Adam Harris is making hotel operators uncomfortable. The distinction he draws is precise: the AI that closes a boardroom sale is not the AI that processes a guest complaint at 2 a.m. in three languages while preventing a chargeback. CEOs signing implementation contracts in mid-2025 are locking in architecture that will look as current as a fax machine lease by 2029, and Harris is betting most of them know it.

The gap is operational, not theoretical. Demo-grade AI handles the easy 80 percent—reservation confirmations, upsell prompts, simple concierge requests. Production-grade AI handles the 20 percent that generates litigation: billing disputes across currency zones, allergy cross-references with kitchen inventory systems, real-time route optimization for housekeeping staff when two rooms call in mechanical failures simultaneously. Harris notes that the $240 million hospitality operators spent on AI tooling in 2024 went largely toward the former. The spend that matters will target the latter, and it requires different vendors, different SLAs, and different liability frameworks.

What separates the two is error cost. A chatbot that misunderstands a cabana reservation costs $180 in comped amenities. An AI that mishandles a payment dispute with a guest holding Centurion status costs $12,000 in lifetime value and generates a compliance file that sits in legal's queue for six months. Harris is pointing at the contracts being signed now—12-month pilots, 24-month rollouts, vendor lock-ins with auto-renewal clauses—and asking which error profile they were built to manage. Most were built for the former. The guest base and margin pressure will demand the latter.

The timing matters because agent-on-agent commerce is no longer speculative. By Q4 2026, Harris expects 15 percent of luxury-hotel bookings to originate from AI agents negotiating directly with property-management systems, no human in the loop on either side. That requires API reliability at 99.97 percent uptime, semantic pricing engines that adjust for 40 variables per request, and fraud detection that distinguishes between a legitimate agent acting on behalf of a family office and a bot farm in Bucharest. The systems being deployed in 2025 were not architected for that load or that threat surface.

Operators should watch three forcing functions. First, the Q1 2026 chargeback data from properties that deployed conversational AI in guest-services roles. If dispute rates are flat or down, the vendor stack is sound. If they are up, the contracts have a flaw that will compound. Second, the Q2 2026 RFPs from the six largest luxury chains for multi-agent orchestration platforms. Those will set the capability floor for 2027. Third, the mid-2027 insurance-policy language around AI-generated guest communication. Underwriters will begin excluding coverage for systems that cannot produce auditable decision logs, and that will split the vendor field.

The decision signed in a May 2025 boardroom will be evaluated in May 2029 not against what the demo showed, but against what the system prevented at 2 a.m. Harris is counting on operators to feel that gap now, while the contracts are still in redline.