass shows that bundles are packaging over inventory. Purchase, activation, and redemption are separate events; economics depend on supply mix, validity, breakage, and redemption—not model sophistication.
Protect Pool B economics. Measure activation, redemption, breakage, margin, and concierge workload before presenting the Pass as defensible innovation.
The pattern summary in five lines
- The most transferable shipped-and-quantified GenAI pattern is supplier-side catalog production.
- AI-first storefronts are scale-dependent demand and acquisition plays, not universal product templates.
- The B2B AI gap sits on top of a more basic constraint: missing or unreliable availability, commercial, and confirmation state.
- Much of SatuSatu’s highest-value near-term work is deterministic: rate integrity, refund rendering, booking chase, sold-count separation, entitlement, and day-of truth.
- Buy or partner is the default while dedicated capacity is absent; build carries the burden of proof.
Extract supply
Brochure, website, form, or message becomes candidate fields.
Verify the promise
Inclusions, rights, rate, refund, location, and capacity model.
Publish canonical state
One catalog record, availability model, and commercial terms.
Request + confirm
Booking state changes only from explicit system or operator acknowledgement.
Learn from outcomes
Summarise edits, incidents, reviews, and search gaps into the next action.
05 · What we already have
The constraint is not tooling. We have more infrastructure than this pack assumed.
This section comes almost entirely from the companion pre-read, and it changes several build/buy calls in the map that follows. The research was written from the outside and assumed a standing start; it is not one.P1
Automation and messaging rails
n8n for workflow automation, Respond.io on the messaging layer, Retool for internal tooling, and the TipTip Content Hub already holding organiser-relevant data. 🔴 This matters: three of the map’s “buy” calls become “configure what we own”, and the AI reply-assist pattern has a host already.
Individual AI adoption
Team members are already using Claude, vibe-coding and workflow automation — but individually and sporadically, with no shared toolstack or approved pattern. The gap is not appetite or access; it is that none of it is embedded in a product flow.
SKU onboarding is manual
Named by 5+ team members as a bottleneck. 🟢 This is the single most useful fact in this section — it converts pattern 01 from an inference off GetYourGuide’s engineering blog into a confirmed internal problem with a known owner and a measurable baseline.
One number to handle carefully
The companion pre-read cites 10 million registered users as a behavioural-data asset. It is not in our research corpus, so it needs one internal confirmation — and more importantly, registered users and transacting users are different denominators. This pack’s central caution is that monthly bookings have never been counted. A large registration base does not answer that, and a recommendation engine trained on the wrong denominator will learn the wrong thing. Confirm both numbers before either is used to size anything.P1, R8
Applied classifier · adopted from the PM pre-read, with control mode added
Our own tasks, classified before the workshop rather than during it.
The middle column is the PM’s call. The right-hand column adds what mode the work actually runs in — and in four of the eight rows the honest answer is deterministic, not AI at all.
| Task / process | Mode | Rationale | Control mode — what actually runs it |
|---|---|---|---|
| SKU description generation | Automate | Well-defined inputs, high volume, easy to QC. A human reviews rather than writes. | AI-assisted, with a named verification gate. This is pattern 01 and the strongest-evidenced move in the pack. |
| CS FAQ responses booking, T&C, refunds | Automate | Repetitive, well-defined answers; low error cost provided an escalation path exists. | Mixed. Policy answers should be rendered from the record, not generated. Only genuinely open questions need a model. |
| Finance reconciliation | Automate | Rule-based, high volume, errors detectable via exception flagging. | 🔴 Deterministic. This is a rules engine with anomaly thresholds. Calling it AI would misprice both the effort and the risk. |
| Ticket pricing strategy | Augment | Needs market context, partner relationships, risk tolerance. AI informs; a human decides. | AI-assisted, human-in-loop. Note the category-wide finding: nobody in the benchmark set has shipped AI into pricing. We would be first, without a reference implementation. |
| Organiser daily performance summary | Automate | Data pull plus templated narrative. Informational, not decision-critical. | 🔴 Deterministic for the numbers, AI-assisted for the prose. Keep the figures rendered; let a model write only the wrapper. |
| Event investment decision | Augment | High-stakes, contextual, relationship-dependent. AI provides signals; leadership decides. | Human judgement, with AI-assisted signals. This is the decision provocation 3 would eventually put capital behind — keep it human until the forecast has a track record. |
| Marketing copy for campaigns | Augment | AI drafts; a marketer shapes tone, brand voice and cultural fit. | AI-assisted. Lowest blast radius on the list — a bad draft costs review time and nothing else. |
| Test case generation (QA) | Automate | Well-defined specifications, high volume, deterministic outputs. | AI-assisted. Returns engineering weeks, which is the binding constraint — so it outranks its apparent size. |
Classification adopted from the companion pre-read; the control-mode column is added here. Four of eight rows are deterministic in whole or in part. That is the single most common way an AI roadmap loses credibility with a CFO — presenting rules work as model work — and it is cheap to avoid by labelling the mode.P1
06 · Decision map · TipTip.id / SatuSatu.com
Where AI may act, where records must decide, and where a person stays accountable — stage by stage.
This replaces a SWOT. The source research explicitly rejected SWOT as its analytical engine and used a value-chain × AI-capability map with build/buy/partner filters and a red-team kill pass instead. A SWOT cannot express a prerequisite, and the pack’s central finding is a prerequisite: availability and confirmation state gate the entire B2B strategy.R5, R8
Must stay deterministic
The correct output follows from a record, contract or explicit state transition. A model adds risk without adding information.
Where AI can assist
Extraction, retrieval, drafting or constrained composition, verifiable before use. AI proposes; a rule or a person commits.
Human-owned
Safety-adjacent, culturally sensitive, ambiguous, exceptional or relationship-dependent. Autonomy here is priced in blast radius, not fluency.
| Value-chain zone | Must stay deterministic | Where AI can assist | Human-owned | Build / buy / partner | Bet it feeds | Gate or kill criterion |
|---|---|---|---|---|---|---|
| Supply & catalog | Canonical record per attraction; content rights; one declared source of truth per SKU; separation of inherited supplier counts from own bookings. | The strongest evidenced pattern in the set. Extract and draft listings from supplier-owned brochures, sites, forms and messages. Pattern 01. | Verify inclusions, terms, rights, location and capacity model before publish. GetYourGuide retains an expert approval gate at 200k SKUs. | BuyOff-the-shelf model; an ops habit, not a build | Pool B Supply Intelligence & Confirmation OS | Kill if escaped-defect rate on AI-assisted listings exceeds 2× manually authored on a paired 15/15 comparison of the next 30 Pool B listings. Manage escaped defects, never auto-pass rate. |
| Availability & pricing | Everything. Declared availability model per SKU, booking state machine, confirm/decline evidence, rate integrity, refund rendered from the supplier field. | Parse operator replies at high precision and flag ambiguity for a human. Never let a parse declare capacity. | Operator negotiation; any judgement about whether a departure can actually run. | BuildNo vendor can supply your own state | Pool B Supply Intelligence & Confirmation OS | Prerequisite, not a bet. No partner-facing surface ships until every Pool B SKU carries a declared availability model. This zone gates the four rows below it. |
| B2B partner surface | Net rate, allowed markup, FX exposure, cancellation terms, confirmation SLA, prepay ledger. A quote that cannot be honoured is worse than no quote. | Compose a draft multi-day quote and explain trade-offs — pre-booking, so an error costs a re-quote, not a wrong booking. | Agent relationship, exception service, commercial negotiation. | PartnerTest through an incumbent’s rails first | Differentiated Partner Yield Layer | Prove demand before building: rate sheet plus WhatsApp to 20–40 named agents. Kill the portal if fewer than three agents transact repeatedly. A portal is table stakes and confers nothing. |
| Connectivity API | Machine-readable products, prices, availability and order state; idempotency; certification-grade reliability. | Nothing yet. Every B2B surface in the benchmark set is pre-LLM — no AI endpoint, no semantic search, no agent manifest, at any of them. | Integration support and partner engineering relationship. | PartnerAdopt OCTO; never become the middleware | Differentiated Partner Yield Layer (later stage) | Do not start. All four conditions must hold, not any: exclusivity proven, a material share of Pool B instant-confirmable, named partners already transacting at volume, and a real multi-quarter budget. Middleware prices at a fraction of the volume it carries. |
| Discovery & planning | Price, availability, inclusions and cancellation shown at the point of choice. Search results may be ranked by a model; the terms may not be generated. | Narrow-intent retrieval, synonym handling, review summarisation, entity-rich content for assistant surfaces. | Curation judgement about what belongs in a Bali shortlist at all. | BuyRetrieval before ranking; no flagship planner | AI Discovery & Merchandising Loop | Instrument first. No ranking or personalisation model is funded until query logs exist and a zero-result baseline is measured. An AI-first storefront is a demand-acquisition play at a scale we do not have. |
| Service & in-destination | Day-of instructions, meeting points, entitlements and refunds rendered from confirmed records — never generated. A fluent wrong meeting point fails the job completely. | Draft itineraries, summarise traveller history, prepare context for the concierge before they open the thread. | All disruption recovery. Water and surf conditions, volcano status, scooter advice and medical or allergen guidance must not be AI-issued at all. | BuildThin: templates over confirmed fields | Human-supervised Travel Orchestration | Measure concierge workload, instruction-error rate and recovery time before funding anything. Containment is not a success metric — silence after a wrong instruction looks like resolution while the traveller walks the wrong way. |
| Internal ops | Audit trail, evaluation thresholds, monitoring, rollback. Whatever a model touches must be reviewable after the fact. | Code assistance, reconciliation triage, content and marketing operations. This is where the constrained resource — engineering weeks — is actually returned. | Ownership of the evaluation bar and the decision to ship. | BuyBorrow group controls and people | Shared AI Product Operating System | Measure returned cycle time and escaped defects. Transfer the production discipline, the people and the controls — never the event-forecasting model, because the demand shape differs. |
Rows use the same seven zones as the §03 value-chain matrix, so the benchmark scan and this decision grid read against one vocabulary. The five bets in column six are the opportunity areas previously listed separately — folded in here so each one carries its control mode, its build/buy/partner call and its kill criterion in one place. Impact is deliberately qualitative: no bet in this map has been sized, and the room should not read an unranked map as a ranked one.R5, R6, A1
Constraints and unresolved dependencies · these gate every cell above
None of these is an AI problem, and all of them can stop an AI bet.
No verified commercial owner
No source in the set records a B2B or business-development function, against a plan whose largest single line item is agent outreach. Hire or borrow one for a quarter, or drop the partner chain.
PII and liability unresolved
Whether traveller conversation data may reach an external model provider is binary and decisive — it flips build-versus-buy across much of the map. So is liability for a generated instruction. One counsel brief closes both.
The demand base is shrinking
Bali arrivals ran negative year on year in Jan–Apr 2026 while Indonesia grew. Several current-basis cases sit on a base that is contracting, and both volume and ticket value are under pressure at once.
Materiality at real volume
Monthly bookings, Pass economics and the cost base are unknown. At plausible low volume many of these opportunities do not clear a materiality floor — which is why measurement, not modelling, is the first commitment.
Our supplier is also our competitor
GlobalTix is simultaneously the Pool A feed, a licensed travel agent, and a direct competitor for the same agents and the same long-tail operators — and it can digitise the operators we call exclusive. It sets the clock.
Second-feed dedup debt
Integrating aggregator #2 “properly” is estimated at 8–29 engineering-weeks we do not have. A non-overlapping territory clause avoids the work entirely — one negotiation instead of twelve fixes.
07 · Provocations · adopted from the PM pre-read, gated here
01
02
03
04
Four bets that would make the business different, not faster.
Everything above this line makes an existing process cheaper or quicker. None of it changes what the business is. These four do — and they are reproduced here because the map is deliberately conservative and the workshop agenda asks for AI-first ideas it can argue with. They are unsized on purpose. Each carries the gate this pack’s research attaches to it, so the argument can be about evidence rather than ambition.P1
Instant booking on supply that owns no software
Strongest research agreementSatuSatu. Global platforms assume instant confirmation requires the operator to run a booking system. Most Indonesian local operators do not and probably never will. Build an AI agent that becomes the booking system on the operator’s behalf — confirming and settling through a WhatsApp or voice conversation in their own language, learning each operator’s patterns well enough to offer instant confirmation without asking every time.
🟢 This is the research’s own central finding, arrived at independently. The decision map puts availability and confirmation state as the prerequisite that gates the entire B2B strategy. The two documents agree here more strongly than anywhere else.
Gate. Exclusivity must hold first — if Pool B operators already self-list elsewhere, the moat is not there to build on. Then: declare an availability model per SKU before any partner surface ships. Do not let “AI confirms it” become a way to skip the state machine.
Stop buying demand. Become the rails other agents book through.
Qualified — timing contestedBoth products. Agentic commerce is framed as a threat: agents bypass platforms and erode distribution. The inversion is to stop defending and instead be the thing the agent transacts against — machine-callable Indonesian experience and event inventory, cleared through us because we hold the local supply, payment and settlement a foreign agent cannot assemble.
⚠️ Two research qualifiers. First, OpenAI deprecated Instant Checkout in March 2026 in favour of merchant-run apps — the agentic booking layer retreated rather than advanced, so the “coming fast” premise needs dating. Second, and more usefully: “SatuSatu, with 253 SKUs and no structured feed, is not in any model’s retrieval set.” The prerequisite is unglamorous and cheap.
Gate. The channel incumbents actually won was a commercial deal, not a schema attribute — Klook and Viator are first-party apps. Structured feed and entity-rich Pool B content are worth doing at near-zero cost; attach no traffic claim until observed. Do not fund this as an AEO project.
From a platform that takes a cut, to an engine that backs events
Highest risk — capital at stakeTipTip. Today the platform earns a percentage of events other people decide to run. If demand forecasting became confident enough, TipTip could selectively back or guarantee events it predicts will succeed — earning upside beyond a flat fee and pricing that risk with data no single promoter has. The role shifts from operating events to underwriting them.
🔴 The research is pointed here. The group’s one production model is event-business sales forecasting, and its headline margin claim is correlational and single-source — every Tier A/B source says “after” deployment, and the same release offers two non-AI explanations. Underwriting on a forecast whose accuracy has never been externally established is the highest-variance move in either document.
Gate. A controlled pilot, never a launch. Before any capital is committed: establish forecast accuracy on held-out events, and separate the model’s contribution from the take-rate pivot that ran at the same time. The prediction has to earn a track record before it earns a balance sheet.
The company a larger competitor cannot afford to match
Most within our controlThe operating model, both products. Not “add AI to each team” but a company-wide target: a cost-to-serve per unit of sales structurally lower than any competitor our size, with the team growing far more slowly than revenue. Support, content, onboarding, reconciliation, price monitoring and first-pass outreach run on AI or rules by default; people hold exceptions and judgement.
🟢 A lower cost-to-serve is not a saving — it is a pricing weapon. It lets us profitably serve thin-margin inventory and price-sensitive customers where larger competitors lose money at the same price. It also funds every other bet by freeing capacity, and it needs no partner and no market condition.
Gate. It is measurable or it is a slogan. Baseline cost-to-serve per booking before anything ships, then track the ratio quarterly against revenue. ⚠️ And note the mode: much of this list is deterministic work, not model work — four of the eight tasks in §05 are. Bill it honestly.
How to use these
These are provocations, not a shortlist — argue with them, break them, replace them. The job is not to pick a winner today. It is to decide which single bet we would most regret not starting, and to name the smallest experiment that would tell us whether it is real. 🔴 One honest caution: none of the four has been sized, and three of them depend on a booking base nobody has counted. That is an argument for measuring first, not for dropping the ambition.
Strategic posture
Use TipTip’s AI operating discipline to build a supervised Pool B learning loop for SatuSatu. Win on direct local supply becoming easier to publish, confirm, and distribute—not on a generic agentic travel interface. Every autonomy step must be earned by reliable state, measured volume, and a bounded blast radius.
Evidence notes, source map, and interpretation guardrails
Guardrails. Market maturity scores are directional, not a league table. Public evidence is much stronger for GetYourGuide than for private companies. Pool A and Pool B percentages use different profit bases and are directional only. No internal number should be invented to fill a data gap. The red-team memo withdraws earlier incompatible headline sums and makes measurement the first decision. Competitor examples are inspiration, not substitutes for observed customer or operational problems.
🔄 Reconciliation with the companion PM pre-read (2026-07-27). Both documents go to the same room, so every point where they differ is named here rather than averaged. Where the PM pre-read corrects us: GetYourGuide shipped a ChatGPT app with its Spring 2026 release, so our reading of that release as “a conversion layer, not an external agent” was incomplete; and Pelago — a named consumer-facing SEA company running AI-native WhatsApp service — falsifies our blanket refutation of “WhatsApp-first SEA players.” It also closes our own open question #20 on Traveloka’s consumer surface. Where we correct the PM pre-read: Klook’s April 2026 announcement is global premium sports (World Cup, Roland-Garros, Wimbledon, F1), not an Indonesian live-events move, so the TipTip threat is category adjacency rather than head-on; and the widely-circulated Traveloka “120M users / ~22% conversion” pair traces to a Tier C SEO domain this research refutes elsewhere, so it is excluded. Where the evidence is genuinely split: Klarna — see §02. Not independently verifiable: Klook’s “40% above industry average” conversion and its “2025 personalisation engine” (the Google Cloud materials carry no conversion metric); the “37% of UK ticket sales by 2028” agentic projection, which sits against the documented March 2026 Instant Checkout deprecation; and the 78%/62% OTA-adoption and “co-pilot +40%” figures, all vendor or consultancy research. A full side-by-side is in differences.md.
Full-source access. Every local citation below opens a complete HTML reading copy of the source file. The rendering preserves the original sections, tables, bullet points, caveats, and links; nothing is condensed for this view.
- M1 · Full content source. workshop-reference-codex-source.html — complete HTML rendering of the editable workshop reference source.
- A1 · Workshop agenda. AI-First Product Workshop — Preparation and Facilitator Guide — timing, session prompts, scoring weights, required outputs, and post-workshop artefacts.
- P1 · Companion PM pre-read. AI-First Product Workshop pre-reads 01–05 (Google Doc, product management). Source for the Plus-AI/AI-Plus framing, the 10× capacity frame, automate-vs-augment, the SEA benchmark, the TipTip-side JTBD, the internal capability inventory, and the four provocations. Claims drawn from it are tagged P1 throughout and are company-reported or third-party unless separately verified — see the reconciliation note below for where the two documents diverge.
- B1 · BCG on AI-first. AI-First Companies Win the Future (Jun 2025); Design Your Company for AI, Not AI for Your Company (Apr 2026); Transforming Into an AI-First Payments Company (Apr 2026); How Companies Can Prepare for an AI-First Future (Oct 2025). Source of the five no-regrets moves. Consultancy perspective, not measured outcome.
- J1 · The jagged frontier field experiment. Dell’Acqua et al., Navigating the Jagged Technological Frontier, HBS Working Paper 24-013 → Organization Science (2025), doi 10.1287/orsc.2025.21838. 758 BCG consultants, 18 tasks, three arms. ✅ All four figures verified 2026-07-27.
- M1 · Mollick. Co-Intelligence: Living and Working with AI (Penguin Portfolio, 2024), and the BCG Henderson Institute podcast (Feb 2025). Source of the “tools are necessary but not sufficient” caution.
- E1 · Pelago. Yellow.ai case study (vendor-published metrics) and the Pelago tech blog (first-party architecture). Added 2026-07-27.
- E2 · Eventbrite. Eventbrite IR on the Bending Spoons agreement; TicketNews on the 10 Mar 2026 close. Added 2026-07-27.
- R1 · Detailed competitor matrix. research/02-competitor-matrix.html — shipped vs announced AI, B2B APIs, SEA distribution, connectivity, evidence asymmetry.
- R2 · Codex competitor synthesis. research-codex/02-competitor-matrix.html — the full source file containing the maturity scale, value-chain matrix, and reusable competitor implications.
- R3 · JTBD maps. research/07-jtbd-maps.html — traveller, B2B buyer, supplier, SatuSatu, and TipTip jobs.
- R4 · Operations scan. research-codex/03-ops-scan.html and detailed red-team version — concierge decomposition, deflection definitions, cost and safety.
- R5 · Red-team executive memo. research/06-exec-memo.html — measurement priority, materiality, deterministic slate, and risk. Its SWOT appendix is explicitly a summary wrapper: the memo states no conclusion derives from it, which is why §05 uses the value-chain map instead.
- R6 · Opportunity executive memo. research-codex/06-exec-memo.html — Pool B operating thesis and three opportunity bets.
- R7 · Landscape. research/01-landscape.html — Bali/Indonesia demand, channel economics, structural forces, and AI discovery timeline.
- R8 · Internal context. research/00-internal-context.html — knowns, unknowns, capacity constraint, Pool A/B/C policy, TipTip transfer boundary.
- R9 · Source ledger. research/99-sources.html — primary URLs, source tiers, raw evidence, and access dates.
- R10 · Klook and GetYourGuide teardown. research/raw/teardown-klook-gyg.html — complete AI-stage, API, connectivity, and Bali Pass evidence.
- R11 · Viator, KKday/rezio, and Headout teardown. research/raw/teardown-viator-kkday-headout.html — complete named-feature, API, ownership, and economics evidence.
- R12 · Indonesia connectivity teardown. research/raw/teardown-connectivity-indonesia.html — complete GlobalTix, aggregator, Traveloka, TBO, and Indonesian-agent evidence.
- R13 · SEA source ledger. research/raw/sources-group45.html — complete primary and secondary URLs behind the regional comparison.
- Chart method. Storytelling with Data guide to charts and graphs — chart selection reference; bars use a common baseline, direct labels, restrained color, and message-led titles.