The How · Thinking Architecture · Framework
AI Cognitive Strategy
Matrix© 2026
Before any decision, your organization makes a prior one: how cognitive work gets allocated. Every organization needs a policy for that, and almost none has one.
The allocation policy: how cognitive work is routed across human and AI systems.
The Organizational Judgment Diagnostic · Three Altitudes
Your organization is already routing its cognitive work between people and AI. If no one designed that routing, people draw the line by habit, and the judgment they should own is the first thing they hand to the machine.
Operating Question “Before any decision about AI, there is a prior decision: what kind of thinking does this require? And who, or what, should be responsible for producing it?”Route Your Thinking
Decompose a real decision.
Route each part.
A real decision is rarely one kind of thinking. It breaks into cognitive steps, and you route each one with my matrix. Deciding that routing is the judgment you settle before AI touches any of it. Pick the decision closest to yours, then tap any step to see where it routes.
This decision has one step that carries the judgment only a human can own. You are routing the whole decision to AI, so that judgment leaves with it.
Tap a step to see where it routes
Take “make the actual call on who to cut” out of the AI pipeline and put a human name on it. Route everything else the way the matrix shows.
Your organization makes this kind of decision hundreds of times a week. When you set the routing as a standing policy, the same line gets drawn under pressure, after turnover, and on the days you are not in the room.
You answer one question at a time and build your verdict as you go. You can run all four frameworks in one live session, and your readings stay in your browser.
Open the Workbench ↗The Apparatus
The evidence, the cases,
the questions.
The full argument behind the instrument. Open what you need.
01 The Evidence How the routing is built, step by step · Figs. 02–03 · the three modes
Step One · Decompose
One decision.
Many kinds of thinking.
You route the parts of a decision, not the decision itself. Every decision is a sequence of distinct cognitive steps, and each step demands a different kind of thinking. Break the work down first, then run each step through the matrix on its own. Decomposition is the prior move the allocation policy depends on.
One decision, five different routings. The blue step (the judgment) is the one you must keep, and the one most often handed to AI by default.
Step Two · Classify
Four cognitive
routing quadrants.
Take one step at a time. Each step maps to one of four quadrants based on two variables: the stakes if it goes wrong, and whether its thinking should be preserved in human hands or enhanced with AI. The quadrant a step lands in determines how it gets routed.
The four routing quadrants
Step Three · Route
Three cognitive
routing modes.
Each quadrant resolves to one of three routing modes. The mode determines what AI may do for that step, and what must stay in human hands.
Keep thinking in human hands. AI functions as a mirror, not an author. The quality of the output depends on the quality of the reasoning, not the quality of the prompt.
Augment thinking with AI scaffolding. Human direction remains central; AI structures, surfaces, and expands, but does not decide. The human remains accountable for the reasoning.
Delegate to AI. Reserve human attention for higher-stakes cognitive work. The discipline is not in how well you offload. It is in how precisely you identify what qualifies.
02 The Karp Reframe Why this matrix changes the hiring question
The Karp Reframe
Why this matrix
changes the hiring question.
Alex Karp, Palantir's CEO, found people who defaulted to explicit reasoning and called it a hiring advantage. I read what he found as a system failure: organizations that rely on implicit, socially mediated cognition are fragile under complexity. Design a system that routes thinking correctly, and you stop filtering for the person. You start producing the outcome from anyone capable of that work.
03 Example Readings Four decisions, each step already routed · open a reading or pick one in Route Your Thinking
Restructure · Exposed
The decision “deciding which team to cut in the restructuring” · handled today: AI mostly generates it
This decision has one step that carries the judgment only a human can own. You are routing the whole decision to AI, so that judgment leaves with it.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may reflect your reasoning back for you to pressure-test. AI may not synthesize, smooth, or author. The call stays yours.
AI may generate and complete the draft. AI may not be the final author. At these stakes you own every claim.
AI may generate and complete it. AI may not ship unreviewed. A quick look is enough.
Take “make the actual call on who to cut” out of the AI pipeline and put a human name on it. Route everything else the way the matrix shows.
You are handing the whole decision to AI, but “make the actual call on who to cut” is a Preserve step. That is exactly where judgment quietly leaves the building.
VP hire · At risk
The decision “hiring a VP of Sales” · handled today: A mix, no clear line
This decision has one step that carries the judgment only a human can own. With no deliberate line, it falls to whoever’s habit wins that day.
AI may generate and complete it. AI may not ship unreviewed. A quick look is enough.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may structure and expand it freely. AI may not be taken as correct without a check. A light check is enough.
AI may reflect your reasoning back for you to pressure-test. AI may not synthesize, smooth, or author. The call stays yours.
AI may generate and complete the draft. AI may not be the final author. At these stakes you own every claim.
AI may generate and complete it. AI may not ship unreviewed. A quick look is enough.
Take “make the final hire call” out of the AI pipeline and put a human name on it. Route everything else the way the matrix shows.
With a mix, people draw the line by habit. Draw it on purpose instead: keep the 1 Preserve step, delegate the 3 Offload steps.
Pricing change · Exposed
The decision “raising prices 8% next quarter” · handled today: AI mostly generates it
This decision has one step that carries the judgment only a human can own. You are routing the whole decision to AI, so that judgment leaves with it.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may reflect your reasoning back for you to pressure-test. AI may not synthesize, smooth, or author. The call stays yours.
AI may generate and complete the draft. AI may not be the final author. At these stakes you own every claim.
AI may generate and complete it. AI may not ship unreviewed. A quick look is enough.
AI may generate and complete it. AI may not ship unreviewed. A quick look is enough.
Take “make the call on the number” out of the AI pipeline and put a human name on it. Route everything else the way the matrix shows.
You are handing the whole decision to AI, but “make the call on the number” is a Preserve step. That is exactly where judgment quietly leaves the building.
Build vs buy · At risk
The decision “building the feature in-house or buying the vendor” · handled today: A mix, no clear line
This decision has one step that carries the judgment only a human can own. With no deliberate line, it falls to whoever’s habit wins that day.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may map options, surface contradictions, externalize the structure. AI may not make the call or own the conclusion. You verify every load-bearing claim.
AI may reflect your reasoning back for you to pressure-test. AI may not synthesize, smooth, or author. The call stays yours.
AI may generate and complete the draft. AI may not be the final author. At these stakes you own every claim.
AI may generate and complete it. AI may not ship unreviewed. A quick look is enough.
Take “make the build-vs-buy call” out of the AI pipeline and put a human name on it. Route everything else the way the matrix shows.
With a mix, people draw the line by habit. Draw it on purpose instead: keep the 1 Preserve step, delegate the 2 Offload steps.
04 FAQ Questions about the AI Cognitive Strategy Matrix · six answers
FAQ
Questions about the
AI Cognitive Strategy Matrix.
What is the AI Cognitive Strategy Matrix?
It is the cognitive allocation policy every organization needs and almost none has. It is my routing framework for how cognitive work is allocated between people and AI, based on two variables: the stakes of the decision, and whether thinking should be preserved in human hands or enhanced with AI.
What problem does it solve?
Before any decision about AI, there is a prior decision: what kind of thinking does this require? And who, or what, should be responsible for producing it? Organizations without a deliberate allocation policy default to blanket AI use. That erases the line between cognitive work that should be preserved in human hands and work that should be enhanced or offloaded. The quality of everything downstream depends on how you route the work.
How do you apply it?
Decompose the work first: a real decision is a sequence of distinct cognitive steps, and you route the parts, not the whole. Then map each step to one of four quadrants based on its stakes and thinking type. Each quadrant corresponds to one of three routing modes: Preserve (keep thinking in human hands, with AI as a mirror, not an author), Enhance (augment thinking with AI scaffolding while human direction stays central), and Offload (delegate to AI and reserve human attention for higher-stakes cognitive work).
Do you apply it to a whole task or its parts?
Its parts. A whole decision is rarely one kind of thinking. It is a bundle of distinct cognitive steps, each with different stakes. Decomposing the work into those steps is the prior move the matrix depends on: you route each step on its own, which is why a single decision can land in several quadrants at once.
How is it different from "use AI everywhere"?
"Use AI everywhere" treats AI as a tool decision. The AI Cognitive Strategy Matrix treats it as a cognitive architecture decision: what kind of thinking does this require, and who or what should be responsible for producing it.
Who is it for?
Leaders and organizations designing how cognitive work is routed: executives, founders, and operators making decisions about where AI belongs in strategy, hiring, planning, and execution.
05 The Essays Every essay that puts it to work · newest first
- Read Your Own Organization July 2026
- What cognitive architecture actually means. June 2026
- The Vatican Named the Moral Stakes. Here Is the Architecture. May 2026
- Cognitive Architecture Is the Deadline May 2026
- I Think I Found the Unit of Cognition April 2026
- The Translation Tax April 2026
- I Experience Cognition the Way AI Computes. That Is Not a Metaphor. April 2026
- The Article Named the Symptoms. Here Is the Architecture. April 2026
One piece of advice every leader already gives, and one skill that makes it work for every kind of mind, in your inbox weekly. Free. No sales pitches. Unsubscribe any time.
Subscribe on Substack →My three frameworks read as one: align the culture, distribute the judgment, preserve it at entry level. Run your organization down the three altitudes and leave knowing whether your judgment is intact and which altitude to fix first.
Read the full system →