Coordination + Orchestration

The Single Front Door: Stop Making People Pick the Model

Ask most people what scaling AI looks like and they will describe a better model. Ask anyone who has actually done it and they will describe a door.

The lived reality inside organisations is not a shortage of capability. It is a surplus of fragments: a chatbot here, a copilot there, a specialist tool for summarisation, another for search, each behind its own login, each requiring the user to know it exists and decide when to use it. People copy text out of one platform and paste it into another. They lose track of which tool is approved. The cognitive tax of choosing is high enough that most people just default to whatever is easiest, sanctioned or not.

A single front door removes the choice. One place to go. The system decides what happens underneath.

Coordination and orchestration are not the same thing

This is worth getting precise, because the two get collapsed constantly and the distinction is load-bearing.

Coordination is the structural question: do your capabilities, models, and tools behave as one system, or as a pile of disconnected islands? It is the anti-fragmentation layer. A single front door is, first, a coordination achievement. It means the pieces cohere enough to be presented as one thing.

Orchestration is the run-time question: given a specific request, which capability and which model should handle it, and how does the answer get delivered back? Routing a cheap, simple query to a cheap, fast model, and a hard one to a frontier model, is orchestration. So is delivering the result in the right format, to the right person, at the right moment. Delivery is orchestration's final act, not a separate reward, and not value.

You can have orchestration without coordination, a clever router bolted onto a mess, and it will feel brittle. You can have coordination without real orchestration, one door that always calls the same expensive model, and it will be coherent but wasteful. Scaling needs both.

What it looks like when it works

Wells Fargo is the clearest example. They describe their approach as poly-model, poly-cloud: one model for the customer assistant, a different one running internally, others tapped as needed, with an orchestration layer that talks to the model and chooses per task. The user never picks. The system does. And critically, the orchestration layer is also where governance lives: it is the filter in front of and behind the model. Coordination, orchestration, and governance, in one place.

JPMorgan's LLM Suite is the coordination story at scale: roughly 250,000 employees, one portal, multiple frontier models behind it, refreshed every few weeks. The point of the portal is not that it is clever. The point is that it is one thing, a single front door instead of a landscape.

Merck's GPTeal is the same shape in pharma: multiple models behind one secure door, with the door doing the governance work. Morgan Stanley's leadership described the destination more crisply than any vendor has: AI as an efficiency-enhancing layer that sits between colleagues and the many applications they use. That sentence is the single-front-door thesis.

The economics are real, but they are a second-order benefit

There is a tidy story about model routing that says send cheap queries to cheap models and save a fortune. It is true. The price spread across models is roughly an order of magnitude, and intelligent routing can cut cost substantially while holding quality, with caching on top. The gateway tooling to do this exists and is mature: self-hostable options that keep everything on your own infrastructure, which matters enormously for European data residency, and managed options with observability and PII redaction built in.

But the cost saving is not the reason to build the door. It is what you collect once the door exists. Build the routing for the savings alone and you will under-invest in the things that actually drive adoption: governance, audit, a coherent interface, and the sense that there is one trustworthy place to go. Start with the door and the coherence. Add the routing economics when your volume justifies it.

Where this sits in the framework

The single front door spans two layers of the AI Value System. Coordination is what makes the door possible: capabilities behaving as one system. Orchestration is what the door does: routing each request and delivering the result. Both sit on top of governed Knowledge and wrapped in Governance, and none of it produces value until it clears the Human Adoption gate. The door is necessary. It is not sufficient. That is the next piece.

Frequently asked

Questions this section answers

One interface through which employees access all sanctioned AI capabilities, instead of a scattered set of separate tools. The system routes each request to the appropriate model behind the scenes, so the user never has to choose.

Coordination is the structural property of your capabilities behaving as one coherent system. Orchestration is the run-time act of routing a specific request to the right model and delivering the result. You need both to scale.

For a regulated organisation the first question is data residency, which favours a self-hostable gateway. Start with governance and observability features, covering audit trails, PII redaction, and cost attribution, and add cost-based routing once volume justifies it.

Less than it did. Frontier model performance has converged, which is precisely why the orchestration layer, not the model, has become the differentiator.

Sources

Wells Fargo and Fargo (VentureBeat; Google Cloud), JPMorgan LLM Suite (CNBC; The Digital Banker), Merck GPTeal (Merck newsroom; IntuitionLabs), Morgan Stanley (OpenAI; Morgan Stanley press releases), LLM gateway tooling (practitioner comparisons of LiteLLM, Portkey, Kong; treat vendor benchmarks with care).

Continue reading