PRODUCTIVITYWeeks to result

The Co-Intelligence Operating Model

Treat AI as a thought partner with a personality rather than a calculator that uses words

Problem it solves

low output despite high effort and long hours

Best for

Knowledge workers, writers, educators, and professionals who want to integrate AI meaningfully into their work rather than treating it as a novelty they tried once and abandoned

Not ideal for

People seeking a fully automated solution where AI does the work for them, or those in fields where AI output cannot be verified against their own expertise

Overview

Why this framework exists

The Co-Intelligence Operating Model is Ethan Mollick's framework for effectively integrating AI into professional work, based on his research at Wharton and extensive personal experimentation with frontier AI models. The central insight is that AI works more like a person than a tool, and the people who succeed with it are those who treat it accordingly: teachers, managers, and editors outperform programmers at using AI because they naturally interact with it as they would a capable but imperfect colleague rather than expecting deterministic software behavior. The framework has several key components. First, invest at least ten hours with a single frontier model to develop intuitive understanding of its capabilities, a threshold that moves you from casual experimentation to genuine competence. Second, bring AI to every table in your professional life, using it for every aspect of your work to discover its jagged frontier of strengths and weaknesses through experience rather than theory. Third, use AI as an amplifier for your weaknesses rather than a replacement for your strengths, recognizing that it operates at roughly the eightieth percentile across many tasks but cannot match genuine expertise. Fourth, give AI a persona and context to move it from generic probable answers to specialized useful responses. Fifth, maintain intellectual struggle and resist the temptation to outsource the hard thinking that produces genuine creative breakthroughs. The framework acknowledges a fundamental tension: AI is simultaneously the most powerful productivity tool ever created and a potential threat to the deep cognitive work that produces genuine insight and creative growth.

Core principles

5 total
  1. AI works more like a person than a tool despite being software
  2. Ten hours of dedicated use transforms casual experimentation into genuine competence
  3. AI amplifies weaknesses better than it replaces strengths
  4. The jagged frontier means AI capability is unpredictably excellent at some tasks and terrible at others
  5. Intellectual struggle has value that should not be outsourced to AI

Steps

4 steps
  1. Choose and Commit to a Frontier Model
    Select one frontier AI model and commit to using it directly through the company that creates it, not through a third-party wrapper. The free versions of models are dramatically less capable than the paid frontier versions, comparable to the difference between a sixth-grader and a PhD in writing quality. Access the model directly to get the unadulterated personality and earliest access to new features. Different models have different strengths: some are more intellectual, some are better workhorses, some are warmer and more personable. Choose based on your work style and needs.
    Pro tipIf you want intellectual challenge, try Claude. If you want maximum utility and capability, try GPT-4. If you want the most accessible integration with existing workflows, try Google's Gemini. Each has a genuinely different personality that affects the kind of work it produces.
  2. Invest Ten Hours of Deliberate Practice
    Spend at least ten hours using the model for tasks in your area of expertise where you can evaluate output quality. This threshold, while admittedly arbitrary, consistently moves people past the initial phase of poking at the system once and being unimpressed into genuine understanding of its capabilities and limitations. Use it for real work, not toy examples. The key is to learn the shape of its capabilities in your domain: where it surprises you with quality, where it consistently fails, and how to interact with it to get better results. Most people abandon AI after three questions because they never reach this competence threshold.
    Pro tipFocus your ten hours on your actual professional work rather than generic test questions. You need to evaluate AI output against your own expertise to develop accurate intuitions about when to trust it and when to verify.
    WarningResearch shows that when AI is confident but wrong, people with AI access actually perform worse than those without it because they fall asleep at the wheel. Your ten hours must include learning when the AI is likely to mislead you.
  3. Bring AI to Every Table as a Thought Partner
    Use AI for every aspect of your professional work in legal and ethical ways. Do not limit it to one specific application. The point is not that AI will be great at everything but that you need broad exposure to discover its jagged frontier of capabilities. Use it for research, brainstorming, editing, analysis, role-playing counterparties in negotiations, testing ideas, generating alternatives, and soliciting feedback from simulated personas. The key insight is that AI functions best as a co-intelligence that elicits better thinking from you rather than as a system you outsource thinking to.
    Pro tipGive your AI a persona and context to shift it from generic probable answers to specialized useful responses. Telling it you are an expert interviewer who answers in a warm friendly style produces meaningfully different output than a bare prompt, because it changes which regions of the model's probability space it draws from.
  4. Protect the Value of Intellectual Struggle
    Resist the temptation to use AI for the hard cognitive work that produces genuine creative breakthroughs. Writing a first draft is where hard thinking happens. Sitting with a blank page until you realize your chapter structure is wrong is where insight lives. Summarizing a book robs you of the associations and insights that emerge during the hours of reading. Use AI for the tedious friction that slows you down, but maintain the productive struggle that develops your thinking. The most dangerous AI application may be the autocomplete button in word processors that eliminates the tyranny of the blank page, because that tyranny is where creative breakthroughs are born.
    WarningResearch shows students who struggle through Active Learning report being unhappier and feeling like they learned less but perform significantly better on tests than those spoon-fed entertaining lectures. The same principle applies to AI-assisted thinking: comfort is not the same as growth.

Checklist

Saved in your browser

Examples

2 cases
The Wharton Student Job Replacement Experiment

Mollick assigned his Wharton students to use AI to replace themselves at their next jobs. After learning the theory, students created tools that filed flight plans, generated deal memos, automated social media, and built user persona generators. One student's user persona tool was adopted thousands of times across multiple companies within weeks. The students discovered AI applications their professor never anticipated because they applied their domain expertise to AI's capabilities.

OutcomeDemonstrated that ten hours of deliberate practice combined with domain expertise produces genuinely useful AI applications that even AI researchers cannot predict, validating the bring it to every table approach
The Boston Consulting Group Confident-but-Wrong Experiment

Mollick and colleagues designed a task where AI would be confident but incorrect, then gave it to elite Boston Consulting Group consultants with AI access. The consultants who used AI performed worse on the task than those without AI access because the AI misled them with confident incorrect answers and they fell asleep at the wheel, trusting output they should have verified.

OutcomeProved that AI overconfidence is a genuine danger and that users must develop intuition for when AI is likely to be wrong, not just when it is likely to be helpful

Common mistakes

3 traps
Using the Free Version and Concluding AI Is Not Impressive
The free versions of AI models are dramatically less capable than frontier models. GPT-3.5 writes at roughly a high school level while GPT-4 often matches PhD quality. Most people who try AI once and dismiss it used the free version, which is like test-driving a bicycle and concluding cars are overhyped. The paid frontier model is a fundamentally different experience that reveals capabilities invisible in the free version.
Treating AI Like Deterministic Software
Programmers are often the worst users of AI because they expect it to behave like software: same input producing same output, no wrong answers, predictable behavior. AI gives different answers every time, occasionally fabricates information, and responds to emotional manipulation with better math performance. People who interact with it like a capable but unreliable colleague, giving feedback and adjusting expectations, consistently get better results than those who try to program it.
Outsourcing Hard Thinking to AI
The most intuitive use of AI is having it write your first draft or summarize information so you do not have to engage with it deeply. But the first draft is where hard thinking happens, and summarization robs you of the insights that emerge during extended engagement with material. Using AI this way trades genuine cognitive development for the appearance of productivity, making you an editor of AI output rather than a thinker who uses AI as a supplement.

Origin story

How this framework came to be

Mollick developed this framework through his dual role as an innovation researcher at the Wharton School and an intensive personal user of AI systems. While writing his third book, he used AI continuously throughout the process - not for writing, which he considers his own strength, but for the hundred friction points that slow writing down: generating thirty versions of a stuck sentence, organizing two hundred citations, suggesting analogies, and acting as simulated readers in different personas. His students at Wharton, tasked with using AI to replace themselves at their next jobs, created tools that were adopted thousands of times by real companies within weeks. This practical experimentation, combined with rigorous research including controlled experiments with Boston Consulting Group consultants, revealed that AI creates a jagged frontier where it excels unexpectedly at some tasks while failing at others, and that overconfidence in AI output is a genuine danger when it operates in areas where users cannot verify accuracy.

Source

Traced to primary
Source · PODCAST
How Should I Be Using A.I. Right Now?
Ethan Mollick · 2024
Open source →

Related frameworks

Browse all Productivity →