← Essays
Essay

The Capability Model: A Map from Intent to Production

Marcelo Leal · June 15, 2026
Share X LinkedIn WhatsApp

Every company now has AI initiatives. Few have a map from the business outcome they want to the platform that would actually deliver it. Between intent and production sits a stack of decisions, and most teams meet them one tool at a time, out of order. The Capability Model is my answer to that: one layered map of what a business needs to take AI from intent to production, and how each layer depends on the one beneath it.

A model, not a checklist

A maturity chart ranks you, level one to level five, as if progress were a single line. A capability model does something different. It shows what has to exist for AI to reach production, arranges those capabilities into layers, and makes the dependencies visible. You read it top down, from the outcome the business wants to the infrastructure that runs it, and at every step the question is the same: what must be true here for the layer above to stand.

The Capability Model: four layers and the constraints band
Four layers, read top down, from the business outcome to the platform that runs it.

It is not a generic maturity ladder, and it is not finished. Like everything I publish, it is a working model, a map I keep reshaping as the field moves and as new implementation patterns show up in real work.

Four layers, one reading

The model has four layers, and each one answers to the layer above and leans on the layer below. Business is the top: the outcomes AI is supposed to serve, customer, operations, product, risk and control, illustrative and never exhaustive. Application is where AI meets the person who uses it, shown as families rather than products: agents, assistants, automations. Data is where meaning lives, the layer that turns records into something a machine can reason over. Technology is the floor: the cloud, the data platform, the AI platform, and the integration that runs all of it.

One layer opened to show the capabilities inside it
Open any layer and the capabilities inside it appear.

The point of the layering is not tidiness. It is that a failure travels upward. If the data layer does not know what a customer is, no agent above it can act on one, however good the model. The map makes that dependency impossible to ignore, which is exactly what a slide of disconnected boxes lets you do.

What cuts across

Some things refuse to sit in a layer. Cost, latency, security, scale, reliability, compliance, these bound every layer at once, and treating them as a footnote is how production systems fail in their second year. So the model gives them their own place, a band that runs across all four layers, named as first-class concerns rather than assumed.

The six constraints that apply across every layer
Six constraints. They do not live in a layer, they cut through all of them.

This is the part a capability map gets right and a feature list gets wrong. A demo answers the question once. A constraint is the same question asked again under load, under audit, under a budget. Naming the six of them up front is the difference between a thing that works in a meeting and a thing that works in production.

Where it goes deepest

One layer carries more weight than the rest, and it is the one most roadmaps skip: data, and inside it, the ontology. Models are fluent in language and blind to meaning. Most systems store records; the ontology stores what those records mean and how they relate, so software and people reason over one shared model. Read from the bottom up, it is a single path: data sources, then binding and entity resolution so the same real-world thing becomes one object, then a semantic model in open standards, then a knowledge graph you can query, then logic and rules that validate and derive, then actions that let the model operate and not just describe, and finally serving, where apps and agents consume it.

The ontology layer, read from sources at the bottom to serving at the top
The data layer, read bottom up: sources become meaning, meaning becomes something you can act on.

This is where I spend most of my depth, because it is the layer everything above inherits. Get it wrong and the confusion is passed up, silently, to every agent and every dashboard. Get it right and the rest gets easier than it has any right to be. The public model stops at the shape of this layer; the detailed implementation is its own, longer story.

Why this, and why in public

The skeleton is borrowed honestly: the layering is adapted from TOGAF, then reshaped by operating-model design and by hands-on AI implementation across different industries. I did not invent the idea of layers. What I am adding is a specific reading of how they apply to AI, where the dependencies actually bite, and which constraints decide whether any of it survives contact with production.

It is a work in progress and a study in public, the same as the Atlas. It will be wrong in places and it will change. I publish it because the thinking is the contribution: the deepest detail can stay private and the map can still be useful, to me for keeping the whole shape in my head, and to anyone trying to get from an AI intent to something that actually runs. The map is the argument.

Open the Framework →