The Agent-Era Supply Problem

The previous essay ended on a simple observation. A great deal of what looks like industry connectivity functioning at scale is, on closer inspection, the customer quietly absorbing the gaps between what the infrastructure actually delivers and what the interaction requires. The traveller re-enters a hotel address. Accepts that availability on the booking page was wrong and picks a different time. Clarifies by email that the pick-up point moved. Confirms the voucher is legitimate by calling the supplier. That absorption has never appeared on any balance sheet, but it has been one of the industry’s most important operating subsidies.^¹

Agentic AI is the first interaction model that refuses to pay this subsidy. An agent does not re-enter the address. It does not accept that the time was wrong. It does not call. It queries a surface, receives a response, and either acts on the response or does not. Whatever the human was quietly compensating for is now exposed, because the new party at the other end of the transaction is not absorbing anything.

This essay asks the obvious follow-on question. If agentic booking becomes commercially material, what does the experiences industry actually have to offer it? The answer is uncomfortable, and it is not primarily a story about AI. It is a story about supply.

What an agent actually needs

Strip away the interface narrative and an agent needs four things from a supply side in order to transact reliably on behalf of a user.^²

Machine-readable product data. Not a marketing page, not a PDF, not an image with embedded text. A structured description of what the product is, what it includes, how long it takes, what restrictions apply, and how it is priced, updated when the supplier changes any of those attributes.

Real-time availability coherence. The answer to ‘is this bookable at 09:00 tomorrow for two adults’ has to be authoritative at the moment the question is asked. Not eventually consistent across a cache that refreshes overnight. Not approximately correct most of the time. Authoritative now, or an explicit signal that it cannot be determined.

Deterministic transaction completion. When the agent commits to a booking, the booking either succeeds and is acknowledged as a unique, unambiguous reservation the supplier is aware of, or it fails and the agent knows it failed. What cannot happen is the ambiguity the industry has long tolerated, where a confirmation is sent, the supplier’s calendar shows something else, and the mismatch is resolved on arrival by the person at the desk.

Commercial reachability. The agent has to be able to reach the supplier at all, by means that do not require a human on either end. A supplier whose only route to market is a listing keyed by hand into a distributor extranet is, for an agent acting on behalf of a user, functionally unreachable.

Four requirements. None of them are exotic. A hotel-chain interface has offered all four for the better part of two decades. A flight-booking interface has offered all four since the GDS layer was externalised to consumer-facing booking.^³ The question is whether the experiences industry, as an industry, offers them.

What the experiences industry actually offers

A product page can be visible online, persuasive to a human reader, well-ranked in search results, and still not satisfy any of the four requirements above. Visibility and executability are different problems. The honest answer is that it offers each of the four inconsistently at the top of the supply curve and not at all through most of the long tail.

Product data. Fragmented across catalogues that are, for a substantial share of suppliers, maintained by hand inside distributor extranets rather than pulled from a supplier reservation system.^⁴ The same supplier can appear with different inclusions, different cancellation terms, and different titles on different distribution surfaces, because each surface’s catalogue was keyed separately.

Availability. Correct authoritatively only at the supplier’s own reservation system. Everywhere else, it is a copy that is more or less fresh, held in channel-manager queues and distributor caches that were designed to refresh fast enough for human browsing, not to satisfy an agent acting on a deterministic query at an arbitrary moment.^⁵

Transaction completion. The place the gap is most visible. A booking made by a human through a distributor has a complete path to the supplier’s calendar only in the subset of cases where the supplier is integrated in real time. In the rest, the confirmation is sent, a secondary process moves it toward the supplier’s calendar later, and the traveller’s standing at the supplier’s front desk is, in effect, a well-formed query about whether those two systems agreed.

Commercial reachability. A four-way asymmetry. The operators that sit inside direct operator-to-distributor integrations are reachable. The operators that sit inside a channel manager are reachable. The operators whose only connection to the distribution graph is an account someone logs into once a week to update prices are, for an automated counterparty, not reachable at all.

Four structural mismatches

Placing the four requirements next to the four realities yields four structural mismatches, and they are not the sort that a clever interface layer can paper over.

Catalogue-first where the agent needs query-first. Catalogues are assembled ahead of time, in a distributor’s staging systems, to support a browsable surface. An agent does not browse. It asks specific questions against specific attributes and expects the answer to reflect the supplier’s current state, not last week’s staging copy.

Eventually consistent where the agent needs real-time authoritative. The industry’s connectivity layer was built to propagate calendar changes fast enough that human-speed browsing rarely sees a stale number. It was not built so that a machine, querying at any moment for a specific combination of date, time, party composition and product, receives an answer guaranteed to agree with the supplier’s own system at that instant.

Loose where the agent needs deterministic. The industry has tolerated a confirmation protocol where the booking is considered ‘good enough’ and residual ambiguities are resolved at the moment of arrival. That tolerance has been sustained by humans on both sides of the interaction who can resolve mismatches in real time. It does not survive the removal of the human.

Network-of-networks where the agent needs a coherent counterparty. The industry’s distribution graph is a sparse mesh held together by channel managers, manual extranets, and bilateral agreements of varying machine-readability. For a human shopper, this mesh is invisible because a distributor’s front-end hides it behind a search box. For an agent, the mesh is the problem. It has to reach every supplier it intends to transact with, and it has to reach them in the same shape.

The uncomfortable diagnosis

The temptation, reading this, is to conclude that agentic AI creates a new set of problems for the experiences industry. It does not. The four mismatches described above are not new. They are the conditions the industry has operated under for as long as it has been digital. What agentic AI does is remove the party that has been compensating for them, which is the human at the interface. Essay 7 named this directly. Human tolerance has been the hidden subsidy making the distribution graph appear to work. The agent is the first participant that refuses to pay the subsidy.

Which means the supply problem the industry is about to discover in the agent era is not a new supply problem. It is the one it already had. The tolerance of the human at the interface has been obscuring it, essay by essay of industry commentary, press release by press release, throughout the modern history of the category.

Who feels this first

The first-order effect falls on the long tail.

In the experiences industry, the long tail is not a minor phenomenon. It is the majority of the supply base. The overwhelming share of operators globally are independent, run by small teams, and occupy a position in the distribution graph that depends on intermediation.^⁶ Many of them are reachable to distributors only through an extranet that a human updates intermittently. This is not a failing of operators. The industry has not rewarded operators consistently for producing complete, structured, machine-actionable supply data, and many have had no rational reason to maintain it. For an agent acting on behalf of a user at commercial scale, those operators are, absent a structural change, functionally invisible.

The point deserves to be stated sharply, not as a neutral observation. A supplier unreachable by an agent is unreachable by a growing share of demand. Whether this becomes a long-run structural problem depends on whether the preconditions for agent reach can be met at the long tail, not only at the top. If they can, the industry converges toward a supply surface that is legible to an agent without flattening into a handful of megasuppliers. If they cannot, the only operators reachable by an agent will be those sitting inside channels already designed for machine-to-machine transaction, which is a much smaller population.

What happens by default

Absent a neutral infrastructure layer capable of presenting a coherent, machine-readable supply surface across the long tail, the agent era’s default outcome is a concentration effect. The reasoning is structural. An agent, at scale, will transact most reliably with whichever counterparty can satisfy the four requirements end-to-end. Where no neutral layer offers that, the position falls to the participants who have built their own. Those participants are the ones for whom demand and supply sit inside a single vertically integrated platform, because that is the only configuration in which product data, availability, transaction completion and reachability can be made coherent by fiat.^⁷

The infrastructure position defaults, in other words, to whoever can present the cleanest supply surface to the agent, and for structural reasons that is the participant whose commercial interests are not neutral to the outcome. This is a first-principles observation about what happens if nothing changes. It is not a prediction about any specific participant. The point is that the agent era has a default, the default is concentration, and the question of whether the industry ends up there is the question of whether a neutral supply layer exists to offer an alternative.^⁸

Close

The experiences industry has, in its modern history, optimised for attention. Interfaces, brand affinity, marketing funnels, conversion. The agent era does not care about any of these. It cares about a structured supply surface that is real-time, deterministic, and commercially reachable across the full length of the supply curve.

That surface does not currently exist at industry-wide standardisation. The tolerance that has masked its absence is being withdrawn. Who ends up holding the infrastructure position is, therefore, not really a question about AI. It is a question about supply, and about which participants are positioned to present supply coherently to a counterparty that will not compensate for incoherence on their behalf.

The next essay takes up the larger pattern this sits inside. When the interface commoditises, value does not disappear. It migrates. The migration has happened before, in search, in social, in cloud infrastructure, in retail distribution.^⁹ Where value goes when interfaces die is, in every previous case, the same place. The experiences industry is about to find out whether it has built the thing that place happens to be.