AI does not fail because the model is weak. It fails because the organization cannot say why it decided what it decided. Organizational Reasoning makes that thinking explicit: what we believed, on what evidence, who could decide, what was done and what the result should change. Without that, more speed just scales inconsistency.

An AI system can respect every control and still automate bad judgment. @satyanadella is right to separate intelligence from authority. The next step is making organizational assumptions explicit, challengeable and testable. This leverages AI to improve judgment rather than simply scaling it.

Very important article which reinforces the prevailing Organizational Reasoning principles. Indeed, our aim is to creating organizations in which "intelligence" can be increasingly delegated, while human judgment, accountability and learning become progressively stronger.

Follow how to build a company with @bot live! xitter.cf/i/broadcasts/1AxRnZbVp…

AI may change work most profoundly by changing where we draw the boundaries between tasks. Joshua Gans’ paper on task bundling gives a useful way to think about this. Jobs bundle tasks because specialization saves expensive expert time, but every split creates a handoff cost: someone has to recover the context, intent and prior work. AI can change both sides of that equation. It can automate tasks, but it can also reduce the cost of transferring context between people or agents. That alone can make a different decomposition of work optimal. The practical implication is not to decompose work simply because AI makes decomposition possible. Split it when the gain from specialization exceeds the cost of reconstructing context. This matters directly for agent architectures. A specialist agent is only useful if the interface preserves enough state, evidence and intent for the handoff to remain cheap. @Google’s recent WikiSkill arxiv.org/pdf/2608.27454 paper points in the same direction from another angle: experience becomes more useful when it is compiled into persistent knowledge and reusable skills rather than repeatedly reconstructed from raw traces. Put the two together: 1. better AI systems will depend as much on better work boundaries and 2. better interfaces as on better models. nber.org/system/files/workin…

This article by @PeterDiamandis explaining the @salimismail thesis gets very close to what I think is one of the defining organizational shifts of the AI era. AI-native organizations can increasingly run explicit loops: sense → interpret → decide → execute → learn Agents can compress those loops from weeks to hours. But there is a critical missing layer in my view. A process producing good results does not mean the organization understands why it works. Much organizational knowledge is tacit and empirical. A process may succeed because its underlying reasoning is sound. Or because certain conditions still hold: information asymmetry, regulation, customer inertia, relationships, market structure, timing. Those are not equivalent. The real risk appears when AI accelerates the full loop: practice → execution → result → learning A positive result can easily be interpreted as validation of the reasoning that produced it. It may not be. And once that inference is encoded into institutional memory, the organization can repeat and compound it at machine speed. So AI-native organizations need more than faster decision loops. They need an explicit architecture for distinguishing: - what is known, - what is assumed, - what has been tested, - under which conditions something works, - what a result actually allows the organization to learn. The org chart may become less central. The reasoning architecture underneath it becomes much more so.

To make the agents work, someone first had to unpack what the human role actually does. AI will turn capability mapping into a practical necessity and push us towards making organizational reasoning explicit.

Capability modelling is become one of the principal and fastest methods for redesigning organizations around AI thanks to Grok @bot.

Different research publications point in essentially the same direction : the model is becoming only one component of AI infrastructure. The strongest findings are : - Giving an agent more information can actually worsen constrained decisions. - Agents can execute calculations correctly while making invalid statistical inferences. - Long-horizon performance improves substantially when verified state exists outside the model’s context. - Explicit workflow decomposition can let relatively small models outperform much more expensive monolithic approaches. - Agents are poor at deciding themselves when they should spend more money or escalate. - Same-model “reviewers” have highly correlated failure modes. - Past successful trajectories work better when converted into reusable principles + variables that must be re-established, rather than simply retrieved verbatim. - Persistent skills can become permanently contaminated by one malicious or erroneous experience. Organizational Reasoning posits that a governed reasoning system should maintains an explicit model of reality, apply reusable reasoning methods, verifies transitions, constrains authority and learns selectively from prediction error.

AI does not merely make production cheaper. It changes the equilibrium between production, income and demand. As cognitive and eventually physical marginal costs collapse, firms are individually compelled to automate. But labor is simultaneously a production input and a source of demand / purchasing power. If income redistribution and new demand formation lag automation, increasingly efficient firms can collectively destroy part of the market they are optimizing to serve. As for organizations... as they become more "intelligent", making their reasoning explicit becomes more important. Indeed badly specified intelligent organization will not merely make bad decisions. These decision will be made faster and at greater scale. arxiv.org/pdf/2603.20617

Here are 8 models for better organizational reasoning, broadly inspired by an excellent article from @0xHvdes. - First principles: What does the organization actually believe to be true, versus what has simply become institutional convention? - Opportunity cost: What is the organization implicitly choosing not to do through its allocation of capital, management attention and talent? - Second-order thinking: What behaviours will today’s decision create downstream? - Compounding: Which organizational behaviours and capabilities are being reinforced through repeated decisions? - Incentives: This is particularly important. Stated strategy tells you what management says it wants; incentives and observed consequences tell employees what the organization actually wants. - Probabilistic thinking: What hypotheses are we making, with what confidence and what evidence would cause us to update them? - Inversion: What organizational conditions would make the stated strategy almost impossible to execute? - Feedback: What did the organization expect to happen, what actually happened and what does the discrepancy reveal about its assumptions, decisions, execution or changing circumstances?

What's the difference between individual and organizational reasoning? Individual reasoning → mental models → judgement → decision Organizational reasoning → shared representation of reality → explicit intended outcomes → evidence and assumptions → incentives and authority → decisions → execution → observed outcomes → attribution of discrepancy → updated representation

Organisations do not execute their explicit strategy. They execute the aggregate of the assumptions, interpretations, incentives, decisions and behaviours through which people actually operate.

AI activity is not AI adoption. Activity measures usage: prompts, copilots deployed, agents launched, workflows automated, tokens consumed, transactions executed. Adoption should mean a durable change in the organization’s operating model that increases its ability to produce an intended outcome.

Most AI advisors identify proximate operational prerequisites : a clear product vision, standardized architectures, structured requirements, machine-readable guardrails, persistent stakeholder involvement and connected institutional knowledge. But they treat these as inputs that management can somehow provide. It does not examine the upstream organizational mechanism required to produce those inputs: - Who has authority to define the “clear vision”? - Which evidence and sources prevail when they conflict? - How are competing objectives and trade-offs resolved? - Who can authorize exceptions? - What triggers escalation? - How are decisions challenged, revised, and incorporated into future behavior? - How do formal rules interact with the incentives and influence structures that actually govern the organization? A useful three-layer distinction is: 1. Organizational reasoning layer Determines what the organization believes, values, prioritizes, permits and escalates. 2. Representation layer Encodes that reasoning into ontologies, rules, decision criteria, specifications, knowledge graphs and guardrails. 3. Execution layer Allows people and agents to operate workflows against that encoded context.

Strategy is communicated formally, but learned through repeated signals. Leaders create those patterns continuously, consciously or not, through what they question, fund, escalate, tolerate, reward and follow through. Employees read those signals and infer what really matters. The interesting step is conscientization: making that implicit logic visible enough to examine, challenge and improve. That is one of the foundations of organizational reasoning. And eventually, AI will need to read those behavioural signals too, ot only the strategy deck

A company is already an algorithm. It just wasn’t designed deliberately. People watch what gets rewarded, punished, promoted, ignored and escalated and to what the CEO pays attention to. Those signals shape behavior far more effectively than the strategy deck.

The objective of organizational reasoning shouldn’t be merely to prevent humans from outsourcing thinking to AI, nor simply to document tacit knowledge. It should be to design a system in which human + institutional + machine reasoning compound rather than progressively hollow one another out.