Essay Seven

AI Doesn’t Remove Infrastructure, It Reveals It

What looked like connectivity functioning at scale was humans absorbing the gap. That subsidy is being withdrawn.

The industry’s AI conversation is almost entirely about the consumer interface. Whether the interface is a generative discovery experience, a personalised search result, an agent booking a holiday on a traveller’s behalf, or a chat window answering a question about a tour, the debates in this sector converge on one layer: the thing the customer touches. “AI will replace OTAs.” “AI will transform marketing.” “AI will disrupt booking.” The three claims differ in detail. They share a shape. Each treats the consumer-facing layer as the place where something significant happens, and treats everything beneath that layer as out of frame.

The frame is too narrow. The structural implication of AI is not that the consumer interface is replaced by a smarter one. It is that the consumer interface itself, as a category, commoditises. When the interface layer becomes infinite and cheap, the thing that defines market position stops being the interface and becomes whatever the interface sits on.^¹ In every previous technology transition, the move has been the same. Value migrates from the layer that is becoming infinite to the layer that remains scarce.^² The layer that was hard to build before the transition becomes easy to build after it, and the constraint relocates.

The pattern, across prior transitions

Search did this to publishing. When indexing and retrieval became a cheap utility in the late 1990s, the scarce resource stopped being the content and started being the ability to match a query to the right page at scale. Publishers who had been the layer discovered they were now above the constraint, and much of the value migrated down to the search layer that had replaced their discovery function.^³

Social did a similar thing to attention. When the cost of distributing content collapsed inside the networks that became dominant through the 2000s, the scarce resource became attention allocation rather than content production. Publishers, the same category, found themselves displaced again, this time by the feed. The generating layer commoditised; the attention layer concentrated.^⁴

Cloud did this to operations. When provisioning, scaling, and managing infrastructure became a commoditised utility in the early 2010s, the scarce resource stopped being operational competence and became scale. The old constraint (running a data centre) was not removed. It was absorbed into a different layer, operated by different companies, under different economics.^⁵ In each of these transitions, the same move is visible. One layer commoditises; the layer underneath becomes load-bearing in a way it was not before. The constraint is not eliminated. It relocates.

What sits underneath, in experiences

The AI-era version of this rule is already in motion elsewhere in travel. Aviation has GDS clearing with a standardised schedule and ticketing vocabulary. Hotels have a central-reservation-system layer with more standardised interface conventions than experiences has developed, and a reasonably well-understood inventory model.^⁶ When AI commoditises the consumer interface in those categories, the layer underneath is reasonably well-formed. It will flex, but it is there. Experiences is the category where this is not true. What sits underneath the interface in this industry is the connectivity stack described in Essays 4 through 6. A booking crosses five to seven organisational boundaries. Each boundary carries its own translation between heterogeneous product vocabularies. The chain is latency-bound, translation-heavy, and missing a neutral settlement layer. The channel-manager category, the clearest partial answer the industry has produced, has been pushed to carry weight it was not built for.

If AI is the mechanism by which the consumer interface is commoditising across travel, then the place where that commoditisation meets a bottleneck in this industry is the connectivity stack.^⁷ Not because AI is doing something new to the stack, but because the stack was already under strain, and the strain was previously absorbed by something that will no longer be present in the same form.

What AI does not do

It is worth stating this part plainly before the next.

The boundaries. AI does not remove the six boundaries of Essay 4. An AI system checking availability still has to cross the distributor, connectivity, reservation-system, operator, and voucher boundaries, in some order, with some latency, across the same three or more organisations. Replacing a human user with a machine does not collapse the topology underneath. It changes who is at the edge of it.

The translation layer. AI does not remove it. Schemas do not converge because a faster reader is parsing them. The mapping between a reseller’s product vocabulary and an operator’s product vocabulary is a structural feature of fragmentation, not a feature of the reader. A faster reader with the same mapping reads the same compromises.

The settlement gap. AI does not remove this either. The absence of a neutral settlement layer, named in Essay 5 as a precondition of the aviation clearinghouse pattern, does not become less of a gap because an agent is placing the booking. Money still moves through the same peer-to-peer arrangements, negotiated out of band between the same participants. A machine at the edge does not change the plumbing behind it.

Whatever the headline version of the AI debate predicts the consumer interface will look like, none of the infrastructure beneath it is removed by the prediction.

What AI does do

What AI does is sharpen the cost of fragmentation. The connectivity stack has, until recently, been a sufficient answer to the industry’s connectivity problem because the edge of that stack was a human. Humans absorb unstructured input, resolve ambiguity across systems, work around schema mismatches, wait longer than they should, and re-read a page when the data arrived slightly wrong. A great deal of what has looked like industry connectivity functioning at scale is this absorption.^⁸ The gap between what the stack delivers and what the interaction requires has been held together, quietly, at the customer’s own expense.

A machine at the edge does not absorb ambiguity. Where a human reader encountering a product with inconsistent age bands across two schemas will pause, compare, and make a judgment call, a structured-data reader will either pick one and commit, or fail. Schema drift becomes decision drift. Product-vocabulary mismatch becomes unbookable inventory. Latency that was tolerated across a twenty-second checkout window is not tolerated inside a machine-to-machine call that expects to complete in a fraction of a second. The translation layer, the compromises it has made, and the inconsistencies it has been quietly accommodating, all become visible in a way they were not before.

This is most obvious at the agentic end of the AI spectrum, which is why it deserves to be named directly. An agent booking on a traveller’s behalf is not merely a faster customer. It is a structurally different kind of edge. It has none of the patience a human booker has. It cannot resolve ambiguity by asking the customer what they meant. It cannot stall the transaction while a human reconciles a policy mismatch out of band. It reads structured data, makes a decision, and books. Whatever the connectivity stack has been quietly hiding behind human accommodation is precisely what stops the agent from completing the transaction reliably.^⁹ The argument does not depend on agents becoming the majority booking path soon. It depends on them becoming commercially material enough that supply which is incoherent, incomplete or unreachable to machines becomes structurally disadvantaged. This is the point at which interface commoditisation meets the infrastructure of this industry. The interface layer is being rebuilt, in parallel, by many actors, across discovery, search, and agentic surfaces. The infrastructure underneath is the layer that was already strained and has now been asked to carry more load at lower tolerance.

Implication

Every piece of infrastructure that sits between the interface layer and the operator’s daily manifest becomes a load-bearing wall in a way it was not before. Those walls were not built with machine-to-machine load in mind. The translation layer in particular was built with human accommodation as a quiet dependency. Remove the accommodation and the compromises show.

The next essay takes up the agentic case in its own right, because the agent-mediated booking is the cleanest single place to watch the structural argument above operate in the specific. The essay after that asks where value actually settles when the interface layer commoditises at scale. This essay has done only the framework work: named the pattern, and named the place in this industry where the pattern meets resistance.

Close

The headline version of the AI debate in experiences is about replacement. It asks which consumer-facing participant is displaced and which one takes its place. The structural version is about revelation. It asks what the industry’s infrastructure has been quietly holding together, and what happens when the thing it was holding together for, a human at the edge, is no longer what arrives. AI does not remove the infrastructure this industry depends on. It reveals which parts of it were never strong enough to hold the weight they were being asked to carry. Machine-legible, current, trusted supply becomes more important, not less. The rest of this phase is about what that recognition implies.

Endnotes

The collapse of marginal cost per token across frontier and open-weight large language models, and the corresponding commoditisation of the consumer-interface layer, is documented in Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2025, on inference-cost curves and the rate at which interface-equivalent capabilities have moved from scarce to abundant. ↩
Christensen, C. M., and Raynor, M. E., The Innovator’s Solution: Creating and Sustaining Successful Growth, Harvard Business School Press, 2003, on the law of conservation of attractive profits: when an adjacent layer commoditises, value migrates to the layer that remains scarce. The same argument is articulated in operator class in Spolsky, J., “Strategy Letter V: The Economics of Open Source,” Joel on Software, June 2002. Christensen and Spolsky together provide the academic and the practitioner statement of the structural move named in this paragraph. ↩
The search transition is treated historically in Battelle, J., The Search: How Google and Its Rivals Rewrote the Rules of Business and Transformed Our Culture, Portfolio, 2005, and quantitatively in Brynjolfsson, E., Hu, Y., and Smith, M. D., “Consumer Surplus in the Digital Economy: Estimating the Value of Increased Product Variety at Online Booksellers,” Management Science, 49(11), 2003, pp. 1580–1596. Together they describe how value migrated from the publishing layer into the matching layer when retrieval became cheap. ↩
The attention argument originates in Simon, H. A., “Designing Organizations for an Information-Rich World,” in M. Greenberger (ed.), Computers, Communications, and the Public Interest, Johns Hopkins Press, 1971, pp. 37–72. The dominant-feed-era account is Wu, T., The Attention Merchants: The Epic Scramble to Get Inside Our Heads, Knopf, 2016. ↩
Carr, N. G., “IT Doesn’t Matter,” Harvard Business Review, 81(5), May 2003, pp. 41–49, and Carr, N. G., The Big Switch: Rewiring the World, from Edison to Google, W. W. Norton, 2008. As a category of capability becomes commoditised utility, competitive position migrates to the layer above or beside it. ↩
The hotel-distribution comparison is treated systematically in Buhalis, D., eTourism: Information Technology for Strategic Tourism Management, Pearson, 2003 (cited in Essay 1), on the development of central-reservation-system infrastructure, interface conventions, and inventory standards across the hotel sector. The contrast with experiences, where no equivalent layer has formed, is the structural point this paragraph rests on. ↩
See Essay 4 on the boundary count and translation burden of a single booking, Essay 5 on the absence of a neutral settlement layer relative to the aviation clearinghouse precedent, and Essay 6 on the structural strain accumulated by the channel-manager category as it absorbs problems beyond its original design envelope. ↩
The argument that human users in fragmented information systems quietly perform the reconciliation work the systems do not perform themselves is the central thesis of Suchman, L. A., Plans and Situated Actions: The Problem of Human-Machine Communication, Cambridge University Press, 1987. The reconciliation argument is reinforced by Bowker, G. C., and Star, S. L., Sorting Things Out: Classification and Its Consequences, MIT Press, 1999, on the labour of bridging classification mismatches at infrastructure boundaries. ↩
Arival, Guide to AI in Experiences, 2024. Field observations on operator and distributor adoption patterns during the early phase of agentic and AI-mediated discovery surfaces. The framing of AI as a layer that exposes rather than removes infrastructure is consistent with this report’s findings on machine-to-machine reliability requirements. ↩