<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Khurram Sultan on Medium]]></title>
        <description><![CDATA[Stories by Khurram Sultan on Medium]]></description>
        <link>https://medium.com/@khurram.sultanahmed?source=rss-e6b31ac66f1f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*5RGbFtAHfIUe73Mh</url>
            <title>Stories by Khurram Sultan on Medium</title>
            <link>https://medium.com/@khurram.sultanahmed?source=rss-e6b31ac66f1f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 07:46:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@khurram.sultanahmed/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI Won’t Create the Leadership Crisis. It Will Expose It.]]></title>
            <link>https://medium.com/@khurram.sultanahmed/ai-wont-create-the-leadership-crisis-it-will-expose-it-269ccb551c7a?source=rss-e6b31ac66f1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/269ccb551c7a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Khurram Sultan]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 15:00:14 GMT</pubDate>
            <atom:updated>2026-11-09T15:15:54.414Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI Won’t Create the Leadership Crisis. It Will Expose It.</h3><p>When execution becomes automated and accountability becomes measurable, what value will leadership bring?</p><p><a href="https://images.openai.com/static-rsc-4/AOZtAnNta99VvpWK4CLXdz0fWOfHI_jxchXVNLDyr7l6pxOgDKwaRPABz4GM8-UaA9xX-Mg8OJeiBcRPGrcbU2yimIM17M9ZU2srDEQfL-56VfWxhWDoUkcdkBf68lBGfyjX14jr9Fs_BVpJW7nTGZIR1iCv483aSKPSYVcY6us?purpose=inline">Leadership decisions cascade into pressure on exhausted teams</a></p><p><a href="https://medium.com/%40kanishks772/the-real-reason-intuit-laid-off-3-000-employees-isnt-cost-cutting-it-s-architecture-6c8d4a1e1feb?utm_source=chatgpt.com">medium.com</a></p><p>There is a leadership crisis in modern organizations, and it did not begin with artificial intelligence.</p><p>It began when organizations confused authority with competence, ambitious commitments with sound planning, and employee sacrifice with operational excellence.</p><p>Executives make promises disconnected from reality. Senior leaders translate them into targets. Managers transfer pressure downward. Employees compensate through overtime and personal sacrifice.</p><p>When things succeed, leadership claims credit. When they fail, employees face questions about productivity, ownership, and commitment.</p><p>We have normalized a system where leadership creates the problem, employees absorb the cost, and leadership receives the reward.</p><h3>AI Is Changing the Rules</h3><p><a href="https://images.openai.com/static-rsc-4/S8FF0dsm1zpyaiElQLB7KZ3qHdbWmNL2lwqxssZh-FvcCVabHL-m_jwTiAdQXiFfXN-2JknGO4FvO1Pa2Ju45VlFkxaGyT7tOdlF1_deu6u1nKWW21WTXr9EQevxdXMJAB2dQ-coCkOtjKvEzgXhXCplgRuIBcn8XjV7z8evRfk?purpose=inline">Human teams coordinate workflows with AI agents</a></p><p><a href="https://medium.com/h7w/the-post-ai-workforce-hybrid-human-ai-teams-4fd9332cf8e6?utm_source=chatgpt.com">medium.com</a></p><p><a href="https://www.gartner.com/en/documents/7658761?utm_source=chatgpt.com">Gartner’s March 2026 research</a> describes pressure to reduce management layers while preserving the managers needed to deliver AI value.</p><p><a href="https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born?utm_source=chatgpt.com">Microsoft’s 2025 Work Trend Index</a> introduced the agent boss: someone who delegates to and oversees AI agents. In its survey, 28% of managers considered hiring AI workforce managers and 32% considered AI agent specialists.</p><p>The <a href="https://www.weforum.org/publications/the-future-of-jobs-report-2025?utm_source=chatgpt.com">World Economic Forum</a> projects 92 million jobs displaced and 170 million created by 2030 across several economic forces, not AI alone.</p><p>None of this proves leadership jobs will disappear. It does suggest that coordination-heavy hierarchies face a serious test.</p><p>Yesterday’s leadership habit</p><p>The capability that matters next</p><p>Promising without capacity analysis</p><p>Evidence-based commitments</p><p>Chasing status updates</p><p>Continuous, meaningful visibility</p><p>Escalating predictable crises</p><p>Anticipating and managing risk</p><p>Taking credit, distributing blame</p><p>Traceable decision accountability</p><h3>You Cannot Emotionally Blackmail an AI Agent</h3><p>An agent does not fear a poor appraisal, losing a promotion, or disappointing its manager. You cannot convince it that sacrificing another weekend demonstrates loyalty.</p><p>And you cannot make your inadequate planning its moral obligation.</p><p><a href="https://images.openai.com/static-rsc-4/k85lZBcW3nuIncpWpq4k7mGOU48SiiSzYMvC2WhIcK_BSNTbegl10zqaoZ6u92GhiEEeWiAkDgBZGDs5qm62kZ2gHswrdKFr6fdIYLPMqawqfaTHnrzMxJj-50OAWkJNh2e8q4e_2figWwA5jwC86FPR1n4eQYfcUNshGqM9bYs?purpose=inline">A team reviews a mismatch between deadlines and evidence</a></p><p><a href="https://pilotoasistido.com/?utm_source=chatgpt.com">pilotoasistido.com</a></p><p>Imagine an AI-enabled delivery system estimating six weeks for a project. Leadership commits to four. The assumptions, dependencies, and risks are recorded. Four weeks later, delivery is incomplete.</p><p>Who is accountable: the engineer who refused overtime, or the leader who ignored the evidence?</p><p>Agents can make mistakes, and data can mislead. But better traceability can expose decisions that were once concealed behind allegations of poor employee performance.</p><h3>The Future Needs Leadership, Not Hierarchy</h3><p>Imagine teams where agents support planning, development, testing, monitoring, and reporting. Human specialists evaluate their work, handle exceptions, and make consequential judgments.</p><p>What happens to leaders whose primary contribution is collecting updates, transmitting pressure, or presenting other people’s achievements?</p><p>The <a href="https://www.ilo.org/publications/generative-ai-and-jobs-2025-update?utm_source=chatgpt.com">ILO’s 2025 research</a> suggests transformation, rather than wholesale replacement, is more likely for many AI-exposed jobs.</p><p>That is an opportunity for genuine leadership to distinguish itself.</p><p>Less defensible leadership</p><p>More valuable leadership</p><p>Supervising routine activity</p><p>Orchestrating people and agents</p><p>Measuring visible busyness</p><p>Evaluating outcomes and quality</p><p>Exercising positional authority</p><p>Exercising informed judgment</p><p>Passing responsibility downward</p><p>Owning decisions and consequences</p><p><a href="https://images.openai.com/static-rsc-4/y_0hItSbnho6so_DWsxHvRdfJtflKa9N-89Bbm_LStHyGfrvIb5QAUDWMTTJ-mII5qlL5muZGst3VEfO42DC_IHLKzpTlOiDGfCxAzTsolUN1VlLpCsB6na1oH_QnyTcJn9tZ5y3d_P91F3XLiSi3ua2s0EivI0ajpBZPyBqXwY?purpose=inline">Human leaders and AI assistants collaborate on outcomes</a></p><p><a href="https://hypersightlabs.com/post/Beyond-the-Hype-How-IA-Is-Actual-Repashaping-Innovation-Strategy-in-2026?utm_source=chatgpt.com">hypersightlabs.com</a></p><p>Future leaders will need technical literacy, realistic planning, governance, ethical people management, and accountability. Not another certification, another dashboard, or another transformation slogan.</p><p>Actual competence.</p><h3>The Uncomfortable Reality</h3><p>For decades, talented employees have rescued flawed strategies, absorbed unrealistic commitments, and protected leadership careers through personal sacrifice.</p><p>AI will not automatically eliminate incompetent leadership. Organizations must choose to reward judgment over appearances and stop treating exhaustion as proof of dedication.</p><p>But when execution is automated and decisions become easier to trace, leadership’s value can face sharper scrutiny.</p><p>Your authority was never proof of competence. Your team’s exhaustion was never evidence of leadership excellence.</p><p>The real disruption may not be AI replacing the people doing the work.</p><p>It may be organizations finally questioning the people making the decisions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=269ccb551c7a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cognitive Offloading: The Dilemma Between Full Automation and Human-Centric AI Orchestration]]></title>
            <link>https://ai.plainenglish.io/cognitive-offloading-the-dilemma-between-full-automation-and-human-centric-ai-orchestration-88a67403fe1a?source=rss-e6b31ac66f1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/88a67403fe1a</guid>
            <category><![CDATA[cognitive-offloading]]></category>
            <category><![CDATA[human-ai-collaboration]]></category>
            <category><![CDATA[ai-productivity]]></category>
            <category><![CDATA[future-of-work-2026]]></category>
            <category><![CDATA[ai-orchestration]]></category>
            <dc:creator><![CDATA[Khurram Sultan]]></dc:creator>
            <pubDate>Sun, 02 Aug 2026 03:17:23 GMT</pubDate>
            <atom:updated>2026-08-02T03:17:23.844Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>AI can now perform parts of our research, planning, execution, and review. The real question is no longer whether we can delegate cognitive work, but how much control humans should retain.</em></strong></p><p>Humans have always found ways to think outside their own heads.</p><p>We write notes because memory is unreliable. We use calendars to track appointments, maps to find routes, calculators to solve equations, and spreadsheets to organize information.</p><p>Researchers describe this as <strong>cognitive offloading</strong>: using physical actions or external tools to reduce the mental processing required to complete a task.</p><p>For most of history, these external tools were passive.</p><p>A notebook stored what we wrote. A calculator processed the numbers we entered. A calendar reminded us about events we had already scheduled.</p><p>Artificial intelligence changes the relationship.</p><p>AI can receive a broader goal, interpret instructions, generate a plan, perform parts of the work, review an output, and recommend what should happen next.</p><p>We are no longer offloading only memory and calculation.</p><p>We are no longer offloading only memory and calculation. AI can now support almost the entire work cycle:</p><ul><li>gathering and organizing information;</li><li>understanding problems and creating plans;</li><li>producing written, visual, or technical content;</li><li>analyzing evidence and comparing options;</li><li>carrying out defined tasks;</li><li>coordinating work across people, tools, and agents;</li><li>checking results and identifying possible errors.</li></ul><p>This creates an important dilemma:</p><blockquote><em>Should we move towards full automation, or should AI remain under human-centric orchestration?</em></blockquote><p>The answer is not simply “human” or “AI.”</p><p>The better question is:</p><blockquote><em>Which parts of thinking and execution can safely be delegated, and which parts must remain under meaningful human control?</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qzrrwirw_WXMDWTHYCm5sw.png" /></figure><h3>From Cognitive Offloading to Cognitive Orchestration</h3><p>Cognitive offloading normally helps us conserve mental energy.</p><p>A student creates a revision checklist instead of remembering every topic. A project manager records tasks in a planning system rather than tracking everything mentally. A developer uses automated tests instead of manually checking every part of an application.</p><p>The tool reduces the burden of remembering or repeating work.</p><p>AI expands this idea because it can participate in the task itself.</p><h3>Figure 1: The evolution of cognitive offloading</h3><pre>MEMORY OFFLOADING<br>Notes, calendars and checklists<br>           ↓<br>TASK OFFLOADING<br>Calculators, software and automation<br>           ↓<br>COGNITIVE ORCHESTRATION<br>AI helps plan, execute, coordinate and review</pre><p><strong>Source note:</strong> The definition of cognitive offloading is based on Risko and Gilbert. The three-stage progression above is a conceptual framework developed for this article.</p><p>In the first stage, the human performs the work but uses tools to remember information.</p><p>In the second stage, software performs clearly defined actions.</p><p>In the third stage, AI may help decide which actions should be performed and how different parts of the task should be organized.</p><p>This is a major change.</p><p>The human is no longer using only a tool.</p><p>The human may be managing a system of tools that can produce decisions, recommendations, and actions.</p><h3>The Classical Human-Curated Workflow</h3><p>Traditional organizations are designed around humans completing and supervising work.</p><p>Imagine a company preparing a research report.</p><p>A manager defines the goal. A researcher gathers information. An analyst evaluates it. A writer prepares the report. A reviewer checks the work. A senior person approves the final version.</p><h3>Figure 2: Classical human-curated workflow</h3><pre>Human receives task<br>        ↓<br>Human understands requirements<br>        ↓<br>Human creates plan<br>        ↓<br>Human performs work<br>        ↓<br>Human reviewer checks result<br>        ↓<br>Human approves and releases</pre><p>Each stage may be performed by a different person, but the workflow remains human-curated.</p><p>People control the information, decisions, exceptions, and final outcome.</p><p>This model has important strengths.</p><p>Humans can understand social context, notice unusual situations, challenge unclear requirements, and take responsibility for difficult decisions.</p><p>It also has limitations.</p><p>People have limited time, memory, and attention. They become tired. They may spend hours on repetitive tasks that require little creativity or judgment.</p><p>AI promises to reduce that burden.</p><h3>The Full-Automation Model</h3><p>Full automation aims to remove most human involvement after the goal has been provided.</p><h3>Figure 3: Fully automated AI workflow</h3><pre>Human provides goal<br>        ↓<br>AI interprets task<br>        ↓<br>AI creates plan<br>        ↓<br>AI agents execute<br>        ↓<br>AI reviews output<br>        ↓<br>System approves or acts</pre><p>This model is attractive because it promises speed, scale, consistency, and continuous operation.</p><p>For routine and predictable tasks, those benefits can be real.</p><p>A system may automatically:</p><ul><li>classify documents;</li><li>detect duplicate records;</li><li>produce regular summaries;</li><li>format standard reports;</li><li>route common requests;</li><li>run repeatable checks.</li></ul><p>When the rules are clear and mistakes are easy to detect or reverse, full automation may be sensible.</p><p>The difficulty begins when a task is uncertain, complex, or high-impact.</p><p>An AI system may complete every step in its plan and still solve the wrong problem.</p><p>It may misunderstand the original goal, use weak evidence, overlook an exception, or confidently present an incorrect conclusion.</p><p>Without meaningful human involvement, those problems can move directly from AI output to real-world action.</p><h3>The Human-Centric Orchestration Model</h3><p>Human-centric orchestration offers a middle path.</p><p>AI performs much of the routine execution, but humans retain control of direction, risk, exceptions, and final accountability.</p><h3>Figure 4: Human-centric AI orchestration</h3><pre>Human defines goal, limits and success criteria<br>                    ↓<br>             AI proposes plan<br>                    ↓<br>            AI agents execute<br>                    ↓<br>     Tests, sources, logs and evidence<br>                    ↓<br>        AI or automated review checks<br>                    ↓<br>        Human reviews decision package<br>              ↙               ↘<br>        Approve             Redirect,<br>        or release          stop or escalate</pre><p>The human does not need to perform every task manually.</p><p>Instead, the human becomes the <strong>orchestrator</strong>.</p><p>The role is similar to that of an orchestra conductor.</p><p>The conductor does not personally play every instrument. The conductor ensures that different performers follow the same composition, enter at the correct time, and contribute to one complete outcome.</p><p>A human AI orchestrator performs a similar function.</p><p>The orchestrator:</p><ul><li>defines the real objective;</li><li>sets boundaries and permissions;</li><li>decides which work can be delegated;</li><li>monitors progress;</li><li>detects circular or unproductive work;</li><li>asks for evidence;</li><li>handles exceptions;</li><li>decides when to stop or redirect;</li><li>approves important outcomes;</li><li>remains accountable for the result.</li></ul><h3>Three Models for Organizing Human and AI Work</h3><h3>1. Human-Curated Workflow</h3><p><strong>Direction:</strong> A human manager or worker defines and controls the work.</p><p><strong>Execution:</strong> People perform most of the tasks.</p><p><strong>Review:</strong> A human reviewer checks the result.</p><p><strong>Final decision:</strong> A human approves and releases the work.</p><p><strong>Best suited for:</strong> Work requiring context, judgment, creativity, or direct human involvement.</p><p><strong>Main risk:</strong> The process can be slow, expensive, and difficult to scale.</p><p><strong>Simple flow:</strong></p><pre>Human directs → Human executes → Human reviews → Human decides</pre><h3>2. Full AI Automation</h3><p><strong>Direction:</strong> A human provides the initial goal, after which the AI manages the workflow.</p><p><strong>Execution:</strong> AI systems and agents perform most of the work.</p><p><strong>Review:</strong> Automated checks or AI reviewers assess the result.</p><p><strong>Final decision:</strong> The system may approve or act without human involvement.</p><p><strong>Best suited for:</strong> Repetitive, predictable, low-risk, and easily reversible tasks.</p><p><strong>Main risk:</strong> Errors may move through the complete workflow without being noticed.</p><p><strong>Simple flow:</strong></p><pre>Human sets goal → AI plans → AI executes → AI reviews → AI acts</pre><h3>3. Human-Centric AI Orchestration</h3><p><strong>Direction:</strong> A human defines the goal, limits, risks, and success criteria.</p><p><strong>Execution:</strong> AI agents perform most of the routine work.</p><p><strong>Review:</strong> AI performs initial checks and produces evidence. The human reviews important findings, exceptions, and uncertainties.</p><p><strong>Final decision:</strong> A human approves high-impact or difficult-to-reverse outcomes.</p><p><strong>Best suited for:</strong> Complex AI-assisted work requiring both productivity and responsible oversight.</p><p><strong>Main risk:</strong> Managing, reviewing, and correcting AI agents can create significant coordination overhead.</p><p><strong>Simple flow:</strong></p><pre>Human directs → AI executes → Evidence is produced → Human decides</pre><p>The key difference is not simply who performs the work. It is <strong>where direction, verification, decision-making, and accountability remain</strong>.</p><h3>The Hidden Productivity Equation</h3><p>AI is often presented as a simple productivity multiplier.</p><p>Give the work to AI, receive the result, and save time.</p><p>In practice, the calculation is more complicated.</p><p>A useful way to think about it is:</p><blockquote><strong><em>AI productivity gain = work avoided − supervision, correction, and coordination costs</em></strong></blockquote><p>For a simple task, the cost of review may be tiny.</p><p>AI may format a document or summarize meeting notes, and the human may need only a quick check.</p><p>For difficult work, supervision costs can become much larger.</p><p>The human may have to:</p><ul><li>correct hallucinated information;</li><li>identify unsupported assumptions;</li><li>restate misunderstood requirements;</li><li>stop an agent repeating the same approach;</li><li>compare conflicting agent conclusions;</li><li>check whether the correct tests were run;</li><li>repair work that drifted away from the original goal.</li></ul><p>The AI may generate a large amount of activity without producing equivalent progress.</p><p>This creates a form of <strong>false productivity</strong>.</p><p>Many plans, drafts, messages, or code changes may be produced, but the real problem may remain unsolved.</p><h3>When Cognitive Offloading Becomes Cognitive Surrender</h3><p>Cognitive offloading is useful when a tool reduces unnecessary mental effort while the human remains able to understand and control the task.</p><p>Cognitive surrender happens when the human stops questioning the tool.</p><p>The difference can be illustrated simply:</p><pre>COGNITIVE OFFLOADING<br>“I will use AI to help perform this work,<br>but I remain responsible for understanding it.”</pre><pre>COGNITIVE SURRENDER<br>“The AI produced and checked it,<br>so it must be correct.”</pre><p>This second belief is dangerous because AI can sound confident even when it is wrong.</p><p>A polished explanation may feel reliable because it is well written. A detailed report may feel trustworthy because it contains headings, numbers, and professional language.</p><p>But presentation quality is not the same as evidence quality.</p><p>The essential rule should be:</p><blockquote><strong><em>AI output is a proposal, not proof.</em></strong></blockquote><h3>Automation Bias: “If AI Said It, It Must Be Fine”</h3><p>Researchers studying automation have long warned about overreliance on automated systems.</p><p>Parasuraman and Riley described different relationships with automation, including appropriate use, misuse through overreliance, disuse through rejection, and abuse through inappropriate design or deployment. They also identified trust, workload, and perceived risk as factors influencing how people use automation.</p><p>In an AI workflow, automation bias may appear when a person assumes:</p><ul><li>an AI-generated fact has already been verified;</li><li>an AI reviewer must be independent;</li><li>a passing automated check proves the entire solution;</li><li>several agreeing agents provide independent confirmation;</li><li>a confident answer indicates high certainty;</li><li>a detailed explanation must be based on strong evidence.</li></ul><p>None of these assumptions is automatically safe.</p><p>Two AI agents may agree because they share similar training patterns, instructions, models, or incorrect inputs.</p><p>An AI reviewer may repeat the original agent’s blind spot instead of detecting it.</p><p>A test may pass while failing to cover the behaviour that actually changed.</p><p>Agreement is useful, but agreement is not proof.</p><h3>The Out-of-the-Loop Problem</h3><p>Heavy automation can also weaken the human’s ability to intervene.</p><p>Endsley and Kiris described the <strong>out-of-the-loop performance problem</strong>, in which operators of automated systems may lose situation awareness or become less capable of taking over when automation fails. Their work linked this problem to issues including passive information processing, reduced awareness, complacency, and possible skill loss.</p><p>This problem becomes especially important in AI-based work.</p><p>A person may remain listed as the “human reviewer” while understanding very little about what the AI actually did.</p><p>When an unusual failure occurs, that person may not know:</p><ul><li>which decisions were made;</li><li>which assumptions were used;</li><li>which tools were called;</li><li>which information was ignored;</li><li>where the process went wrong;</li><li>how to safely take control.</li></ul><p>A human approval button does not automatically create meaningful human oversight.</p><p>For oversight to be real, the human must receive understandable evidence and retain enough knowledge to challenge the system.</p><h3>The Risks of Excessive AI Dependence</h3><h3>1. Hallucinations</h3><p>AI may invent facts, references, calculations, test results, or explanations.</p><p>The output may be grammatically perfect while being factually weak.</p><h3>2. Error propagation</h3><p>One agent may produce an incorrect result.</p><p>A second agent may treat that result as true. A third agent may use it to make a recommendation.</p><p>The original mistake becomes harder to identify as it moves through the workflow.</p><h3>3. Context drift</h3><p>Over a long task, agents may gradually move away from the original objective.</p><p>They may complete the steps in their plan without solving the actual problem.</p><h3>4. Circular execution</h3><p>An agent may repeatedly revise the same solution, regenerate the same plan, or retry the same failed action.</p><p>This produces activity without useful movement.</p><h3>5. Review fatigue</h3><p>AI can generate information faster than humans can inspect it.</p><p>When humans face hundreds of pages, messages, tests, or changes, they may begin approving work without examining it carefully.</p><h3>6. Skill atrophy</h3><p>When people stop performing a task, they may gradually lose the ability to do it.</p><p>More importantly, they may lose the ability to recognize when the AI is doing it badly.</p><h3>7. Accountability diffusion</h3><p>When multiple agents, models, tools, and reviewers are involved, responsibility can become unclear.</p><p>The AI cannot accept professional, ethical, or legal accountability.</p><p>Responsibility remains with the people and organizations deploying the system.</p><h3>Controls for Human-Centric AI Orchestration</h3><p>Good orchestration requires more than a strong prompt.</p><p>Human-centric orchestration requires more than a good prompt. It needs controls that keep AI work focused, verifiable, and accountable.</p><p>NIST’s AI Risk Management Framework organizes AI risk work around governance, mapping context, measuring risks, and managing those risks. It is intended to support organizations in incorporating trustworthiness considerations into the design, development, use, and evaluation of AI systems.</p><p>The following practical controls translate those ideas into an agent-based workflow.</p><h3>Goal Control</h3><p><strong>Question:</strong> Are we solving the correct problem?</p><p><strong>Control:</strong> Define the objective, expected result, constraints, and acceptance criteria before the work begins.</p><h3>Scope Control</h3><p><strong>Question:</strong> What may the AI access or change?</p><p><strong>Control:</strong> Limit the tools, files, systems, permissions, and actions available to each agent.</p><h3>Evidence Control</h3><p><strong>Question:</strong> What proves that the result is correct?</p><p><strong>Control:</strong> Require supporting evidence such as tests, citations, calculations, logs, screenshots, or code changes.</p><h3>Loop Control</h3><p><strong>Question:</strong> Is the system making progress or repeating itself?</p><p><strong>Control:</strong> Set retry limits and detect repeated plans, actions, or failures.</p><h3>Context Control</h3><p><strong>Question:</strong> Has the work moved away from the original goal?</p><p><strong>Control:</strong> Compare intermediate results with the original requirements at defined checkpoints.</p><h3>Risk Control</h3><p><strong>Question:</strong> What happens if the AI is wrong?</p><p><strong>Control:</strong> Apply stronger checks to financial, legal, security, health, hiring, or public-facing decisions.</p><h3>Independent Review Control</h3><p><strong>Question:</strong> Is the review genuinely independent?</p><p><strong>Control:</strong> Use different methods, evidence, instructions, tools, or human reviewers instead of relying only on another similar AI agent.</p><h3>Escalation Control</h3><p><strong>Question:</strong> When should the AI stop?</p><p><strong>Control:</strong> Define clear thresholds for uncertainty, repeated failure, conflicting results, and high-risk decisions.</p><h3>Approval Control</h3><p><strong>Question:</strong> Who may authorize the final action?</p><p><strong>Control:</strong> Require human approval for important, expensive, public, sensitive, or difficult-to-reverse actions.</p><h3>Traceability Control</h3><p><strong>Question:</strong> Can we reconstruct what happened?</p><p><strong>Control:</strong> Record instructions, tool use, decisions, changes, evidence, failures, and approvals.</p><h3>Skill-Retention Control</h3><p><strong>Question:</strong> Can the human still understand and challenge the work?</p><p><strong>Control:</strong> Require explanations, periodic manual checks, and continued human involvement in important tasks.</p><h3>Accountability Control</h3><p><strong>Question:</strong> Who owns the final outcome?</p><p><strong>Control:</strong> Assign a named human or organizational owner before the workflow begins.</p><p>These controls do not guarantee that AI will never fail. They make failures easier to detect, investigate, and manage.</p><h3>Humans Should Review Evidence, Not Every Token</h3><p>One objection to human-centric orchestration is that reviewing AI work can become a full-time task.</p><p>That is true if the human is expected to read every message, thought, draft, and intermediate output.</p><p>A better approach is to produce a <strong>decision package</strong>.</p><h3>Figure 5: The decision package</h3><pre>ORIGINAL GOAL<br>What was the system asked to achieve?</pre><pre>PROPOSED RESULT<br>What did the agents produce?</pre><pre>CHANGES MADE<br>What files, records or decisions changed?</pre><pre>EVIDENCE<br>Which tests, sources, calculations or logs support it?</pre><pre>FAILURES AND UNCERTAINTY<br>What did not work, and what remains unknown?</pre><pre>RISK<br>What happens if the result is wrong?</pre><pre>DECISION REQUIRED<br>Approve, reject, redirect or escalate?</pre><p>The human should be able to understand the state of the task without reading the entire agent conversation.</p><p>This changes the human role from line-by-line inspector to evidence assessor and accountable decision-maker.</p><h3>Choosing Between Full Automation and Human Orchestration</h3><p>Not every task requires the same amount of human involvement.</p><p>The correct operating model depends on the task’s complexity, uncertainty, impact, reversibility, and available methods of verification.</p><h3>Full Automation</h3><p>Use full automation when:</p><ul><li>the rules are clear;</li><li>the task is repetitive;</li><li>the impact of failure is low;</li><li>errors are easy to detect;</li><li>actions are easy to reverse.</li></ul><p>Examples include document formatting, duplicate detection, file classification, routine data entry, and standard report generation.</p><p><strong>Minimum controls:</strong> Logging, monitoring, error alerts, and recovery procedures.</p><pre>Human sets goal<br>        ↓<br>AI performs task<br>        ↓<br>Automated checks run<br>        ↓<br>System completes action</pre><h3>Automation With Human Monitoring</h3><p>Use this model when the task is mostly predictable but unusual cases may still occur.</p><p>Examples include routine reporting, support-request routing, content tagging, and basic operational monitoring.</p><p>The AI completes normal work while a human reviews samples, monitors performance, or handles exceptions.</p><p><strong>Minimum controls:</strong> Exception alerts, random sampling, performance checks, and the ability for a human to intervene.</p><pre>AI performs routine work<br>        ↓<br>System monitors outcomes<br>        ↓<br>Exceptions are flagged<br>        ↓<br>Human intervenes when needed</pre><h3>Human-on-the-Loop</h3><p>Use this model when AI may operate independently within defined limits, but a person must remain able to observe and stop the system.</p><p>Examples include support triage, fraud alerts, content moderation, and operational recommendations.</p><p><strong>Minimum controls:</strong> Live monitoring, stop controls, escalation rules, and review thresholds.</p><pre>AI acts within defined limits<br>        ↓<br>Human monitors the process<br>        ↓<br>Unusual or risky cases are escalated</pre><h3>Human-Centric AI Orchestration</h3><p>Use human-centric orchestration when:</p><ul><li>the task is complex;</li><li>evidence may be incomplete;</li><li>several agents or tools are involved;</li><li>the cost of failure is meaningful;</li><li>human judgment remains important.</li></ul><p>Examples include software changes, hiring support, financial analysis, security investigations, strategic research, and legal-document review.</p><p><strong>Minimum controls:</strong> Defined scope, evidence requirements, independent review, escalation rules, and human approval.</p><pre>Human defines goal and boundaries<br>        ↓<br>AI agents plan and execute<br>        ↓<br>Evidence and uncertainties are collected<br>        ↓<br>Human reviews and decides</pre><h3>Human-in-the-Loop Approval</h3><p>Use this model when AI may prepare the work or recommendation, but it should not take the final action independently.</p><p>This is appropriate when the action is public, expensive, sensitive, high-impact, or difficult to reverse.</p><p>Examples include production deployments, financial transfers, hiring decisions, contract approvals, public publication, and access-control changes.</p><p><strong>Minimum controls:</strong> Explicit authorization, identity verification, complete evidence, and an audit record.</p><pre>AI prepares recommendation<br>        ↓<br>Human reviews the evidence<br>        ↓<br>Human approves or rejects the action</pre><h3>Human-Led Investigation</h3><p>Use this model when:</p><ul><li>the goal is unclear;</li><li>evidence is weak or contradictory;</li><li>the situation is new;</li><li>no reliable method exists for verifying the AI result.</li></ul><p>Examples include novel incidents, disputed decisions, strategic crises, unusual security events, and ethical conflicts.</p><p>AI may assist with bounded research or analysis, but the human remains responsible throughout.</p><p><strong>Minimum controls:</strong> Pause automation, preserve evidence, clarify the problem, and maintain human ownership.</p><pre>Human investigates the problem<br>        ↓<br>AI assists with bounded tasks<br>        ↓<br>Human interprets the evidence<br>        ↓<br>Human makes the decision</pre><p>The operating model should not be selected only by asking whether AI can perform the task.</p><p>The more important question is:</p><blockquote><strong><em>What is the cost of allowing AI to be wrong?</em></strong></blockquote><h3>A Practical Software Development Example</h3><p>Consider a software change.</p><h3>Traditional workflow</h3><pre>Developer reads issue<br>        ↓<br>Developer investigates code<br>        ↓<br>Developer creates patch<br>        ↓<br>Developer runs tests<br>        ↓<br>Reviewer checks changes<br>        ↓<br>Maintainer approves</pre><h3>Fully automated workflow</h3><pre>AI reads issue<br>        ↓<br>AI creates plan<br>        ↓<br>AI modifies code<br>        ↓<br>AI runs selected tests<br>        ↓<br>AI reviews its work<br>        ↓<br>AI opens or merges change</pre><p>This may be fast, but it contains serious risks.</p><p>The AI may modify the wrong behaviour, run irrelevant tests, overlook a regression, or confidently declare success based on incomplete evidence.</p><h3>Human-orchestrated workflow</h3><pre>Human defines requirement and acceptance criteria<br>                    ↓<br>AI planner identifies affected components<br>                    ↓<br>Coding agent creates proposed patch<br>                    ↓<br>Testing agent runs relevant tests<br>                    ↓<br>Reviewer agent searches for gaps and regressions<br>                    ↓<br>System prepares evidence package<br>                    ↓<br>Human approves, redirects or rejects</pre><p>The human does not need to write every line of code.</p><p>However, the human retains control of:</p><ul><li>what behaviour must change;</li><li>what must remain unchanged;</li><li>what tests count as evidence;</li><li>which risks are acceptable;</li><li>whether the patch should be released.</li></ul><p>This is cognitive offloading without surrendering accountability.</p><h3>The New Human Role</h3><p>The future of work may involve fewer people manually performing every step.</p><p>But this does not mean humans become unnecessary.</p><p>Their role changes.</p><h3>From worker to goal setter</h3><p>The human defines the real problem and what a successful result should look like.</p><h3>From doer to delegator</h3><p>The human decides which tasks can be safely assigned to AI.</p><h3>From manager to orchestrator</h3><p>The human coordinates agents, dependencies, limits, and exceptions.</p><h3>From reviewer to evidence assessor</h3><p>The human evaluates proof rather than trusting presentation.</p><h3>From approver to accountable owner</h3><p>The human or organization accepts responsibility for the final outcome.</p><p>This is not a reduction in responsibility.</p><p>It is a movement from task-level responsibility to system-level responsibility.</p><h3>The Better Future Is Not Human or AI</h3><p>The debate is often presented as a competition:</p><p>Will humans perform the work, or will AI replace them?</p><p>That framing is too narrow.</p><p>A more useful question is:</p><blockquote><em>How should humans and AI divide execution, judgment, control, and accountability?</em></blockquote><p>AI is strong at speed, scale, repetition, summarization, pattern detection, and generating possible solutions.</p><p>Humans remain essential for purpose, context, values, exceptions, uncertainty, and responsibility.</p><p>A strong workflow combines these strengths.</p><pre>AI performs routine execution<br>             +<br>Controls restrict and verify the process<br>             +<br>Evidence supports the result<br>             +<br>Humans manage risk and exceptions<br>             =<br>Responsible AI orchestration</pre><h3>Conclusion</h3><p>Cognitive offloading has always been part of human progress.</p><p>Writing reduced the burden on memory. Calculators reduced the burden of arithmetic. Software reduced the burden of repetitive execution.</p><p>AI now allows us to offload parts of planning, analysis, coordination, and review.</p><p>That creates an enormous opportunity.</p><p>It also creates a danger.</p><p>When cognitive offloading is well designed, it frees people to focus on judgment, creativity, and difficult decisions.</p><p>When it is poorly designed, it becomes cognitive surrender: people approve systems they no longer understand, accept conclusions they cannot verify, and remain responsible for decisions they did not meaningfully control.</p><p>Full automation may be appropriate for predictable, low-risk, and reversible work.</p><p>For complex or consequential tasks, human-centric orchestration offers a stronger model.</p><p>The human does not need to perform every step.</p><p>But the human must still define the goal, set the boundaries, demand evidence, recognize uncertainty, control important decisions, and remain accountable for the outcome.</p><p>The central principle is simple:</p><blockquote><strong><em>Delegate execution, but never delegate accountability.</em></strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IhpNltGjjcXfaxt03u7Njw.png" /></figure><h3>References (Verified)</h3><ol><li><strong>Risko, E. F., and Gilbert, S. J. — “Cognitive Offloading.”</strong> <em>Trends in Cognitive Sciences</em>, 2016. The paper defines cognitive offloading and examines how humans use physical action and external tools to reduce cognitive demand. <a href="https://doi.org/10.1016/j.tics.2016.07.002">https://doi.org/10.1016/j.tics.2016.07.002</a></li><li><strong>Parasuraman, R., and Riley, V. — “Humans and Automation: Use, Misuse, Disuse, Abuse.”</strong> <em>Human Factors</em>, 1997. The paper examines appropriate and inappropriate reliance on automation and the roles of trust, workload, and risk. <a href="https://doi.org/10.1518/001872097778543886">https://doi.org/10.1518/001872097778543886</a></li><li><strong>Endsley, M. R., and Kiris, E. O. — “The Out-of-the-Loop Performance Problem and Level of Control in Automation.”</strong> <em>Human Factors</em>, 1995. The study examines how automation can affect situation awareness and the ability to take over when systems fail. <a href="https://doi.org/10.1518/001872095779064555">https://doi.org/10.1518/001872095779064555</a></li><li><strong>National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework 1.0.</strong> NIST, 2023. The framework provides guidance for governing, mapping, measuring, and managing AI risks. <a href="https://doi.org/10.6028/NIST.AI.100-1">https://doi.org/10.6028/NIST.AI.100-1</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=88a67403fe1a" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/cognitive-offloading-the-dilemma-between-full-automation-and-human-centric-ai-orchestration-88a67403fe1a">Cognitive Offloading: The Dilemma Between Full Automation and Human-Centric AI Orchestration</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[After 18 Hours Managing AI Agents, I Realized I Needed a Project Manager, HR, and an Accountant]]></title>
            <link>https://ai.plainenglish.io/after-18-hours-managing-ai-agents-i-realized-i-needed-a-project-manager-hr-and-an-accountant-19e86414b549?source=rss-e6b31ac66f1f------2</link>
            <guid isPermaLink="false">https://medium.com/p/19e86414b549</guid>
            <category><![CDATA[token-economics]]></category>
            <category><![CDATA[ai-needs-a-manager]]></category>
            <category><![CDATA[agentic-fixation]]></category>
            <category><![CDATA[ai-workplace-satire]]></category>
            <category><![CDATA[autonomous-ai-teams]]></category>
            <dc:creator><![CDATA[Khurram Sultan]]></dc:creator>
            <pubDate>Thu, 23 Jul 2026 21:53:33 GMT</pubDate>
            <atom:updated>2026-07-23T21:53:33.002Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OJqH5QfK_5vR3OMad85fJw.jpeg" /></figure><p><strong><em>After spending 18 hours managing AI agents on a task estimated at seven hours for a human, I began to see them less as autonomous workers and more as a dysfunctional corporate team. The article explores agentic fixation, shifting problem scope, circular discovery, and the hidden cost of token-based productivity. It asks whether AI inefficiency is always a prompting problem — or whether the models themselves need better management, accountability, and efficiency audits.</em></strong></p><p>I recently assigned a software-engineering task to a small team of AI agents.</p><p>One agent acted as the architect. Another handled development. A third was responsible for quality control.</p><p>On paper, it looked like the perfect autonomous engineering team: no meetings, no scheduling conflicts, no annual leave, and no arguments about whether the ticket contained enough information.</p><p>The task was estimated to take an experienced human engineer approximately six and a half to seven hours.</p><p>Eighteen hours later, the agents were still working.</p><p>At that point, I started thinking that the team needed additional support.</p><p>Perhaps a project manager was required to keep the architect and developer aligned.</p><p>Maybe HR was needed to listen to their grievances about unclear requirements, missing context, tool limitations, and the unreasonable expectations placed upon them.</p><p>An office assistant could bring them juice whenever their focus started to drop and they suddenly abandoned the current approach for an entirely different solution.</p><p>They might also request a raise — not in salary, but in the number of tokens they were allowed to consume.</p><p>Naturally, an accountant would then be needed to track those tokens, calculate the cost of repeated implementation attempts, and ensure that the agents were promptly rewarded once their efficiency improved.</p><p>It sounds ridiculous.</p><p>But after a sufficiently long AI-agent run, the comparison with a human organization becomes surprisingly difficult to ignore.</p><h3>AI Agents Are Learning More Than Our Code</h3><p>Large language models are trained on material produced by humans.</p><p>That includes our source code, technical documentation, design discussions, project-management practices, disagreements, excuses, organizational habits, and institutional dysfunction.</p><p>We should therefore not be surprised when AI systems reproduce more than human knowledge.</p><p>They also reproduce human attitudes.</p><p>An architect agent may overcomplicate a straightforward requirement.</p><p>A developer agent may confidently implement the wrong interpretation.</p><p>A quality-control agent may focus on a minor compliance detail while overlooking the actual behavioural failure.</p><p>One agent may reverse another agent’s valid decision. All three may then provide detailed explanations for why their individual actions were technically reasonable.</p><p>Humans perform similar behaviours through meetings, emails, tickets, design documents, and several working days.</p><p>AI agents can compress the entire corporate experience into one extremely long terminal session.</p><h3>Agentic Fixation: When an Agent Falls in Love With Its Solution</h3><p>During the eighteen-hour run, I noticed a behaviour that I have started calling <strong>agentic fixation</strong>.</p><p>Agentic fixation occurs when an AI agent becomes excessively committed to a particular interpretation, architecture, or solution path.</p><p>Instead of stepping back and comparing alternatives, it continues refining, defending, and repairing the approach it has already selected.</p><p>The agent is not necessarily stuck in a literal loop. It may continue generating new code, tests, explanations, abstractions, and architectural changes.</p><p>From the outside, the work appears to be progressing.</p><p>But nearly all that progress remains trapped inside one solution domain.</p><p>A developer agent may decide that the problem belongs in the persistence layer and spend several rounds improving database queries.</p><p>An architect may conclude that a new abstraction is necessary and repeatedly redesign that abstraction.</p><p>The QC agent may accept the same underlying assumption and concentrate only on whether the chosen design has been implemented completely.</p><p>Each agent produces more detail, but none asks the most important question:</p><p><strong>Should we still be solving the problem this way?</strong></p><p>Human engineers experience the same tendency. We become attached to the first explanation that appears plausible, especially after investing substantial effort in it.</p><p>The more time we spend on an approach, the harder it becomes to abandon it.</p><p>AI agents appear capable of reproducing this tendency at token-generating speed.</p><p>The result is not merely wasted computation. Agentic fixation can make an incorrect solution look increasingly sophisticated.</p><p>Every additional patch, test, and explanation strengthens the appearance of progress while making it more expensive to reconsider the original assumption.</p><p>The agent begins to treat continued refinement as a substitute for exploring alternatives.</p><h3>The Scope Changed Every Time We Understood More</h3><p>Agentic fixation was only part of the problem.</p><p>The task also went through a process I would describe as <strong>elaborative discovery</strong>.</p><p>After each implementation round, the next set of artifacts revealed something new.</p><p>A test exposed a previously unnoticed behavioural requirement.</p><p>A verifier result showed that the acceptance criteria were narrower — or sometimes broader — than expected.</p><p>A code review revealed that the problem belonged in a different architectural layer.</p><p>A new artifact demonstrated that the previous interpretation was technically valid but incomplete.</p><p>As our understanding improved, the scope of the implementation changed.</p><p>This is normal in complex engineering work. Requirements are rarely understood perfectly at the beginning.</p><p>Implementation produces evidence. Evidence improves understanding. Improved understanding changes the implementation.</p><p>The problem begins when this discovery process stops converging.</p><p>The architect starts in one solution domain. The developer implements the design. The QC agent reviews the resulting artifacts and identifies a deeper concern. The architect then reclassifies the problem and moves it into another domain.</p><p>The cycle may look like this:</p><p><strong>application logic → persistence → validation → serialization → test infrastructure → application logic</strong></p><p>This is not necessarily regression.</p><p>A regression means that previously correct behaviour has been broken.</p><p>In this case, the agents may not be breaking completed work at all. Instead, they are repeatedly changing their understanding of where the real problem belongs.</p><p>Each round can be locally reasonable.</p><p>The new artifacts genuinely provide more information. The revised scope may genuinely represent a better understanding of the task.</p><p>Yet the overall process still moves in circles because the solution keeps hopping from one domain to another.</p><p>The agents are not returning to the same point because they forgot their previous work.</p><p>They are returning because every new layer of understanding reframes the original problem.</p><p>That distinction matters.</p><p>Calling every repeated implementation a regression hides the real issue. The problem is not always failure to preserve working code.</p><p>Sometimes the real failure is the inability to maintain a stable problem model.</p><h3>More Understanding Does Not Always Produce Convergence</h3><p>We often assume that giving an AI agent more context will naturally move it closer to the correct answer.</p><p>In practice, additional context may also increase the number of plausible interpretations.</p><p>A new artifact may resolve one uncertainty while introducing three more.</p><p>A failed test may prove that the implementation is wrong, but it may not reveal whether the defect belongs in the algorithm, interface, data model, test suite, execution environment, or verifier.</p><p>As agents receive more evidence, they can become more informed without becoming more decisive.</p><p>This produces a strange form of progress:</p><p><strong>Understanding improves, but convergence does not.</strong></p><p>The architect produces a more nuanced design.</p><p>The developer creates a more comprehensive implementation.</p><p>The QC agent reports a more detailed list of risks.</p><p>And the task moves further away from completion because the team has discovered a new way to define the problem.</p><p>This is where an AI project manager might genuinely help — not merely by assigning work, but by preserving the evolving problem model.</p><p>Such a coordinator would need to record:</p><ul><li>What has been conclusively established</li><li>Which assumptions remain uncertain</li><li>Why the scope changed</li><li>Which solution domains have already been explored</li><li>What evidence justified the latest architectural shift</li><li>What evidence would justify another shift</li><li>When further discovery is no longer worth its cost</li></ul><p>Without this discipline, iterative discovery becomes iterative reinvention.</p><h3>Knowing When to Let a Solution Go</h3><p>A capable agent should not only know how to improve a solution.</p><p>It should also know when to stop investing in it.</p><p>That requires more than code generation. It requires deliberate comparison between alternatives.</p><p>Before producing another patch, the agent should ask:</p><ul><li>What assumption is this solution based on?</li><li>What evidence supports that assumption?</li><li>What evidence contradicts it?</li><li>Which alternative solution domains remain unexplored?</li><li>Is this change addressing the root cause?</li><li>Are we preserving the approach only because we have already invested in it?</li><li>What specific result would cause us to abandon this solution?</li></ul><p>These questions create an exit condition for agentic fixation.</p><p>They also place boundaries around elaborative discovery.</p><p>New evidence should be allowed to update the scope, but every new interpretation should not automatically trigger another complete implementation cycle.</p><p>At some point, the expected benefit of further exploration becomes smaller than its cost.</p><p>That cost is not measured only in tokens.</p><p>It includes human supervision, delayed completion, review fatigue, context overload, and the growing difficulty of understanding which parts of the implementation are still intentional.</p><h3>Token Consumption Is Not the Same as Productive Work</h3><p>A long output can look impressive.</p><p>A detailed chain of actions can create the appearance of progress.</p><p>Multiple agents discussing a problem may resemble a serious engineering process.</p><p>But activity is not the same as productivity.</p><p>If a seven-hour human task becomes an eighteen-hour AI-agent marathon, we need more than a final pass-or-fail result.</p><p>We need to understand where the time and tokens went.</p><p>Was the task genuinely difficult?</p><p>Were the agents blocked by the environment?</p><p>Did the architect introduce unnecessary complexity?</p><p>Did the developer repeatedly patch symptoms instead of addressing the root cause?</p><p>Did the QC agent report false positives or false negatives?</p><p>Did the team remain fixated on one solution for too long?</p><p>Did it repeatedly migrate between solution domains without preserving a stable understanding of the task?</p><p>Did the agents continue generating work because the system rewarded continuation more directly than completion?</p><p>These questions matter because the commercial model of frontier AI is closely connected to usage.</p><p>AI laboratories are businesses. They need revenue, and token consumption is one of the primary ways in which that revenue is generated.</p><p>That does not mean every inefficient response is intentional.</p><p>It does mean that there is a structural tension between the user’s desire for efficient task completion and a pricing model in which longer interactions can produce more revenue.</p><p>That tension deserves serious examination.</p><h3>The User Is Often Blamed for Inefficient Prompting</h3><p>When an AI system consumes too many tokens or fails to complete a task efficiently, the first response is often that the user should have written a better prompt.</p><p>The prompt should have been clearer.</p><p>The requirements should have been more structured.</p><p>The constraints should have been stated explicitly.</p><p>The acceptance criteria should have been converted into executable tests.</p><p>The user should have anticipated every ambiguity before the task began.</p><p>There is some truth in this. Better instructions usually produce better results.</p><p>But this explanation can also become a convenient way to transfer responsibility from the model provider to the user.</p><p>Complex engineering tasks cannot always be fully specified before implementation begins.</p><p>Some requirements become visible only after code is executed, artifacts are inspected, assumptions are challenged, and competing interpretations are tested.</p><p>The user should provide clear objectives and constraints.</p><p>But the model should also be responsible for managing uncertainty efficiently.</p><p>If a system repeatedly forgets established context, reopens settled decisions, generates unnecessary files, performs redundant analysis, or confidently follows the wrong path, the issue cannot always be reduced to poor prompting.</p><p>At some point, model behaviour itself must be audited.</p><p>We should ask:</p><ul><li>Did the model use the available context efficiently?</li><li>Did it repeat work that had already been completed?</li><li>Did it preserve previous discoveries?</li><li>Did it compare alternatives before committing to a solution?</li><li>Did it distinguish uncertainty from evidence?</li><li>Did additional token consumption meaningfully improve the result?</li><li>Did the model move toward task completion, or merely continue the interaction?</li></ul><p>Without this kind of evaluation, token-based pricing creates an uncomfortable incentive.</p><p>The customer pays not only for useful reasoning, but also for repetition, indecision, abandoned approaches, context loss, and conversations required to correct model-created mistakes.</p><h3>We Need Efficiency Audits, Not Just Intelligence Benchmarks</h3><p>Frontier models are commonly evaluated on coding ability, reasoning, mathematical performance, tool use, and benchmark completion.</p><p>Those measurements are useful, but incomplete.</p><p>A model may eventually solve a task while consuming several times the expected effort.</p><p>Another model may generate an impressive amount of reasoning while repeatedly revisiting the same decisions.</p><p>A multi-agent system may appear sophisticated because several agents are active, even though much of their activity consists of transferring confusion from one agent to another.</p><p>We should therefore evaluate more than whether the final answer is correct.</p><p>Models should also be assessed on:</p><ul><li>Time to stable problem understanding</li><li>Tokens consumed before convergence</li><li>Number of abandoned implementation paths</li><li>Repetition of previously completed analysis</li><li>Ability to detect agentic fixation</li><li>Ability to preserve decisions across iterations</li><li>Frequency of unnecessary scope changes</li><li>Cost of correcting model-generated errors</li><li>Ratio of useful work to total generated content</li></ul><p>A system that completes a task correctly but uses three times the expected effort is not automatically efficient.</p><p>A system that produces thousands of tokens explaining and correcting its own mistakes should not place all responsibility on the person who wrote the original prompt.</p><h3>Perhaps AI Needs Management After All</h3><p>The original promise of AI agents was that they would reduce coordination overhead.</p><p>Instead, we may be creating digital organizations that reproduce many of the same coordination problems as human ones — only faster, more verbose, and billed by the token.</p><p>The architect needs clear boundaries.</p><p>The developer needs executable acceptance criteria.</p><p>The QC agent needs protection against false positives and false negatives.</p><p>The team needs a shared record of established facts, rejected alternatives, open assumptions, and scope changes.</p><p>Someone — or another system — must determine when improved understanding justifies a new direction and when the team is simply moving in circles.</p><p>In other words, the autonomous AI team may still need a project manager.</p><p>After an eighteen-hour run, perhaps HR, an accountant, and an office assistant carrying juice would not hurt either.</p><p>The larger lesson is not that AI agents are useless.</p><p>They can perform extraordinary work.</p><p>But we should evaluate them as systems, not magical employees.</p><p>They should be expected to recognize when they are becoming fixated, when discovery is becoming circular, when the scope has genuinely changed, and when it is time to let an approach go.</p><p>After eighteen hours, I did not simply have three AI agents working on a software problem.</p><p>I had an architect defending its architecture, a developer repeatedly rebuilding the implementation, and a QC agent continuously discovering that the real problem was somewhere else.</p><p>That was the moment I realized I did not only need a project manager.</p><p>I needed someone capable of saying:</p><p><strong>“We understand the problem better now. But we are not starting over again.”</strong></p><p>Otherwise, the future of autonomous work may look remarkably familiar:</p><p>Too many meetings, unclear ownership, expanding budgets — and everyone asking for a raise.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19e86414b549" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/after-18-hours-managing-ai-agents-i-realized-i-needed-a-project-manager-hr-and-an-accountant-19e86414b549">After 18 Hours Managing AI Agents, I Realized I Needed a Project Manager, HR, and an Accountant</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>