From Models to Minds
A Cognitive Mode Framework for Designing AI
Photo by srihari kapu on Unsplash
When a user begins using a leading consumer AI today, they’re immediately faced with the “pick your model” problem—clearly illustrated in the ChatGPT UI below.
Having a menu of models is useful for developers building with the product, but for users a growing list of models introduces a 'paradox of choice' i.e. more options makes it harder to choose. Mainstream users - for whom gen AI is a tool to tackle problems with, not in of itself a curiosity to tinker with - won't care to remember which model is good at what; they'd just want powerful, useful output. Their thrill of a new model, if any, will wear off soon. Hence they will (if not already, like me) prefer to simply submit their prompt and let the software figure out - via conversation with the user - which model to deploy, and the kind of output that's right. So the user's perception of value will shift from 'range of models' and 'type of model' to simply 'time to value'.
So how do we build for this shift in user expectation?
As AI capabilities mature and models proliferate, the real challenge product teams face today is not which model to choose - but how to design the cognition their AI product should exhibit.
To support this shift, I propose a two-dimensional cognitive mode framework that helps shift thinking from model selection to the AI’s cognitive design. The framework doesn’t describe the user’s mental model—it describes the kind of intelligence the AI system must execute.
It is built on two axes:
Axis One: Determinism
The dimension maps predictive versus generative cognition, the spectrum that most traditional generative AI have strived to perform accurately in.Axis 2: Reasoning
Inspired by Daniel Kahneman’s work on Thinking, Fast and Slow, this axis captures the spectrum from intuitive (fast) to analytical (slow) cognition. Recent thinking on this dimension include the Sequoia paper, Gen AI’s Act Two (1). This is the deterministic dimension, for example ‘chain of thought’ processes in current AI, which are exploding in popularity and lead the current frontier of the field.
Together, these dimensions define four key cognitive modes—distinct ways your AI product might need to operate.
Mode 1. Intuitive + Predictive:
Instant Judgment
Think fast, narrow, and confident. These systems act like trained instincts—using historical data to make quick decisions.
Example prompts/tasks:
"Approve all reimbursement claims under €5000 unless flagged for anomaly."
"Auto-tag inbound support tickets into predefined categories."
"Predict the next best action in a customer journey."
Design principle: Lightweight and responsive. Trust the defaults.
Mode 2. Intuitive + Generative:
Creative Improv
This mode powers fluid, idea-generation workflows. The focus isn’t on correctness—it’s on momentum and tone.
Example prompts/tasks:
"Draft a friendly rejection email to a job candidate."
"Generate a list of slogan variations for a health brand campaign."
"Suggest subject lines for a promo email about a winter sale."
Design principle: Fast creativity with personality.
Mode 3. Analytical + Predictive: Deliberate Reasoning
Here the system slows down. It weighs, evaluates, compares, and justifies.
Example prompts/tasks:
"Rank potential warehouse sites based on cost, resilience, and proximity to demand centers."
"Forecast Q4 revenue factoring in promotional campaigns and seasonality."
"Evaluate insurance claims for fraud risk based on pattern deviations."
Design principle: Traceable logic and depth over speed.
Mode 4. Analytical + Generative: Structured Imagination
Here, the system creates—but with structure, constraints, and intent.
Example prompts/tasks:
"Generate a detailed onboarding plan for a new engineering hire, based on their role and experience."
"Draft a compliance policy based on ISO standards and internal controls."
"Build a personalized lesson plan for a learner struggling with algebra."
Design principle: Rich outputs that can be inspected, edited, and reused.
What About the In-Betweens?
In practice, real-world workflows often straddle quadrants. For instance:
"Generate personalized marketing copy based on past customer behavior."
— This blends predictive + generative modes, intuitive if automated, analytical if heavily tailored."Summarize legal risk in a contract and suggest clause improvements."
— This demands both predictive insight and generative rigor, and leans analytical."Draft a business strategy based on current performance data and competitor benchmarks."
— Squarely generative + analytical, but might involve quick instinctive heuristics at points.
The fact that we may run into edge cases within each quardant may be a good thing. A sign of increasing sophistication of the AI to would be its ability to straddle hybrids modes and synthesise between them in its outputs. Keep in mind:
A quadrant is a representation, not a fixed reality—it serves as a shared mental model to help product teams align on what they’re aiming for.
The prompt determines the cognitive mode, helping the AI shift into a specific ‘mindset’ to respond best.
The AI responds according to the chosen mindset and learns through reinforcement from the user.
Towards a generalised Cognitive Architecture that shifts effortlessly between the four modes
In the early stages of AI product development, different quadrants might require different models, data pipelines, or even teams. But as systems mature, we’ll move toward prompt-based orchestration—a single engine (or architecture) dynamically switching modes based on what’s being asked.
At that point, building AI will feel less like shipping models and more like designing minds—where prompts guide not just inputs and outputs, but the type of thinking the system performs.
And that’s what product strategy in AI will ultimately become: not just choosing the right model—but choosing the right cognition to deploy for the moment.
References
(1) Gen AI’s Act Two, Sequoia Capital
