# Man-Days Are Dead: We Need a New Productivity Metric for the Age of AI Collaboration > When 40 minutes of focused cognitive effort produces the equivalent of 15 man-days of output, most organizations are still measuring productivity by attendance. This article proposes the AI Collaboration Portfolio, a five-dimension framework paired with a three-layer verification architecture, giving enterprises the first auditable system for assessing AI talent. Includes a complete case study and a free assessment tool. - Author: Paul Kuo（郭曜郎） - Date: 2026-04-08 - Pillar: Intelligence & Order - Tags: AI Productivity, The Mythical Man-Month, AI Collaboration Portfolio, Anthropic Economic Index, Five-Dimension Framework, Super Individual, Human-AI Collaboration, Performance Measurement, Enterprise Talent Assessment - Canonical: https://paulkuo.tw/en/articles/beyond-man-days/ - Language: English - Translation: AI-translated from the zh-Hant (Traditional Chinese) original - Original (zh-Hant): https://paulkuo.tw/articles/beyond-man-days.md - Site index for LLMs: https://paulkuo.tw/llms.txt --- > **TL;DR**: Man-days measure attendance, not value creation. This article proposes the AI Collaboration Portfolio five-dimension framework (Command / Delivery / Leverage / Quality / Influence), paired with a three-layer verification architecture, giving enterprises the first auditable method for assessing AI talent. Full case study included. → [Assess your AI collaboration level for free](https://paulkuo.tw/tools/ai-collab-portfolio/)
▶ Listen to Summary
AI-synthesized voice · cloned from the author's own voice
## 40 Minutes to Do 15 Man-Days of Work: What Can Man-Days Still Measure? *If 40 minutes of deep thinking can produce output equivalent to 15 traditional man-days, has the way we measure work already become obsolete?* --- Three AI agents running in parallel. One analyzes market data and generates a market-entry strategy. A second takes that strategy, restructures the backend architecture, and deploys a working prototype. A third integrates both outputs into an executable operating plan. The operator spent roughly 40 minutes: defining the problem, assigning tasks, correcting course in real time, and reviewing final quality. When it was over, the table held a fully functional product prototype, a market-entry strategy analysis, and an automated operating plan. In the past, that would have required a five-person cross-functional team working three to four days. Fifteen to twenty man-days. This is not the norm. It is a snapshot of what happens when limits are pushed to an extreme. But the trend it signals is structural: as AI dramatically lowers the execution threshold for standardized tasks, output is no longer determined solely by how much time you put in. It depends on how the operator defines the problem, distributes tasks, and governs quality. The human is still there, but the role has shifted from operator to curator. So what, exactly, does "man-days" measure in this context? It measures time, not value. When the metric lags behind reality, what you end up measuring is not productivity. It is presence. The [Anthropic Economic Index](https://www.anthropic.com/research), published in January 2026, provides the empirical foundation. Their analysis of over one million Claude conversations found a 0.92 correlation between instruction quality and output quality. The implication is worth sitting with: in the AI era, the ability to define a problem well has become directly equivalent to final performance. Tasks requiring college-level comprehension are accelerated 12x by AI; high-school-level tasks, only 9x. AI amplifies gaps, and it does so exponentially. Two candidates both list "proficient with AI tools" on their résumés. That single phrase can correspond to two entirely different levels of productive capacity. One person can ship a full-stack application over a weekend using AI. The other may have run many rounds of prompting, generated summaries, and filled a NotebookLM, looking busy the whole time, and produced nothing truly usable. The harder problem: until now, we have had no way to identify, measure, or even name that gap. --- ## When Hours No Longer Equal Output, How Do Organizations Rebuild Their Performance Framework? In 1975, computer scientist [Fred Brooks](https://en.wikipedia.org/wiki/Fred_Brooks) published [*The Mythical Man-Month*](https://en.wikipedia.org/wiki/The_Mythical_Man-Month). Its most enduring insight was demolishing the myth of the "man-month": headcount and time have never been interchangeable variables. Every person added to a project brings not only more capacity but also more communication, coordination, and integration overhead, costs that can easily swallow the efficiency gains you imagined. A project that is already running late typically does not recover when you throw more people at it. It slows further. Brooks was talking about human teams. The AI era has validated his argument from the opposite direction: AI has dramatically lowered the coordination overhead for standardized tasks. No onboarding, no alignment meetings, continuous 24-hour availability, near-zero context-switching cost. When coordination cost approaches zero, parallel execution becomes genuinely efficient for the first time. And the metrics built on the assumption that "coordination is expensive" lose their meaning at exactly the same moment. Some have proposed replacing man-days with "attention duration," a concept that breaks down cognitive engagement in AI collaboration into four types: activation attention (translating vague requirements into precise instructions), supervisory attention (reviewing AI output and correcting in real time), integrative attention (coordinating the distributed outputs of multiple AI agents), and incubation attention (allowing background thinking to surface and generate insight). The taxonomy is useful. But attention duration is fundamentally an input metric. It tells us how much cognitive resource was invested; it does not tell us what those resources actually produced, how good the output was, or how much value was created relative to conventional approaches. The gap is visible if you map the landscape. On the input side, we have "attention duration." At the macro level, the [Anthropic Economic Index](https://www.anthropic.com/research) estimates AI's contribution to US labor productivity growth at roughly 1.0 percentage point per year. At the individual output level, the level where hiring decisions are made, where freelancers are evaluated, there is almost nothing. The measurement vacuum sits right at the center. --- ## The World Is Testing AI Literacy. Why Is No One Testing Individual Output? This vacuum exists not from indifference. Over the past two years, governments, international organizations, and academic researchers have produced an enormous body of AI competency frameworks. Every one of them answers the same question: can you use AI? None of them ask: what have you built with it? The [AI Fluency Framework](https://www.anthropic.com/research), developed by Anthropic with Rick Dakan and Joseph Feller, defines four core competencies: Delegation, Description, Discernment, and Diligence. This is probably the closest existing answer to "what does it look like to use AI well." But it describes the qualities of good AI collaboration; it does not quantify what that collaboration produces. Anthropic's own Economic Index approaches the question from a different angle, analyzing millions of real conversations to estimate AI's impact on labor productivity. The data is immensely valuable, but it operates at the scale of a national economy, not your next hiring decision. The [US Department of Labor](https://www.dol.gov/) has published an AI Literacy Framework. The [Alan Turing Institute](https://www.turing.ac.uk/) released the third version of its *AI Skills for Business Competency Framework*. [UNESCO](https://www.unesco.org/) has competency frameworks for both students and teachers. Academic researchers have proposed Collaborative AI Literacy and Collaborative AI Metacognition scales. Many frameworks, but they all answer the same category of question: do you understand AI, can you use it responsibly, do you know its limitations? They measure literacy, whether your knowledge of AI is sufficient. What organizations actually need to know is performance: what did you produce with AI, how good was it, and how much faster was it than conventional methods? ![Positioning map of global AI competency frameworks: the lower-left quadrant is crowded with literacy-oriented frameworks; the lower-right quadrant (individual performance) remains empty](/images/ai-framework-matrix.svg) Place the existing frameworks on a 2×2 matrix with "literacy vs. performance" on the horizontal axis and "individual vs. macro" on the vertical, and a clear picture emerges. The left side is crowded: UNESCO, DOL, Turing, and various academic scales. The upper right holds the [Anthropic Economic Index](https://www.anthropic.com/research). The lower right, individual-level, performance-based, verifiable, is empty. Everyone is testing AI literacy. No one is testing individual AI productivity. This article proposes a framework to fill that gap. It continues the line of thinking from my earlier piece on the [AI capability gap in 2026](/en/articles/ai-capability-gap-2026/): that article established that the gap exists; this one attempts to quantify it. --- ## The AI Collaboration Portfolio: Five Dimensions, and What Your Team Is Actually Missing The AI Collaboration Portfolio is a five-dimension model designed to measure what an individual actually produces through AI collaboration. Each dimension must meet three design requirements: a theoretical basis, quantifiable metrics, and a verifiable data source from a third party. Before getting into the metrics, consider a scenario. You are interviewing a candidate who says "I'm very good with AI." How do you tell whether that's true? Five questions are all you actually need. Can they direct AI effectively? (Command.) What have they concretely delivered? (Delivery.) How much do they multiply their own capacity in a given amount of time? (Leverage.) Does the output hold up under scrutiny? (Quality.) Has anyone else adopted their methods? (Influence.) Here is each dimension in detail. ### Command (25%): For People Who Can't Ask the Right Questions Does your team have people who have all the tools, all the accounts, but still produce manually? The problem is not the tool. It is the direction. Command measures whether you can make AI do the right thing. Anthropic's research shows a strong relationship between instruction quality and output quality. People who ask better questions get better results from AI, reliably. High Command is not merely about crafting prompts. It is about building reusable workflows, designing automation pipelines, and translating ambiguous requirements into tasks AI can execute precisely. Verification signals: number of reusable workflows and system prompts, automation pipelines built, breadth of AI tool integrations, complexity of multi-step task decomposition. ### Delivery (25%): For People Who Only Talk About AI In meetings, they are fluent on AI trends. At the end of the quarter, what actually shipped? Delivery measures not what you tried, but what you launched. Verification signals: commit frequency and volume, number of deployed services or tools, published content, complete projects taken from zero to launch, code quality metrics. ### Leverage (20%): For the Bottleneck That More Headcount Won't Fix This is the heart of the "40 minutes vs. 15 man-days" question. Leverage answers the thing organizations care about most directly: if this person joins, what is the capacity multiplier? How much does the same cognitive effort expand through AI collaboration? Verification signals: ratio of actual cognitive time invested to conventional time estimates, number of parallel projects, number of AI tools orchestrated, automation coverage rate. ### Quality (15%): For the Risk That Volume Creates In an era when AI can generate content at scale, quality gatekeeping matters more than ever. Anthropic's research found that accounting for task success rates reduces productivity-gain estimates by roughly one-third. A third of the apparent "acceleration" is actually quality loss. Quality measures whether your output survives contact with reality. Verification signals: user adoption or traffic figures, system stability, number of quality-control mechanisms (automated tests, review processes, audit workflows), external citations or shares. ### Influence (15%): For the Problem That Strong Individuals Can't Be Replicated A person can be exceptional, but if their methods exist only in their own head, their organizational value has a ceiling. Influence measures whether others have learned from or adopted your approaches. It is the dividing line between being a high-performing individual contributor and being someone who accelerates an entire organization, the highest-leverage form of value creation. Verification signals: open-source project engagement (stars, forks, contributors), workflow template adoption count, reach of instructional content, number of times methodologies are cited externally. ### How the Dimensions Relate To be straightforward: the five dimensions are not fully independent. High Quality is often a precondition for Influence. A powerful automation pipeline (Leverage) is itself a form of delivery. The five capabilities reinforce each other, forming a value loop. This is a design feature, not a flaw. Genuine value creation is multidimensional by nature. A financial statement's revenue, gross margin, and net profit are causally related, but we do not collapse them into a single number, because each reveals something different. The five dimensions are five observational lenses, not five isolated variables. The goal is richer perception, not a perfect mathematical decomposition. --- ## How Do You Prevent the Portfolio from Becoming a Scoring Game? The most common objection to any competency framework is whether it degrades into score optimization. The concern is entirely valid. Goodhart's Law states it plainly: when a measure becomes a target, it ceases to be a good measure. No quantitative system is immune. The real question is not whether the system can be gamed, but whether the cost of gaming it is high enough to make manipulation unprofitable. The AI Collaboration Portfolio addresses this with a three-layer evidence architecture. **Layer one: automated retrieval.** With user authorization, the system pulls data directly from third-party platforms: commit histories, repository structures, automation workflow files, website analytics, package download counts, community metrics. These records are kept by independent platforms the user cannot alter. The number of commits in your GitHub history is the number of commits. This layer forms the "hard evidence" floor. **Layer two: structured self-reporting.** Items that cannot be retrieved automatically, number of parallel projects, conventional man-day estimates, AI tool integration approaches, are collected through a structured form. Every self-reported field sits next to an "evidence link" field. Entries with a link are marked *Evidenced*; entries without are marked *Self-reported*. Anyone reading your Portfolio sees those labels and can weigh them accordingly. **Layer three: AI cross-check.** With user authorization, AI analyzes repository structure and code patterns, then independently suggests dimension scores. The AI-suggested score and the user's self-reported score are displayed side by side. If you report 90 and the AI suggests 50, the discrepancy is visually flagged. All three layers work together: automated data cannot be fabricated, self-reports carry transparent evidence labels, and AI provides an independent reference. The goal is not to eliminate gaming. It is to raise the cost of gaming above the point where it makes sense to try. The framework's deepest defense, though, is external reality. The evidence required by Quality and Influence comes from outside what any individual controls: user adoption rates, community engagement, client feedback, market outcomes. You might coordinate internal votes, but you cannot compel a market to buy, or fabricate genuine open-source adoption. This system is not designed to render automatic verdicts. It is closer to forensic accounting applied to talent assessment: rather than listening to the story a candidate tells in an interview, it audits the digital evidence trail: how they defined problems, distributed tasks, corrected course, and achieved delivery through AI collaboration. --- ## Three Sharp Objections, and Our Responses We ran this framework through multi-model adversarial debate across three rounds of stress-testing. Here are the three strongest attacks and how we respond. **Q1: AI can help someone fabricate an entire evidence chain.** AI can fabricate a map, but not the memory of having walked the terrain. In a framework-guided in-depth interview, the interviewer pursues not "what did you do" but "how did you make decisions": why did you abandon that approach? How did you weigh the tradeoffs? What did you do when the AI produced an error? Someone who actually did the work can answer three layers of follow-up. Someone reading from a script breaks down at the third. **Q2: Leverage uses man-days as a baseline, the same unit you've argued is obsolete. That's circular reasoning.** Leverage should not be read as a static multiplier against a fixed baseline. Its core value is the rate of change over time, your rate of improvement. An engineer who achieved 10x leverage with GPT-4 in 2024 and still achieves 10x with far more powerful models in 2026 has revealed something: their adaptation has stalled. AI tools evolve and baselines shift. What actually predicts future performance is how quickly you adapt to each new paradigm, not your absolute score at any single point. **Q3: The framework focuses on individuals, but the highest-value contributions in the AI era are often shared cognitive assets.** This is a genuine limitation of the current version. Someone who builds a prompt library that an entire department uses, or designs a workflow that doubles everyone's efficiency, creates network effects that an individual-level dimension cannot fully capture. Influence touches this partially, but not sufficiently. A team-level framework, one that measures how a person amplifies the entire system's capacity, not just their own, is the most important direction for the next iteration. > 💡 **Additional objections and responses:** On why Adaptability and Ethics are not separate dimensions, on causal attribution between dimensions, and on whether using the author as the case study constitutes survivorship bias, see the [Appendix](#appendix-further-objections-and-responses) at the end of this article. --- ## Case Study: How Does One Person Run Eight Projects at Once? To validate that the framework is actually operable, I ran my own data through it. Every figure below is publicly verifiable. Rather than listing scores dimension by dimension, here are three contrasts that best illustrate the framework's diagnostic power. **Contrast one: a four-language content system.** My personal site [paulkuo.tw](https://paulkuo.tw) has accumulated over 80 articles, each available in Traditional Chinese, English, Japanese, and Simplified Chinese. A single commit triggers the automated translation pipeline that generates all four versions. This is not only Delivery (how much content shipped), it reflects Command (whether you can design the process that makes AI do it for you). The conventional approach: four translators, one editor, one project manager. **Contrast two: simultaneous cross-domain operations.** I maintain eight or more active projects at the same time, spanning software development, content publishing, e-commerce operations, and enterprise consulting, entirely distinct domains. Time-tracking data from Timing App feeds into a live dashboard on the site, providing actual cognitive hours invested. This is the direct expression of Leverage: conventionally, it would require a team of five to eight people. **Contrast three: what the weakest dimension reveals.** My Influence score is the lowest of all five. Open-source projects exist but community engagement is limited. Social media presence is active but not yet at scale. Significant external talks and citations are nearly absent. This is where the framework's diagnostic value is sharpest: it does not just confirm strengths, it tells you, with uncomfortable precision, what to work on next. The more important observation is not the scores themselves but the nature of the evidence. Every claim above can be independently verified through a public URL, a GitHub repository, or a website analytics API. Nothing rests on self-assertion alone. That is the fundamental difference between this framework and any questionnaire-based assessment: the evidence exists outside the system. A framework's author using their own data as the case study clearly carries selection bias. But the value of this case is not statistical, it is architectural. It is a complete, auditable, replicable blueprint. Anyone can run their own data through it. Think of it as an open-source release, not a clinical trial. --- ## The Cost of Not Measuring AI Productivity: The Capability Gap Is Widening Invisibly Why does this matter? Because the cost of not measuring is higher than the cost of measuring imperfectly. The [Anthropic Economic Index](https://www.anthropic.com/research) documents a deskilling effect: AI preferentially takes over the high-skill components of work and leaves the low-skill portions behind. Without a way to identify who genuinely excels at AI collaboration, high-capability individuals get systematically undervalued, while people who interview well but produce little get systematically overvalued. The productivity gap is real and growing. College-level tasks are accelerated 12x; high-school-level tasks, 9x. This is not a small gap accumulating slowly over time. It is a structural divergence, and it is currently invisible to the labor market, because no measurement system exists to surface it. This framework also has limits. Three deserve honest acknowledgment. **Individual vs. systemic value.** The framework measures individual output. But some of the highest-value contributions in the AI era are shared cognitive assets: a prompt library an entire department relies on, an automation workflow that accelerates everyone. How to measure a person's amplifying effect on a system's capacity, rather than just their own output, is the most important extension for future versions. **Privacy boundaries of digital footprints.** The first layer of the three-layer architecture is automated retrieval. The line between "automated" and "surveillance" depends entirely on who controls the data. If employees authorize access themselves and the Portfolio belongs to them, it is an empowerment tool. If an employer mandates it and managers have unilateral access, it becomes digital monitoring. The technical architecture must lock data ownership to the individual from the start. Otherwise the Portfolio becomes a new kind of time clock. **The commercial reality of B2B pricing.** The framework can change how organizations assess talent internally, but it cannot change how markets currently buy. Most clients' procurement processes remain tightly bound to man-day invoices. You can demonstrate with a Portfolio that you completed 15 man-days of work in 40 minutes, but the client's system has no field for that kind of nonlinear output. Internal talent assessment can move first, but replacing man-day pricing in the market requires the entire commercial ecosystem to shift its conventions. That is not a problem any single tool can solve; it is an industry consensus that takes time. > 💡 **More on limitations:** On capturing adaptability, the difficulty of cross-organizational standardization, and fairness and accessibility, see the [Appendix](#appendix-further-discussion-of-limitations) at the end of this article. These are not defects to conceal. They are the research agenda for the next iteration. The choice, ultimately, comes down to two options: use this framework, or continue with nothing. The nothing option looks like the present: organizations screening by résumé keywords, making decisions based on interview impressions, with no way to distinguish "uses AI tools" from "can carry an entire team with AI alone." What you cannot measure is invisible at the moment of decision. AI collaboration capability is invisible right now. The AI Collaboration Portfolio is an attempt to make the invisible visible. It is an imperfect map, but your alternative is to keep running blind across a new continent. The current version targets individual users. An enterprise version, supporting team dashboards, cross-department benchmarking, and hiring-process integration, is the next development direction. If your organization is thinking about how to assess AI talent, start with the individual version to understand the framework's logic. → *Assess your AI collaboration level for free: [paulkuo.tw/tools/ai-collab-portfolio/](https://paulkuo.tw/tools/ai-collab-portfolio/)* --- ### Appendix: Further Discussion of Limitations **How is adaptability captured?** Adaptability is not a sixth dimension. It is the rate of improvement across the five dimensions over time. A person whose Command, Delivery, and Leverage scores continue to rise across successive iterations of AI tools is highly adaptive. The framework captures this indirectly through the time trajectory of scores. Future versions may promote self-iteration speed to a first-order signal. **The difficulty of cross-organizational standardization.** Command can be verified through commit history for a software engineer, but how do you measure it for a marketing director or a designer? Definitions of each dimension can vary significantly across industries, functions, and organizational sizes. The five dimensions provide an observational structure, but specific quantitative metrics and weighting will need localized calibration by industry and function. **Fairness and accessibility.** People with access to the latest AI tools hold a structural advantage within this framework. If the measurement system amplifies the digital divide rather than revealing genuine capability, it has failed. --- ### Appendix: Further Objections and Responses **Q4: Why not add Adaptability and Ethics as dimensions?** Adaptability is not a standalone dimension. It is the rate at which the five existing dimensions improve over time. A person whose Command, Delivery, and Leverage scores rise continuously as AI tools evolve is, by definition, highly adaptive. The same logic applies to the "reflection loop": the meta-skill of using AI to analyze and improve one's own work patterns. It shows up as an upward trajectory across the five dimensions, not as a static score in a sixth column. Ethics is a baseline constraint, not a performance metric. Ethical violations should be disqualifying, not a 15-point deduction. **Q5: The dimensions are causally related, so precise attribution is impossible.** Acknowledged. The framework aims for richer observation, not mathematically precise decomposition. Where attribution is ambiguous, the three-layer evidence architecture provides raw data for evaluators to interpret themselves. This is intentional: the framework does not replace human judgment; it gives human judgment better material to work with. **Q6: Using yourself as the case study is survivorship bias.** Methodologically, this is entirely valid. A framework's designer will naturally select evidence favorable to their own story. The response is not to deny the bias but to make it structurally unimportant: every claim in the case study can be independently verified through a public URL, a GitHub repository, or a third-party analytics API. The value of this case is not statistical. It is architectural. It provides a complete, auditable, replicable blueprint that anyone can run with their own data. The initial version inevitably reflects the creator's context, but its value depends on whether the community finds it worth forking, attacking, and improving. --- ### References 1. Brooks, F. P. (1975). *[The Mythical Man-Month: Essays on Software Engineering](https://en.wikipedia.org/wiki/The_Mythical_Man-Month)*. Addison-Wesley. 2. Dakan, R. & Feller, J. (2025). "Framework for AI Fluency." Ringling College of Art and Design / University College Cork. Version 1.5. 3. Anthropic. (2026, January). "[Anthropic Economic Index report: Economic primitives](https://www.anthropic.com/research)." anthropic.com/research. 4. Anthropic. (2026, March). "Anthropic Economic Index report: Learning curves." anthropic.com/research. 5. Anthropic. (2026, March). "Estimating AI productivity gains from Claude conversations." anthropic.com/research. 6. Anthropic. (2026, March). "Labor market impacts of AI: A new measure and early evidence." anthropic.com/research. 7. [US Department of Labor](https://www.dol.gov/). (2026, February). "AI Literacy Framework." 8. [Alan Turing Institute](https://www.turing.ac.uk/) / UK DSIT. (2025). "AI Skills for Business Competency Framework." Version 3. 9. [UNESCO](https://www.unesco.org/). (2026). "AI competency framework for students." unesdoc.unesco.org. 10. Schleiger, E. et al. (2025). "Generative AI in Human-AI Collaboration: Validation of the Collaborative AI Literacy and Collaborative AI Metacognition Scales." *Interacting with Computers*. Taylor & Francis. 11. Chee, K.N. et al. (2025). "A Competency Framework for AI Literacy." *British Journal of Educational Technology*. Wiley.