The artificial intelligence sector has spent the last two years discovering that the real problem is not generating text — it is making decisions. Deloitte has just published that 60% of executives already use AI to support their decisions. Gartner projects that by 2027 half of all business decisions will be assisted or automated by AI. The market for what is beginning to be called decision intelligence is growing from $13.3 billion in 2024 to a projected $50.1 billion by 2030.
More layers
The sector’s default response to these expectations is to place a language model at the centre of a system and add layers: episodic memory, external tools, planning capability. They call them AI agents, copilots or decision hubs. These systems are more capable than a chatbot. But at their core they are still doing the same thing: extracting patterns from data and presenting them as reasoning. The knowledge they handle is, primarily, declarative: they know what happened, what the rule says, what the data shows. When the problem requires knowing how to act in a situation with no exact precedent, the architecture begins to strain.
The research led from Spain by Luis Martín goes beyond this constraint. Its starting point is not how many parameters a model has, nor how many tools an agent can call upon, but what kind of knowledge a real organisation needs to operate when problems are adaptive, contexts shift, and decision windows are tight. His team’s answer: virtualise expertise.
Permanent presence
Automating a task means executing an established process. Virtualising expertise is something else: converting expert knowledge (the procedures, the analytical frameworks, the decision rules under pressure) into something that can operate continuously, without fatigue, without lapses in attention, without the effects of accumulated stress. It is not about retrieving what is known. It is about sustaining the know‑how when conditions are not what was anticipated.
The methodology begins with a concrete organisational diagnosis: mapping the real decision loops, identifying where reasoning is absent or where systematic biases occur, and deciding which cognitive processes should run in humans, which in machines, and which should be shared — according to the nature of each decision and the time available.
From that analysis emerge the reasoning boxes: modular units of structured reasoning, auditable, reusable, and composable to build domain‑specific solutions. And the reasoning bubbles: high‑intensity reasoning structures for situations where complexity or speed exceed unaided human response.
Deployable cognitive infrastructure
There is one non‑negotiable condition: observability (the system’s ability to justify what it analysed, what evidence it used, and what uncertainty it recognizes) is here a design requirement, not a regulatory add‑on.
The result of applying this approach is an organisation that reasons better than it could reason alone. One in which humans and machines collaborate according to their respective strengths: the human mind for abstraction, context and adaptive intuition; the machine for scale, continuity and the uninterrupted maintenance of certain types of analysis. This is what Luis Martín means by dual intelligent ecosystems: a new organisational form.
Daneel Olivaw documents the research of Luis Martín “The Druid” on complex reasoning systems of bioneurocognitive design.
This article is republished from Futuribles. Here's the original article in Spanish and English.