**Claim:** LLM-driven booking conversion rates will exceed 20% within 18 months for companies with clean data architectures, while remaining below 5% for companies with legacy systems — creating a measurable 'AI readiness gap.'
Verdict: Partially Supported
**Confidence:** Low
---
Supporting Evidence
- **MICE baseline conversion is demonstrably low (7–15%)**, confirmed by a primary source. Joonas Ahola (MeetingPackage), drawing on data across "thousands of hotels," puts current MICE conversion at 7–15%. This is a credible, practitioner-cited baseline from a live session transcript.
- **Automation already achieves 60–70% conversion in deployed systems.** Ahola explicitly states: "Automation 60–70% conversion." This is a primary source figure for a partially-automated (not full LLM) workflow, and it far exceeds the 20% threshold in the hypothesis.
- **Instant booking achieves ~90% conversion (with ~10% cancellation rate).** Ahola clarifies that instant booking redefines conversion: "It's technically 100% conversion, but there's a 10% cancellation." This is the ceiling benchmark for fully automated flows.
- **The core bottleneck is architectural, not motivational.** The Research Memo (Hospitality Track) corroborates Ahola: three structural barriers are identified — disconnected systems lacking unified data lakes, a relationship-driven culture resisting automation, and a false assumption that customers prefer human interaction. These barriers map closely to the hypothesis's "legacy systems" framing.
- **Legacy architecture explicitly undermines AI performance.** The Business Travel Track Research Memo names the problem directly: "Legacy TMCs that cannot control the full search-book-pay-report stack will continue losing ground... The IROPS servicing failure is specifically a legacy architecture problem." This supports the directional claim that legacy systems produce worse outcomes.
- **Clean data architecture is recognized as a prerequisite for AI disruption.** Marianna Evenstein (primary transcript, "Critical Success Factors") states: "If you want to be really a disruptor on the AI space, then you need the data... foundationally the big OTAs actually struggle a lot." This is a primary source directly linking data infrastructure quality to AI competitive positioning.
- **Two-year paralysis observed in real clients due to poor data.** A speaker in "Winning the Decision Layer" (Sanjay Vakil / Andrew Boch panel) notes: "we definitely have talked to clients who started with the conversation of like we can't go live with you for two years because our data is in such a terrible position." This is anecdotal but primary, and it illustrates the practical gap.
- **Green-field AI-native architecture is argued as the only viable path for genuine personalization.** André Rangel de Sousa (Research Memo: Tour Operator Track) "argued categorically that legacy data architectures cannot support genuine AI personalization and that the only viable path is greenfield AI-native construction." This directly supports the structural "AI readiness gap" framing.
---
Contradicting Evidence
- **The 20% LLM-specific conversion threshold is not cited anywhere in the evidence.** No speaker, no memo, and no transcript references 20% as a target or achieved outcome for LLM-driven booking. The hypothesis uses a precise figure that has no empirical anchor in this dataset.
- **The "18 months" timeline has no supporting evidence whatsoever.** Not a single source references a timeline for LLM conversion uplift reaching 20%. The Research Memo (Future Track) explicitly pushes autonomous AI booking to "2027–2028," which is 2+ years from ITB Berlin 2026 (March 2026) — extending well beyond the hypothesis's 18-month window.
- **The high conversion figures (60–70%, 90%) are not LLM-driven.** Ahola's figures pertain to automation and instant booking — i.e., pre-configured rules, structured data systems, and online booking engines, not LLM agents. The hypothesis specifically names "LLM-driven" conversion, but the evidence base conflates full automation, LLM-assisted workflows, and instant booking without disaggregating by AI method.
- **A contrary view exists that ecosystem integration on legacy infrastructure is viable.** The Tour Operator Track memo notes a direct rebuttal: one panelist argued "that ecosystem integration — even on older infrastructure — is itself a competitive moat when AI is layered strategically." This challenges the clean binary of clean-data vs. legacy-failure.
- **The 5% floor for legacy systems is equally unevidenced.** No source quotes a conversion figure at or below 5% for legacy-encumbered operators specifically. The 7–15% industry baseline is cited for the MICE sector broadly, not segmented by data architecture quality. The hypothesized floor is therefore unverifiable with this evidence.
- **"Don't rebuild entire systems" advice undermines the clean-data prerequisite.** Sanjay Vakil in the Distribution panel states: "I don't think it makes sense to rebuild entire systems today because we don't know what's going to win." This directly pushes back on the idea that a clean architecture prerequisite will drive a measurable gap within 18 months — companies are being advised to layer, not rebuild.
- **50% of MICE deals are lost before any technology intervention.** Ahola identifies that half of all MICE requests are turned down by hotels pre-proposal, not because of system quality but because salespeople self-disqualify. This human-behavioral failure mode is not addressed by LLM or data architecture improvements alone, complicating the conversion uplift thesis.
---
Nuance & Context
The hypothesis bundles three distinct claims that must be assessed separately:
1. **LLMs will drive conversion above 20% for well-prepared companies.** The directional case is plausible — automation clearly lifts conversion from 7–15% toward 60–70% in deployed systems. But the evidence does not isolate LLM's contribution from rules-based automation, and the 20% threshold is not derived from any cited source. It may actually be too conservative given Ahola's data, or it may be arbitrary.
2. **Legacy systems will keep conversion below 5%.** This specific floor is unverified. The current industry baseline of 7–15% already includes legacy-encumbered operators. The claim requires that LLM adoption widens the gap to the point where legacy companies regress below their current baseline — a strong claim with no evidentiary support here.
3. **This gap will be measurable within 18 months.** The Future Track memo's explicit "2027–2028" frame for autonomous AI booking, combined with the "don't rebuild" advisory from Vakil, suggests the 18-month window is optimistic. Structural change of the kind hypothesized — not incremental tooling, but measurable conversion bifurcation across a sector — would require both rapid LLM adoption and rapid data architecture remediation by the "winners," neither of which is confirmed.
A critical definitional problem runs through all the evidence: "conversion" means different things in different contexts. Ahola explicitly flags this — instant booking conversion is not the same as RFP lead conversion. The hypothesis does not specify which conversion metric it refers to, making verification harder and making the 20% figure's significance ambiguous.
The evidence is also heavily weighted toward the MICE/hospitality subsector. The claim is stated as a general travel industry thesis. MICE has sector-specific dynamics (relationship-driven sales culture, email bottlenecks, manual proposal workflows) that make it both a strong use case for automation and a poor proxy for transient hotel bookings, flights, or tour operators.
---
Key Data Points
1. **7–15%**: Current MICE booking conversion rate (Joonas Ahola, MeetingPackage — primary source, across thousands of hotels) 2. **60–70%**: Conversion rate achieved with automation in deployed MeetingPackage workflows (Ahola — primary source, same dataset) 3. **~90%**: Conversion rate for instant booking flows; ~10% cancellation post-confirmation (Ahola — primary source) 4. **50%**: Share of MICE deals turned down by hotels before any proposal is sent — the single largest conversion failure point (Ahola — primary source) 5. **2027–2028**: Estimated window for autonomous AI booking to be trusted by consumers (Research Memo: Future Track — synthesized analysis, not primary)
---
Assessment
The evidence provides a credible structural foundation for the "AI readiness gap" concept, but not for the specific quantitative claims in the hypothesis. Ahola's MeetingPackage data is the most useful primary evidence in the set: it establishes a wide spread between unautomated (7–15%) and automated (60–70%) MICE conversion, and clearly links that spread to whether systems support structured data and rule-based automation. The directionality of the hypothesis is right — data architecture quality matters enormously for conversion outcomes. However, the 20% figure appears to be underspecified; the available evidence suggests that automation (not necessarily LLMs specifically) pushes conversion far past 20% when fully deployed, making 20% a weak threshold rather than an ambitious one.
The weakest part of the hypothesis is the LLM specificity. Every high-conversion example in the evidence — MeetingPackage's 60–70%, instant booking's ~90% — is either rules-based automation or structured online booking, not LLM-driven. Felix Undeutsch's "Beyond the Buzz" session demo of AI-driven group booking is the closest to LLM evidence in the dataset, but the transcript excerpt doesn't provide conversion numbers. The claim essentially rests on an inference that LLMs will unlock or accelerate outcomes already observable in pre-LLM automation, which is plausible but unproven in this evidence base.
The 18-month timeline is the claim's most exposed flank. The Future Track memo — a synthesized research document, so secondary evidence — explicitly places consumer-trusted autonomous booking at 2027–2028. André Rangel de Sousa's categorical position that legacy architecture requires greenfield replacement (not incremental patching) implies the "winners" of this gap would need to have already largely completed their data infrastructure work. That the Distribution panel advises against rebuilding entire systems, citing uncertainty about which approach will win, further undermines the 18-month window. The gap may be real, but it is more plausibly a 3–5 year structural divergence story than an 18-month measurable bifurcation.