If Your Moat Got 280 Times Cheaper in Two Years, It Was Never a Moat

September 14, 2026 · Vera.Support

Where value actually accrues in vertical AI — and the four conditions under which this argument is wrong, stated up front rather than discovered in diligence.

Stanford's AI Index reports that the inference cost for a system performing at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024. Over the same stretch, AI hardware costs declined around 30 per cent annually and energy efficiency improved around 40 per cent.

You can watch the same thing inside a single vendor's price list. A more capable model tier, released later, priced below its predecessor.

For anyone building on this capability, that is extraordinarily good news: the largest variable cost in the business falls every year without management doing anything.

For anyone whose business is having the capability, it is the opposite.

A capability whose price falls by two orders of magnitude in two years is an input. It is not an advantage, and any competitor can reach your level for a fraction of what it cost you.

Orchestration is not the moat either

The next answer people reach for is the layer above the model: prompt design, conversation flow, retrieval plumbing. This is real engineering, and it is not defensible.

It is weeks of work. It is documented in survey papers — a survey in Transactions on Machine Learning Research, with Yann LeCun among its authors, describes the whole pattern of models calling external modules as an established research direction rather than a proprietary technique. And crucially, it improves for every competitor simultaneously whenever a model ships.

The test is simple. Does this layer get better when somebody else releases something? If yes, it is not a moat.

What that test leaves standing

Apply it to the hospitality AI stack and three of four layers fail.

Foundation models: commoditising fast, by two orders of magnitude. Orchestration: already commoditised. Channel surfaces — WhatsApp, Instagram, web, voice: owned by platforms, governed by their terms, subject to their pricing. Rented, never held.

What is left is the layer nobody writes about: the authenticated, maintained connection into each property's own reservation system.

It does not commoditise, for a specific reason. No model release improves it. It is built per software vendor and frequently per property. It requires credentials and operational trust. It breaks silently when a schema changes. It is slow, local, and unglamorous — which is precisely why it accumulates.

Why it is slow, concretely

It is worth being specific, because "integration is hard" is the kind of claim founders make and investors discount.

Once per reservation-system vendor, you build: an authorisation flow and token lifecycle; schema mapping for room types, rate plans, inventory and occupancy rules; pricing semantics covering child-age brackets, minimum stay and multi-room distribution; a representation of the cancellation policy matrix; a safe write path for an actual booking; and an error taxonomy that can tell "no availability" apart from "the call failed."

Then, once per property: credentials issued by the hotel; network authorisation, where IP allow-listing routinely blocks foreign-hosted callers; that property's own rate plans mapped; its policies verified against what the staff actually tell guests; test bookings run through to the record; and a cutover with a human fallback running alongside.

Then, continuously and forever: schema drift, credential rotation, seasonal policy changes, version upgrades, and monitoring for the thing nobody instruments — whether the price the guest was shown matched the price in the system at that moment.

Only the first column amortises across customers. The second and third do not. We attach no duration or cost figure to any of it, because we have no sourced basis for one — the argument is about the count and nature of the steps, not their length.

Reliability, not capability, is the product

This is the part most technology theses get backwards.

τ-bench evaluates language agents against real API tools and written policy documents in customer-service domains, scoring by whether the final database state matches an annotated goal. Its published finding: "even state-of-the-art function calling agents (like gpt-4o) succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail)."

Read those two numbers separately, because they measure different things. pass@1 asks whether the agent can do the task. pass^8 asks whether it does the task eight times running.

A hotel is buying the second one. The gap between those two numbers is the entire product problem — and it is not solved by a better model.

WebArena reports a comparable picture for open web tasks: a best agent success rate of 14.41 per cent against human performance of 78.24 per cent. Both benchmarks tested earlier model generations and both will have moved; we cite them for the shape of the problem, not as current capability estimates.

Closing that gap is operational work: constrained action scope, validated arguments, refusal behaviour when the record is silent, escalation paths, and instrumentation that catches the failure before the guest does. None of it appears in a model release. All of it accumulates in a codebase and a set of live installations.

The demand side: what hotels currently spend money losing

Booking Holdings reported $26.9 billion of revenue on $186.1 billion of gross bookings in 2025 — a blended take rate of 14.5 per cent. Expedia Group reported $14.7 billion on $119.6 billion, or 12.3 per cent. Neither is pure accommodation commission, so read them as blended.

Meanwhile HOTREC's European distribution study, surveying 3,089 hotels across 35 countries, puts direct bookings at 50.9 per cent of overnight stays against 29.1 per cent for OTAs — down from 57.6 per cent direct in 2013. Over the same decade Booking Holdings' share of the European OTA market rose from 60.0 to 71 per cent.

One structural change is worth noting. Booking.com was designated a gatekeeper under the Digital Markets Act in May 2024, with full compliance required from November 2024. Parity clauses are prohibited, and the platform "must not introduce other measures with the same effect." Removing rate-parity constraints makes a direct channel economically expressible for the first time in years.

And a regulatory forcing function nobody designed

Three developments between 2024 and 2026 push the market toward systems that can show where their answers came from.

In February 2024 a tribunal found that Air Canada "did not take reasonable care to ensure its chatbot was accurate," applying an ordinary negligence standard and rejecting the argument that a chatbot is a separate legal entity. From 2 August 2026, Article 50 of the EU AI Act requires that people be told they are interacting with an AI system. And a field experiment across 6,200+ customers, published in Marketing Science, found that disclosure cut purchase rates by more than 79.7 per cent — because customers perceived the disclosed bot as less knowledgeable.

That 2019 figure will not replicate today, and we do not forecast it. The mechanism is what transfers. Compulsory disclosure plus a knowledge penalty leaves demonstrable accuracy as the only remaining lever — and demonstrable accuracy is a property of the integration layer, not of the model.

What our own numbers do and do not show

Across two hotel properties under common ownership, over roughly seven months, our deployment handled about 424,000 guest messages — 60,611 in a measured 30-day window, at 99 per cent autonomous resolution and an average 7.9-second first response.

Costed on the least favourable assumption available — top-tier plan plus an AI usage estimate computed for intensive use, long prompts and heavy tool calling — that runs to roughly $2,399 a month for both properties, or about $0.04 per guest message. The fixed platform line is small and far from its ceiling; the variable line is the one falling with the cost curve above.

The four conditions under which this is wrong

This thesis is published by a company operating at the layer it calls defensible. That is a conflict, not a disqualification, and the correct response is to check the falsifiable parts. So here they are.

1. Incumbent reservation platforms ship it natively. They are the system of record; their integration cost is zero. Watch for a major PMS or channel manager announcing native conversational booking against its own inventory, bundled at no incremental price. This is the primary risk, and nothing in our argument answers it.
2. Reliable generic agents plus a universal connector standard. If pass^k reliability approaches usable thresholds and a connector protocol reaches hospitality systems broadly, integration collapses from engineering into configuration. Currently pass^8 sits below 25 per cent in the published retail domain — but this is the fastest-moving variable in the whole thesis.
3. Channel platforms disintermediate. Meta disclosed on its Q1 2026 earnings call "more than 10 million conversations each week being facilitated through Business AIs, up from 1 million at the start of the year." If a messaging platform offers native booking against hotel inventory on its own commercial terms, the channel layer absorbs the value.
4. Demand does not materialise. Hotels may simply keep paying the take rate. Watch whether European direct booking share stabilises in the next HOTREC study now that parity clauses are prohibited. It fell 6.7 points over the decade to 2023; the post-DMA reading is what settles this.

And the risk that is ours rather than the market's: the operational evidence above comes from two properties under one ownership, with no control group and no independent audit. It shows the architecture runs at volume. It does not show that anyone will pay for it.

If you read only one section of the underlying report, read the risk register. It is the part that can actually be checked.