Pizza, Permission, Proof

Pizza

In 2007 a pizza chain had two ways to let people order online.

It could list itself on one of the ordering sites. SeamlessWeb had been taking consumer orders since 2005, Grubhub since 2004. Listing was an afternoon of work and a line item in the marketing budget. Or it could build its own ordering system — years of engineering to do a thing the telephone already did perfectly well.

Nobody in 2007 was choosing between dependence and independence. They were choosing between a cheap experiment and an expensive one.

Nobody except Domino's.

They built online ordering from the ground up across their consumer experience. Ordering in 2007, the Pizza Tracker in 2008 — and a tracker only works if the order, the oven, and the driver are one system reporting on itself. More than 85% of US retail sales run through those channels today.

The other path worked too. Some Pizza Hut operators stopped employing delivery drivers altogether. It was an operator solving a hiring problem by Friday. Today, Papa John's takes roughly 17% of its sales through aggregators against 35% direct.

In July 2023 Domino's listed on Uber Eats anyway, and on DoorDash in 2025. But even then, the pizza still arrives carried by a Domino's driver, and the loyalty program still only works if you order direct.

Both strategies work. But notice what you know without being told. Whether or not you have ever ordered from Domino's, you have heard of the Pizza Tracker. You can picture it. Nobody can picture how a Papa John's order reaches the kitchen, because there is nothing to picture — it gets there through somebody else's software.

That is what the infrastructure actually bought, and it wasn't efficiency. When an aggregator raises its rates, Domino's can decline. The operator who gave up his drivers can pass the increase to the customer or absorb it, and those are the only two things they can do.

Permission

A health system buying an AI tool this year is standing in front of the same two doors, and they look as different now as they did then. One is a purchase order — for every tool you could want, each with its own integration project and its own integration cost, placed as a bet on which one turns out to work. The other is an infrastructure program that builds AI governance into the foundation of the organization and has nothing to show for it this fiscal year.

In February 2026 Epic released AI Charting — a native feature that listens to the visit, drafts the note, and queues the orders, drawing on the patient's entire record to do it. It comes with the system. Every AI scribe charging a few hundred dollars per provider each month, plus the integration project to get it running, is now competing with free.

They are not losing on model quality. They are losing because free, today, with almost no integration cost, is not a hard decision for a hospital. Why take the risk? Why hire more drivers?

But Epic's version is still a tool bolted onto a record. A better one, with a larger permission set — it can queue the order because it owns the record. What it cannot do is hold the thing that matters later, because the record underneath it was built to store documents, not decisions.

It is a good answer to this year's question.

The questions that matter come two years from now, or five. What did the AI tell us to do? Why did we follow it? What did it see that our team missed?

And the one nobody is asking yet: what's our Pizza Tracker?

Once the small wins are banked — the faster notes, the cleaner coding, the shorter prior auth queue — what comes next? The honest answer is that eventually AI has to go deeper into the systems it runs on. That means a serious upgrade and another round of sunk cost to reimplement, on top of every integration already paid for. The bolt-on path does not avoid the foundation work. It postpones it and charges interest.

That is how the hangover keeps going.

Proof

Every one of those questions has the same answer, and it is not a better tool.

Build the record of the decision into the substrate every tool runs on. What was asked. What data the model was permitted to see. Which model answered. What it returned. The licensed human who decided what to do with it. One unbroken, tamper-evident record, written by the foundation rather than by the vendor.

That is the Compute Chain, and it is a compliance record every organization using AI in care should already have.

The Pizza Tracker is not a feature anyone could buy. It exists because the order, the oven, and the driver were already one system. The Compute Chain works the same way. When every AI decision lives in one record, "what did the AI see that our team missed" stops being an investigation and becomes a question you can ask.

And because the chain is built into the substrate rather than bolted on top of a product, it does not care which AI you buy. Run the vendor's scribe. Run three of them, and replace all three in two years. The chain holds, because it was never the vendor's to keep.

If DoorDash disappeared tomorrow, a Domino's driver would still be at your door.

For an organization standing in front of those two doors this year, the conversation starts at info@golden-i.io.