The Junction
2026Merchant of record · Agentic commerce · Group payment · Forecasting in the open
Who owns the sale when an AI agent books? A thesis about where the doors into a purchase end — and a set of methods for holding a strategic argument to the same standard as code: every figure tiered by how it is known, models with their assumptions on the page, and predictions frozen before the results.
Every door ends at someone’s checkout
There are many ways into a cinema ticket now: the cinema’s own app, a discovery site, an AI agent in a chat, a card network’s agent token. Each one ends at a checkout, and whoever runs that checkout is the merchant of record. They keep the sale, the fee and the guest; everyone upstream becomes a referrer.
Agents arrived on both sides at once. In the US, one chain’s ChatGPT app hands each booking to the chain’s own checkout, while an aggregator sells tickets inside ChatGPT on its own. In Australia and New Zealand, the first purchases made through Mastercard Agent Pay were completed at a cinema chain’s own checkout. Both ways of selling are live, and the default is still undecided. That moment — while agents are still choosing their integrations — is the junction.
Many doors, one rail
The argument is for a neutral junction: every door hands off to the merchant’s own checkout, on a payment rail the merchant contracts directly, and nobody in the middle ever holds the money. Owning a front door turns out to be the wrong unit of strategy. A model of one discovery platform becoming merchant of record came out negative in every one of its 2,000 draws, and it left the small cinemas it sold for worse off. The better move is to arm the doors that already hand off: deep links to the seat, attribution on every hand-off, and the merchant’s own rail as the default.
- The doors. Exhibitor apps, discovery, AI agents, group chats, card networks.
- The junction. The group’s yes becomes a booking: a block is held, each member claims and pays their own seat. No settlement verb; the merchant of record keeps the sale; nobody in the middle holds money.
- Behind it. The rail: each claim a payment on the cinema’s own processor. The record: every attendee a known guest of the cinema that sold the seat. The graph: who brought whom, with consent.
Agents are built for one buyer; cinema is bought by groups
The emerging agent payment standards are single-signer: a mandate is signed by one user, and split payments split instruments, not buyers. But going to the cinema is a group decision, increasingly made in a group chat, and today one person fronts the money for everyone. The open design problem is a group mandate: hold an adjacent block of seats, let each member claim and pay for their own at the merchant’s checkout, and measure whether that brings people who would not have come — formation, not convenience.
Hold an argument to the standard of code
Every figure says how it is known. Each number carries a tier: observed in a primary document, reported by the press and unconfirmed, modelled by me with stated assumptions and sensitivity, or argued — my judgment. A claim can’t borrow the authority of a tier it hasn’t earned.
Three models, each answering one question. A decision model in Excel, 2,000 Monte Carlo draws discounted at the owner’s own published rate, asks whether a door is worth building. A floor-up model built from a census of 129 New Zealand cinemas, with an evidence ledger behind every input, sizes what a payment rail actually earns — and reports that its take rate is weakly identified rather than pretending otherwise. A revenue model in Python, 20,000 runs, sizes the upside.
Predictions frozen before the results. The forecasts were written down with P10, P50 and P90 ranges before the company’s results were published, with one pre-set falsifier: a number outside a stated band means the thesis was wrong about the thing it depends on most. Alongside them, the conditions under which the whole argument fails — re-coupling stalls, the rail turns out not to matter, agents pick aggregators, or a group pilot brings fewer than about 15% new customers.
Calibration, said out loud. The customer simulation behind the forecasts scored a mean Brier score of 0.27 — no better than saying 50% on every call, which scores 0.25. That goes on the page, next to the predictions it qualifies.
Three ways the junction could be held
Toll road: aggregator checkouts live inside agents, and the cinema fulfils orders behind someone else’s fee. Patchwork: each chain signs its own agent deals, one at a time, with no shared endpoint. A merchant-sovereign rail: one agent-ready checkout per merchant, on its own rail, paid on throughput. Each future comes with the signals that would show it arriving, so the argument can be checked against the world rather than defended against it.
Written in October 2026; checked when the results land
The forecasts are scored against published results in February 2027, whichever way they fall. The work behind it — the models, the evidence ledgers and the full argument — was prepared privately, and is available on request.