<?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 Munjae Song on Medium]]></title>
        <description><![CDATA[Stories by Munjae Song on Medium]]></description>
        <link>https://medium.com/@workinthenow?source=rss-22231eb9f058------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*iG2hXtjkW-huy-82MEySrg.jpeg</url>
            <title>Stories by Munjae Song on Medium</title>
            <link>https://medium.com/@workinthenow?source=rss-22231eb9f058------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 22:03:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@workinthenow/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[Map, Assign, Pilot, Sustain: The Four Steps of AI Transformation in Detail]]></title>
            <link>https://medium.com/@workinthenow/map-assign-pilot-sustain-the-four-steps-of-ai-transformation-in-detail-85b2ad604929?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/85b2ad604929</guid>
            <category><![CDATA[operations-management]]></category>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 11:21:01 GMT</pubDate>
            <atom:updated>2026-11-02T11:21:01.492Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What each step must answer, how it is evaluated, and what it must hand off to the next</em></p><p><em>Part 3 of 4 in a series on the MAPS+GV framework</em></p><p>In the previous article, [<a href="https://medium.com/@workinthenow/why-ai-transformation-proposals-stall-and-the-framework-i-use-to-keep-them-moving-5b59a1d924bd">Why AI Transformation Proposals Stall, and the Framework I Use to Keep Them Moving</a>], I described MAPS+GV as four steps carried out in order with domain experts, the frontline staff who know the day-to-day work best, and two ongoing conditions that only executives can secure. This article looks closely at the four steps, Map, Assign, Pilot, and Sustain, and explains for each one the question it answers, how it is evaluated, and what it has to hand off before the next step can begin.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TwVkR9PH6h6LwpulKwW8eA.png" /></figure><h3><strong>Map: W</strong>here we are now and where we want to go?</h3><p>Map builds a current-state overview of the work the company wants to improve and sets the direction it wants to move in. On the current-state side, it checks how far earlier improvement efforts have gone and what resources are available, and surveys how employees already use AI on their own. Knowing which tools they use, and for which tasks, gives Assign a practical starting point when it decides which tools and environments the company will support. On the direction side, Map identifies the problems worth solving and the opportunities worth pursuing, and from them sets the broad goal that later tasks must serve, which Assign then turns into a specific target for each task.</p><p>Within that assessment, each candidate task is plotted on the [<a href="https://medium.com/@workinthenow/before-you-start-ai-transformation-find-out-which-quadrant-your-work-is-in-8be4b6c26d61">readiness quadrant from the first article</a>], which places data accumulation on the horizontal axis and codified judgment criteria on the vertical axis. Where a task sits on the chart shows what preparation it still needs and sets the order in which the work should start.</p><p>The hardest part of Map is judging the data rigorously. Depending on where they stand, a consultant may treat the existing data as more valuable than it is or dismiss it as nearly worthless, and domain experts can lean either way just as easily. These reactions are normal, since each side reads the data through its own situation, but they are exactly what makes it hard to decide where a task sits on the readiness quadrant. A table can look complete long before it is reliable, and gaps discovered only after a plan has been built on the data are costly. The consultant and domain experts therefore have to work together to establish what the company’s situation really is, and building that shared picture is the most important and the hardest work in Map.</p><p>My client’s quality management in late 2019 shows how hard that placement can be. When a quality claim arrived from the appliance maker my client supplied, the domain experts at first could not accept that there was a quality problem at all. The defect rate was already below one part per million, so the claim concerned a handful of parts out of an enormous volume shipped without trouble over several years, and after producing that much for so long, the people who made them found it all the harder to believe that anything was wrong. They had managed quality the way they always had, and the sampling inspections, the only results they kept, showed nothing wrong. Behind that reaction was also an unspoken worry that the claim would force a large change in how they worked, an understandable response from people who had been doing their jobs as they were expected to. Looking at the same process, I saw what those records could not show. The machines inspected every part, but each measurement disappeared the moment the pass or fail verdict was made, so the data that could have answered the claim had never been kept. Put together, the two views led to a clear verdict for Map: the judgment criteria were already in place, but the data was not, which placed the task in Quadrant 2. I explained to the domain experts what sampling inspection can and cannot show, and proposed a way to keep the result of every inspection with as little change to their work as possible. They agreed, and collecting inspection data became the first step of everything that followed.</p><p>The data that began to accumulate in 2020 still had a gap. The first-tier vendor my client supplied did not use parts in first-in, first-out (FIFO) order, so whenever a product failed, the trace from the failed part back to its records kept breaking at the shipping stage before it could reach the day the part was produced. In 2021 we added QR codes to every shipping box to close that gap. When another defect claim arrived from the appliance maker in 2023, the trace went through within hours. It connected the shipping box to the exact production lot and its inspection records, showed that the delivered parts were sound, and established that the failure had come from a different component on the circuit board. As these cases show, the point of Map is to establish clearly where the work stands now. For each candidate task, the consultant and domain experts work out what the task needs in data and in judgment criteria, which of those needs are already met, and which are not, and that comparison is what places the task on the readiness quadrant.</p><p>Map ends by producing the broad goal that later tasks must serve, a readiness chart, a starting order, and a list of candidate tasks, along with the gaps found in data and judgment criteria, a survey of how employees already use AI, and an inventory of the people, equipment, and budget available. All of these carry into Assign.</p><h3><strong>Assign: What will we do, how will we do it, and who will do it?</strong></h3><p>Assign turns the candidate tasks from Map into a concrete execution plan. It decides which data and judgment criteria each task will rely on, divides the roles, and commits the resources the pilot will need.</p><p>I structure this planning around three questions, a framework I developed and call 3Q:</p><blockquote>- <strong>What problem are we solving?</strong> This defines the task and narrows its scope.</blockquote><blockquote>- <strong>What result do we want?</strong> This sets the target outcome and the indicators that will show whether it was reached.</blockquote><blockquote>- <strong>How will we achieve it?</strong> This decides how code, AI, and humans will share the work, where it will run, who holds approval authority, and what resources will be committed.</blockquote><p>Working through the three questions turns a vague sense that something is not working into a written task definition. In practice, the best tasks to start with are often problems that people across the company, from executives to domain experts, have recognized for a long time but set aside for lack of staff and time.</p><p>The most important technical output of Assign is the division of roles among code, AI, and humans. When AI is brought into operations, the line between what it will and will not do has to be built into the structure of the system. In the systems I have built, I first separate the work that can be automated from the decisions that must stay with humans, and then divide the automated work between code and AI according to what each can do reliably:</p><blockquote>- <strong>Automated by code:</strong> calculation and verdicts. Anything that must return the same result every time against a fixed standard, such as calculating the process capability index (Cpk), checking pass/fail specifications, or recalculating an operating schedule, belongs to code.</blockquote><blockquote>- <strong>Automated by AI:</strong> summary, interpretation, explanation, and investigation. AI summarizes code outputs, interprets what they mean, and turns them into plain explanations that non-specialists can read, and at the agent stage it selects analysis tools to trace causes and propose actions.</blockquote><blockquote>- <strong>Human decisions:</strong> judgments that carry responsibility. Not every decision needs a person, and routine ones with little at stake can be left to code and AI. A decision stays with a human when someone has to answer for its consequences, or when it requires grasping a complex situation as a whole, as in approving a schedule change, deciding an order quantity, or restarting a machine.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n_zUpQPccP0auAcM5orSIA.png" /></figure><p>This division of roles among code, AI, and humans is enforced in code rather than written into a guideline. In a system I built to reschedule downstream equipment when one machine on a line stops, code recalculates the operating plan and AI agents explain the options and risks for each machine. The agents work only with the numbers and results the code has produced, and any explanation that brings in a figure of its own is rejected, which closes off the possibility of hallucination. If the agents fail, the plan from the code is still delivered unchanged. Restarting the equipment stays with a human, because only someone who can see the floor can confirm that it is safe to start the machines again.</p><p>Before the pilot begins, Assign also settles two commitments within the organization. The first is ownership. Executives and domain experts agree on who will own the task and how much of that person’s working time will go to it, because a plan without an internal owner rarely survives implementation. The second is the set of KPIs that will be used to evaluate Pilot and Sustain. Domain experts and executives may have different KPIs in mind, and when the two conflict, executives make the call. They finalize the KPIs as a matter of Governance before Assign ends, even though the KPIs are measured later.</p><h3><strong>Pilot: Does it hold up in real work, on a limited scope?</strong></h3><p>Rather than rolling the plan out across the company at once, Pilot tests it under real working conditions within a small scope the team can manage. Domain experts use the system in their actual workflow and bring operational problems to the surface before any wider rollout. Throughout, the KPIs finalized in Assign guide the work: they set the direction from the current way of working (AS-IS) toward the target state (TO-BE), mark how far the pilot has moved along the way, and serve as the gauge against which each improvement is measured.</p><p>Pilot KPIs fall into three groups:</p><blockquote>- <strong>Results:</strong> time and workload saved, reduced burden on staff, consistency of outcomes, and processing cost per case</blockquote><blockquote>- <strong>Actual use:</strong> monthly active users and the number of cases processed</blockquote><blockquote>- <strong>Knowledge captured:</strong> how much tacit know-how was turned into codified judgment criteria during the pilot</blockquote><p>Knowledge captured is the group that sets MAPS+GV apart. The vertical axis of the readiness quadrant measures how much of the judgment behind a task has been written down as explicit criteria. A pilot produces two things. One is the time and effort the tool saves, which the results KPIs measure. The other is less visible, and it comes less from interviews than from use. Because the KPIs require the tool to be used in real work, domain experts keep running into the cases it handles badly, and each time they correct an output, add a condition that was missing, or remove one that does not hold, a piece of their unwritten judgment becomes an explicit criterion. Those criteria stay with the company after the pilot ends, and they move the task up the vertical axis, so the work starts from a higher point in the next round of AI transformation.</p><p>My own pilots produced clear numbers in two of these groups. On results, an AI-assisted quality report cut analysis time per case from 30 minutes to two, and a raw material price analysis tool cut report preparation from 40 minutes to one. On knowledge captured, the way a QC manager investigated anomalies had lived only in that one person’s experience. Building it into a quality investigation AI agent meant writing that know-how down as a seven-step investigation sequence and 13 analysis tools, and from that point the method belonged to the company, not only to the person who had developed it.</p><p>Pilot ends by producing the measured KPI results, the documented improvement over the previous workflow, the judgment criteria written down during use, the operational problems found during testing, and a clear decision on whether to expand. These carry into Sustain.</p><h3><strong>Sustain: Can domain experts run it day to day without relying on AI specialists?</strong></h3><p>In Pilot, the new way of working was tested on a limited scope, usually with the consultant or an AI specialist close at hand. Sustain has two aims. The first is to weave AI into the company’s work processes until the two are chemically bonded rather than merely mixed, so that taking the AI out would mean redesigning the work itself. The second is to make the system self-sufficient. Domain experts should be able to run it without the consultant or AI and IT specialists, and that means more than keeping it running, since they also need to be able to add features and improve existing ones as the work changes.</p><p>Sustain KPIs fall into three groups:</p><blockquote>- <strong>Results:</strong> growth in the number of cases handled and the savings realized as adoption widens</blockquote><blockquote>- <strong>Integration:</strong> the share of the target work that now runs through the system, the exception rate, and the rework rate</blockquote><blockquote>- <strong>Self-sufficiency:</strong> how often the consultant or an AI or IT specialist has to step in, how many staff use the system on their own, how many features domain experts add or improve themselves, and how widely those features are shared and reused across the company</blockquote><p>The exception rate and rework rate also do a second job. They show where judgment criteria are still incomplete, which gives the team concrete evidence when it chooses the next candidate tasks.</p><p>My client’s quality system shows what the first aim looks like in practice. In 2024 we added AI reports that interpreted the statistical indicators automatically, and the company changed its workflow rule so that an AI-generated report, once reviewed by staff, could be filed as the official quality control report without being rewritten, with the reviewer taking responsibility for its content. That rule change is what bonded the tool to the work. From then on, the AI report was the company’s official quality record, and taking it out would have meant rewriting the workflow.</p><p>Training belongs to Sustain as well, because AI transformation has to keep going after the project ends. Executives need to understand Governance and Viability well enough to keep setting the rules and committing the resources the work depends on. Domain experts need to learn how to put their own experience and judgment criteria into the system through AI, adding features and improving existing ones, which also means the system has to be built to let them do so. Sustain closes by producing a rollout plan, an operations manual, training materials for each role, and a list of next candidate tasks.</p><p>That last output is where the next round begins. The candidate tasks go back into Map, and the second round starts from a better position than the first, because the records gathered during the pilot and daily operation have filled in the data axis, and the criteria domain experts wrote down while using the system have raised the other. The same area of work therefore sits higher on the readiness quadrant when Map looks at it again. MAPS is meant to be run this way, again at regular intervals, so that AI transformation becomes a routine the company repeats rather than a one-time project.</p><h3><strong>KPIs are rules, not scorecards</strong></h3><p>Treating KPIs as a scorecard for grading a project after it ends leads to poor design, because the moment a metric is set, it tells the people being measured how to behave. The pilot already showed this at work: because its KPIs required the tool to be used in real work, domain experts kept correcting it, and their judgment was written down along the way. A KPI therefore has to be designed by asking what employees will do in order to raise that number.</p><p>If you measure how many automation programs a team builds, people will build many simple tools that are easy to finish and rarely used. If you measure monthly active users and completed runs instead, the team will first look for work that domain experts actually need help with.</p><p>That is why KPIs are measured in Pilot and Sustain but finalized before Assign ends, and why setting them belongs to Governance, the first of the two conditions that only executives can secure and the subject of the final article in this series.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=85b2ad604929" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why AI Transformation Proposals Stall, and the Framework I Use to Keep Them Moving]]></title>
            <link>https://medium.com/@workinthenow/why-ai-transformation-proposals-stall-and-the-framework-i-use-to-keep-them-moving-5b59a1d924bd?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b59a1d924bd</guid>
            <category><![CDATA[ai-transformation]]></category>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[smart-factory]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 11:01:02 GMT</pubDate>
            <atom:updated>2026-10-31T18:44:19.045Z</atom:updated>
            <content:encoded><![CDATA[<p><em>MAPS+GV: four steps taken in order, and two conditions only executives can secure</em></p><p><em>Part 2 of 4 in a series on the MAPS+GV framework</em></p><p>I have received consulting reports, commissioned them as a business owner, and implemented recommendations as a consultant. Across those roles, I watched some projects take root while others quietly stalled. MAPS+GV is my attempt to explain what made the difference.</p><p><strong>Three things I learned on the receiving end</strong></p><p>The first lesson is that a plan must fit the company that will carry it out. When a project stalls, it is easy to blame the technology, a lack of technical skills, or the chosen direction. But in my experience, the problem often lies in a plan that demands more than the company’s available resources can support.</p><p>The second lesson is that executives often already have a rough idea of the problem and possible solutions. When they had fewer sources of information, consultants could introduce an entirely new direction. Today, executives have many ways to explore how to improve their businesses, and they bring that knowledge to the discussion. The consultant’s job is to test and refine their thinking, then turn it into a concrete execution plan. Executives must understand that plan well enough to explain it in their own words and give their teams the direction and support to carry it out.</p><p>Third, an improvement project is more likely to succeed when the frontline staff who know the day-to-day work best help shape it from the start. They know where data is missing, which exceptions recur, and how much extra work they can absorb. Involving them from the diagnosis onward helps the team check proposed changes against those realities and agree on how the work will be done. When implementation begins, the people responsible have already helped shape the tasks and understand why they were designed that way. They are then better prepared to handle problems and keep the work going after the consultant has left.</p><p><strong>Four questions and two conditions</strong></p><p>If you turn those lessons into questions and ask them in the order a project unfolds, you get four:</p><p>- <strong>Map:</strong> Where are we now?</p><p>- <strong>Assign:</strong> What will we do, how will we do it, and who will do it?</p><p>- <strong>Pilot:</strong> Does it hold up in real work, on a limited scope?</p><p>- <strong>Sustain:</strong> Can frontline staff run it day to day without relying on AI specialists?</p><p>These four are the steps. Alongside them sit two conditions that belong to no single step:</p><p>- <strong>Governance:</strong> Who is responsible for what, and under which rules? Executives set the principles for using AI, the roles, who approves, and how errors are handled, and agree on goals, performance evaluation, and rewards so that frontline staff have a reason to take part.</p><p>- <strong>Viability:</strong> Can the company provide the resources the project needs to survive and develop? Like the company itself, an AI transformation project needs a continuing supply of resources throughout its life. Executives must commit time for frontline staff and AI specialists, provide funding, and adjust that support as needs change. These commitments support every MAPS step, from the initial diagnosis to sustained daily operations.</p><p>The project team works through the four MAPS steps in sequence, together with frontline staff. Throughout that process, executives must keep responsibilities and decision rules clear and ensure that the necessary resources remain available. Governance and Viability therefore run across all four steps: they are ongoing conditions that allow the team to carry out the work and keep it going.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*beMj3CQiq8Kf6_bj5l3drA.png" /><figcaption><em>MAPS+GV: four sequential steps (Map, Assign, Pilot, Sustain) with Governance and Viability running beneath all four</em></figcaption></figure><h4><strong>From Map to Sustain: assessment methods and outputs</strong></h4><p>For each MAPS step, I explain what to assess, which tools or measures to use, and what the team should produce.</p><p><strong>Map</strong> starts by identifying the problems the company wants to solve and the opportunities it wants to pursue. It builds a current-state overview by examining how AI is already being used, how far existing transformation efforts have progressed, and what resources are available. Within this assessment, the readiness quadrant from the previous article helps evaluate individual tasks along two axes: data accumulation and codified judgment criteria. Based on these findings, a starting order identifies which tasks to address first and what preparation they need. A list of candidate tasks carries specific improvement opportunities into Assign for further examination and definition.</p><p><strong>Assign</strong> turns the candidate tasks from Map into an execution plan by deciding how to allocate roles, rules, and resources. I structure this planning around three questions (3Q). “What problem are we solving?” defines the task and its scope. “What result do we want?” establishes the target and the indicators that will show whether it was reached. “How will we achieve it?” determines how people, code, and AI will share the work, where it will be done, who can make which decisions, and which rules apply. It also specifies how many people are needed, what other resources to commit, and for how long. Executives and frontline staff agree on these decisions, forming a concrete plan for the pilot.</p><p><strong>Pilot</strong> is judged with KPIs that cover results, actual use, and knowledge captured. It produces measured results, the improvement over the current way of working, the operational problems it exposed, and a decision on whether to expand.</p><p><strong>Sustain</strong> is judged with KPIs that cover results and how firmly the new way of working has settled in: how often it is actually used, how often someone has to step in by hand, and whether operators can maintain it without the expert. It produces a rollout plan, an operations manual, training materials, and a list of next candidate tasks.</p><p>That last output matters more than it appears to. The next candidates go back into Map. Because the pilot and daily operation add records and sharpen criteria, the same work sits higher on both axes when it returns to Map, which is what makes MAPS a cycle rather than a single pass.</p><p><strong>Two cases, one difference</strong></p><p>Two of my own consulting cases show why the framework takes this shape.</p><p>The first is the supplier I have advised since 2019. When a large home appliance maker that had just become number one in its market demanded a factory audit and strong countermeasures, the supplier’s defect rate was already below one part per million, yet the customer was also considering a switch to Chinese sourcing. Its inspection criteria were already written down, but the results were not being kept, so it could neither prove its quality nor trace a cause. Diagnosis therefore placed it in Quadrant 2, and we started by collecting inspection data. Over the following years the system traced parts down to individual boxes, answered a defect claim within hours instead of days, and allowed an auditor to finish an audit report on the same day as the visit. In 2024 we added AI reports that cut a 30-minute analysis to two minutes, the second of the three stages of AI use described in the previous article, and in 2026 an AI agent that investigates process changes which pass/fail verdicts would miss, the third stage, in which the agent leads the investigation and a person makes the final call. Early in 2026 the appliance maker asked comparable suppliers when they would reach the same level of quality management, and my client, already there, had nothing further to do. The system has now been in operation for seven years.</p><p>The second is a solar inverter manufacturer listed on KOSDAQ, which I advised in 2021. It had just opened a second plant and needed to coordinate production schedules and delivery dates between its headquarters and the new site. The problem was clear and a solution was ready, but nobody at the company was responsible for carrying it out, and at the time I was running my own factory and could not provide the training and hands-on support myself. It never reached implementation.</p><p>What separated these two cases was the set of conditions for execution rather than the quality of the solutions. At the inverter manufacturer, no one at the company owned the work, which is a gap in Governance, and no one, including me, had time set aside for training and hands-on support, which is a gap in Viability. That case is why Assign now settles with executives which frontline staff will be committed and for how much of their time, and why both conditions run across every step instead of appearing once as an approval or a budget line.</p><p>The next article walks through the four steps in detail, and the one after that covers Governance and Viability.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b59a1d924bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Before You Start AI Transformation, Find Out Which Quadrant Your Work Is In]]></title>
            <link>https://medium.com/@workinthenow/before-you-start-ai-transformation-find-out-which-quadrant-your-work-is-in-8be4b6c26d61?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/8be4b6c26d61</guid>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 11:46:01 GMT</pubDate>
            <atom:updated>2026-10-01T11:46:01.549Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Two axes, data and documented judgment, decide where AI transformation starts, what it costs, and what to do first</em></p><p>Most AI transformation projects I have seen, including some of my own, go through the same three conversations. At the start someone says, “We don’t have the data we need, and what we do have is scattered across different systems,” so the company spends money to integrate it. A few months later the complaint changes to “We can access all our data now, but we are not sure what to use it for.” The third conversation is where the real problem finally comes out, and it usually sounds like this: “The know-how is in people’s heads, and almost none of it is written down.”</p><p>These projects go wrong when a company starts before it knows where it stands. By the time the third conversation happens, the integration budget is gone, and the team has to go back and start over. The quadrant I use is a way to find that starting point. It places a piece of work on two axes, so you can see where you are today and which direction to move first. The quadrant comes out of eleven years of running a components factory and advising a client on quality since 2019.</p><h3><strong>Two ingredients AI needs to work in your company</strong></h3><p>For AI to take over or support a piece of work that people do today, it needs two ingredients. The first is data, the numbers and records a person looks at before deciding. The second is the criteria for judgment, the rules that tell that person what those numbers mean and what to do next. The quadrant places these two ingredients on two axes, and on both axes only what AI can reach counts, so a paper binder or a file on one person’s PC scores low.</p><p><strong>The horizontal axis is data accumulation.</strong></p><p>Data here means anything a person consults before making a judgment: measurements, records, documents, emails, past cases, and the revision history behind those documents. The axis measures whether the organization knows which of this data actually drives the outcome of the work and is collecting it, not how much data it has. In this series I use quality control as the running example, because it is the work I know best. A plant can record thousands of measurements a day and still score low on this axis. If the one that decides yield is not among them, all it has is a stack of numbers.</p><p><strong>The vertical axis is the codification of judgment criteria.</strong></p><p>It measures how much of the tacit knowledge in employees’ heads has been sorted and written down as documents, rules, or code. The easiest part to write down is a pass/fail threshold. The hardest is the order in which an experienced engineer checks things when a process starts to go off track. The more of both that is documented, the higher the score.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FKO99W7IQcKvsQb9U4LjpQ.png" /></figure><h3><strong>The four quadrants</strong></h3><p>I number the quadrants the way a coordinate plane does, starting at the top right and moving counterclockwise. Each axis is split at the point where it is enough for AI to take on the work in question.</p><ul><li><strong>Quadrant 3: little data and little codified criteria.</strong> Records are kept by hand and the rules live mostly in people’s heads, which is common in small organizations. Both axes have to be built at the same time, which calls for equipment investment and employee time together, so this is the slowest place to start from.</li><li><strong>Quadrant 2: codified criteria, but little data. </strong>Companies that have been through certification audits, or that have written down their internal rules, handover notes, or training materials, often sit here, knowing how to judge a result but keeping few of the values behind it. The client I began advising in 2019 was in exactly this position. It had a clear pass/fail standard but collected no data. Its inspection machines judged each part, and the measured values were lost on the spot. Our first job, finished in 2020, was to add data collection to that equipment. Over the next few years the measurements piled up, and by 2024 the client’s inspection work had both axes in place.</li><li><strong>Quadrant 4: data, but little codified criteria. </strong>Most organizations that already have data collection systems are here. The sensors and databases exist, and basic criteria such as pass/fail limits are usually in place. The gap is the deeper judgment the work depends on, such as what an unusual number means, what to check first, and what to do next. Some of that judgment is in paper manuals, in files on one person’s computer, or in people’s heads, all out of AI’s reach. This is the quadrant where people say, “We have gathered all our data, and we still don’t know what to use it for.” The first two tasks are to turn that judgment into explicit criteria that AI can find and use, and to choose the useful data from what has piled up and connect it to what it means for the business.</li><li><strong>Quadrant 1: data and codified criteria.</strong>Code and AI can now reproduce the judgments people used to make. From here a company can validate an application with a pilot and then move toward more advanced AI transformation, including AI agents that carry out work on their own initiative. That is where my client’s quality work stood in 2024, and by 2026 it had advanced to where an AI agent decides for itself how far an investigation should reach.</li></ul><h3><strong>Why Quadrant 4 comes before Quadrant 2</strong></h3><p>The two axes are built with different resources. Data takes investment in equipment and time for it to accumulate, while codified criteria take employees’ time. That difference decides which axis is cheaper to fill, and from that comes the order in which to start.</p><p>Work in Quadrant 4 is the cheapest to move, because the data collection equipment is already in place and what is missing is employee time, which can be arranged within the existing organization. Work in Quadrant 2 comes second, because the data collection equipment is missing, which means an investment decision first and, even after that decision, time for the data to accumulate. Quadrant 3 comes last because both axes must be built, and work already in Quadrant 1 can go straight to a pilot.</p><h3><strong>Certification is a ready-made checklist</strong></h3><p>Collecting data and organizing tacit knowledge is expensive, so few companies take it on as a standalone project. In practice they start when a regulator, a large customer, or a government program requires a certification such as ISO 9001, IATF 16949, or HACCP, or a national quality standard. Doing the groundwork alongside that preparation lowers the cost. A system rollout or a wave of retirements can force the same work.</p><p>You can borrow the methodology even if you never apply for the certificate. These schemes define what must be documented, to what level, and which records must be kept, so using their items as a checklist makes time and cost far easier to predict than building the list from scratch.</p><p>There is a limit to this, however. Certification asks only for the data needed to separate pass from fail, so even certified companies rarely use those values for anything beyond the verdict. That is why my client sat in Quadrant 2 despite having solid standards.</p><h3><strong>Quadrant 1 is where it begins</strong></h3><p>On the chart Quadrant 1 looks like the destination, but I treat it as the starting line, the point where AI can begin to take on the work. Once AI becomes part of daily operations, both axes keep rising: the records of AI-assisted work become new data, and every correction a person makes to an AI output becomes a sharper criterion. As the axes rise, the range of situations AI can handle widens from routine cases toward unusual ones.</p><p>You also don’t need to wait for Quadrant 1 before beginning. The quadrant shows where you stand, and that position is your starting point. From there I classify AI use into three stages by how much of the work AI takes on. They run from automating routine tasks, to assisting a person’s judgment, to leading the work as an agent.</p><ul><li><strong>Stage 1, automating routine tasks.</strong> AI works with the data you already have. A typical use is employees asking a messenger-based assistant, in plain language, to collect information or check email. This stage needs no new data or criteria, so it can begin from any quadrant.</li><li><strong>Stage 2, assisting a person’s judgment.</strong> AI works on accumulated data together with even partial criteria. When an employee uploads a photo of a receipt, for instance, AI files the expense claim and assigns the account code by following the company’s accounting rules. In quality control, code calculates indicators such as the process capability index (Cpk) and makes the verdicts, while AI interprets the results and drafts a report.</li><li><strong>Stage 3, leading the work as an agent.</strong> Both axes are in place, and AI agents decide what to check and which tools to use, while a person makes the final decision. An agent can audit company spending on its own, looking for unusual changes or irregular payments and tracing each one to its cause. In quality control, an agent can find process shifts that pass/fail verdicts miss and work out why they happened.</li></ul><p>Each stage leaves behind the records and criteria the next one needs.</p><h3><strong>How to use the quadrant</strong></h3><p>Two rules keep the quadrant useful. The first is to start from the smallest unit, the individual task, and work upward. In the same company, the inspection task on the quality team may sit near Quadrant 1 on years of machine data, while the supplier selection task on the purchasing team sits in Quadrant 3 on the instinct of the person in charge. Placing the company as one point would average them into a position that matches neither. Place each task first, then roll the results up to teams, departments, and finally the company.</p><p>The second is to use the quadrant to reach agreement with executives, not to walk them through every detail. Executives may not know how AI works or how to apply it, but years of running the business have shown them exactly where it falls short. So rather than explaining the direction of each task, show them the big picture and how it addresses the gaps they already feel. The task-by-task map stays with the working team, and leadership agrees on the coarse picture and a next step clear enough to decide on at a glance.</p><p>Once the position is agreed, detailed execution moves to a four-step framework I call MAPS+GV, which is the subject of the next article.</p><p>#AI #Transformation #AI_Transformation #Manufacturing #Digital _Transformation #Knowledge_Management #Leadership</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8be4b6c26d61" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Did Not Know How to Build It. AI Told Me How.]]></title>
            <link>https://medium.com/@workinthenow/i-did-not-know-how-to-build-it-ai-told-me-how-cd2b25352a83?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd2b25352a83</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Mon, 10 Aug 2026 11:21:01 GMT</pubDate>
            <atom:updated>2026-08-10T11:21:01.509Z</atom:updated>
            <content:encoded><![CDATA[<p>Being ready was a daily job. It is now a one-minute job, on the day it matters.</p><p>In the last article, I described a purchasing employee who explained a routine order in three sentences. Walking through an actual order with him turned those three sentences into twelve decisions. We did not try to solve all twelve. One carried more weight than the rest, and this is the first half of the tool we built from it.</p><p>Our factory buys raw materials once a month. The order goes out in the first week, the materials arrive three months later, and the price locks in on the day we order. The quantity, though, depends on where prices appear to be heading. Order more before a climb, lean on inventory before a dip, and the difference lands directly on our operating margin.</p><p>That judgment needs a view of direction over time, which is why a morning ritual existed. Someone on the purchasing team checked WTI crude oil, aluminum, and copper on a financial portal and copied the numbers into a spreadsheet. Before each monthly order meeting, that spreadsheet was rebuilt into something presentable: recent prices lined up against past orders, with a line or two at the bottom about where things seemed to be heading.</p><p>The monthly meeting was not the only reason for the daily habit. A producer announces a cut, or a shipping route gets disrupted, and it makes the morning news. I ask the purchasing team how our prices are looking. They already know, in general terms. What they cannot do is put a number on it in the moment, so what I get is that prices are up a little and the actual figures will follow. Keeping the sheet current every morning was how they closed the distance between those two answers.</p><h4><strong>Why the portal was never enough</strong></h4><p>Yahoo Finance and Investing.com do this well. Investing.com alone offers more technical indicators than our team will ever use. But everything there is built to be looked at on their screen, and looking at a screen is not the same as having the data. The moment a number has to go into our own report, someone types it in by hand.</p><p>Which is why the first build is not a chart. It is a price history in a file, which is the form that anything else can be built from. The next article builds the first thing.</p><h4><strong>Three questions on paper, and one I could not answer</strong></h4><p>The build started the way every build in this series starts: three questions on paper, a method I call [3Q](https://medium.com/@workinthenow/my-employees-couldnt-describe-their-own-jobs-063c4b9b6d1b).</p><p><strong>What is the problem?</strong> The team orders monthly under a three-month lead time, and the order quantity depends on price direction. Price snapshots exist in spreadsheets and notes, but keeping them current and analysis-ready is daily manual work, done every morning against the chance that someone asks that day.</p><p><strong>What result do you want?</strong> Our own database holding at least two years of daily price history for WTI crude oil, aluminum, and copper, stored as one reusable file, brought fully current by a single run whenever anyone needs it, and ready to feed charts and analysis later.</p><p><strong>What approach will you use?</strong> I did not know. That third answer is the reason I wanted to write this build up. Every earlier article in this series had an approach I could state before I started: read measurement data from a database, calculate the indicators the quality team already trusts, display them as charts. Not this time.</p><p>The 3Q worksheet does not ask you to. It has a slot for not knowing.</p><p>The first two answers contain no technology, and they were the two I was qualified to write.</p><p>The 3Q worksheet does not ask you to. It has a slot for not knowing.</p><p>The first two answers contain no technology, and they were the two I was qualified to write. The third one I could not even research.</p><p>A search box needs a word to start from. Search engines handle vague questions better than they did a few years ago, and someone practiced at it would have found a way in. I was not going to. I knew my situation and I knew what I wanted out of it. What I did not have was the word.</p><p>AI does not need one.</p><p>So I pasted the first two answers into AI and asked what they implied.</p><blockquote><strong>Here is what I need and why. I am not an IT professional and I work in Google Colab. What is the simplest practical way to do this?</strong></blockquote><p>The reply split the job in two, which I had not thought to do. Collecting is one problem and storing is another, and they have different answers. For collecting, yfinance: an open-source package that pulls prices from Yahoo Finance with no account and no API key. For storing, SQLite: the same single-file database behind the quality dashboard in the first article. It sits in Google Drive like a spreadsheet. Nothing to install, nothing to set up, and nobody in IT has to be involved.</p><p>I had heard of neither. A few follow-up questions were enough to know they were the right answers, which is a different thing from knowing how to use them. And notice what I asked for. Not a solution, not code — just how this kind of work is normally done. The input was the two answers I had already written myself.</p><p><strong>The whole prompt, and the two parts that mattered</strong></p><p>With the third answer filled in, the worksheet was complete, and a completed worksheet is most of a prompt. What I sent to AI used TCREI, the five-part structure behind every build in this series. Here is all of it, because the previous articles have named the five parts without ever showing them together.</p><p><strong>Task — what the notebook does.</strong> Collect daily prices for WTI crude oil, aluminum, and copper using yfinance, and store them in a SQLite file on Google Drive, each commodity kept separately. The first run downloads up to two years of history. Every later run adds only what is new since the last update, with no duplicates and no gaps. Show progress while it runs, so the person watching can see what is being processed and whether it finished.</p><p><strong>Context — who it is for.</strong> The purchasing team needs a price history that is ready whenever it is needed, not only before the monthly meeting. Nobody on that team is a developer, so the Colab notebook has to run with one click after the initial setup. The file lives on Google Drive rather than inside the session, because a Colab session takes everything with it when it closes.</p><p><strong>References — what AI cannot guess.</strong> Two years of history. The exact folder on Google Drive, the same one the quality dashboard already uses. The file name, commodity_prices.db. The tools by name, yfinance and SQLite. And the fields each daily record has to keep: date, open, high, low, close, and volume where available.</p><p><strong>Evaluate — how the notebook proves itself.</strong> On the first run, confirm the database was created and the records were stored, and print a short summary. From the second run onward, compare the most recent stored price for each commodity against the current value on Yahoo Finance. Matching values mean the data can be trusted. A mismatch prints a warning with enough detail to describe the problem to AI.</p><p><strong>Interact — what happens after it runs.</strong> Run it, read the output, and bring whatever went wrong back to AI in plain language.</p><p>Two of those five did the real work here.</p><p><strong>References is where the domain knowledge went.</strong> It is short here, shorter than in any other build in this series, and that is worth noticing. A quality dashboard needs to be told which measurements are equipment noise and what the specification limits are, because those facts exist only inside our factory. A commodity collector needs almost none of that. The information is public and the same for everyone, so what References had to supply was mostly location: which folder, which file, how far back.</p><p><strong>Evaluate is the section people skip.</strong> Read it again and notice who does the checking. Not me. I cannot open a database file and confirm that a column holds what it should, and neither can anyone on the purchasing team. So the verification had to be built into the notebook and reported in a sentence a person can read.</p><p>That is a different instinct from writing a checklist for yourself. A checklist assumes you can perform the check, and here nobody can.</p><p>You do not need to understand how something works to tell whether it is correct. Nobody knows how their electricity is metered, and everybody can read the meter. What makes the reading useful is having something to compare it against — last month’s number, or the month you actually had.</p><p>The cross-check against Yahoo Finance is that line. It runs on every update, forever, without anyone remembering to do it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p_Ypjz4Pz1_kgDtjyxkf8Q.png" /></figure><p><strong>Interact, before and after</strong></p><p>Interact had already done half its work before anything was built. The conversation that produced yfinance and SQLite was the Interact step, running before there was anything to fix.</p><p>Interact is usually described as what you do after the first draft comes back. It works just as well before, when the problem is clear and the path is not.</p><p>The other half never came up. The notebook ran on the first try. That surprised me less than it would have a year ago, because there was not much left to get wrong — the tools were already chosen, the folder and the file name were specified, the update rule was written out, and the checks were defined before a line of code existed. A prompt that specific does not leave many places for the first draft to go astray.</p><p>This is the part I would push back on if I were reading someone else’s write-up, so let me be precise. Nothing here was hard. Collecting public price data into a file is a small problem, and a small problem with a complete brief is exactly what current AI handles well. The reason to write the brief carefully is not that it guarantees a clean first run. It is that when something does break, you have a written record of what you asked for, and that is what makes the next conversation short.</p><p>A few days later I ran the notebook again, which was the real test. It reported the missing days for each commodity and nothing else. A few days later I ran the notebook again, which was the real test. It reported the missing days for each commodity and nothing else.</p><p><strong>What changed in purchasing</strong></p><p>The morning update is gone, and so is the reason for it. Nobody keeps the sheet warm now. When the question comes, whoever is asked runs the notebook, waits about a minute, and answers from data that is current as of that minute. Being ready stopped being something a person has to maintain.</p><p>That is the honest scope of this build. The data is collected, verified, and current on demand, in a file the team controls. Turning that history into charts and a read on price direction is a separate build, and it is next.</p><p>What this stage bought is the foundation. The price history now accumulates in a file we own, which is what makes the next build possible at all. Collecting is not analyzing. The file answers what the price was, not what to do about it. That is the next build.</p><p><strong>Build your own in a minute</strong></p><p>Commodity prices are public data, so nothing in this build touches company information. The notebook runs in Google Colab, saves to your own Google Drive, and produces your own file. Nothing to install.</p><p><a href="https://github.com/Munjae-Song/medium_Build_Business_Tools_with_AI/tree/main/002_collect_daily_prices">The notebook and the full TCREI prompt →</a></p><p>Download the notebook, open it in Colab, and run it. You will have two years of daily prices for three commodities in a database you control. Change the ticker list at the top and it collects whatever you watch instead. Open the result with DB Browser for SQLite if you want to see the tables without writing a line of Python.</p><p>One note on why you have to run it yourself. yfinance reads Yahoo Finance’s public endpoints, and its authors are explicit that it is meant for personal use. So the repository has the notebook and the prompt, and no database file. Collect your own.</p><p>If it ever becomes something your company depends on, that is when you pay for a data feed.</p><p>The full TCREI prompt is in the same repository. That matters more than the code. The code is what AI produced from the prompt; the prompt is what I actually wrote.</p><p><strong>The part that transfers</strong></p><p>Swap WTI, aluminum, and copper for the exchange rates, freight indexes, or component prices your business watches, and the same prompt builds your collector.</p><p>The part worth copying is not the collector. It is that I started without knowing how. The first two questions were mine to answer. The third one I handed over.</p><p><a href="https://github.com/Munjae-Song/medium_Build_Business_Tools_with_AI/tree/main/002_collect_daily_prices">The notebook and the full TCREI prompt: [GitHub]</a></p><p><a href="https://medium.com/@workinthenow/i-run-a-factory-ai-built-my-quality-dashboard-0ad2b752d099">The first build in this series: [I Run a Factory. AI Built My Quality Dashboard.]</a></p><p><a href="https://medium.com/@workinthenow/my-employees-couldnt-describe-their-own-jobs-063c4b9b6d1b">Where the three questions came from: [My Employees Couldn’t Describe Their Own Jobs]</a></p><p><a href="https://munjaesong.substack.com/">I publish one verified build at a time: [newsletter link]</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cd2b25352a83" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My Employees Couldn’t Describe Their Own Jobs]]></title>
            <link>https://medium.com/@workinthenow/my-employees-couldnt-describe-their-own-jobs-063c4b9b6d1b?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/063c4b9b6d1b</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[problem-solving]]></category>
            <category><![CDATA[productivity-tools]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 11:01:01 GMT</pubDate>
            <atom:updated>2026-07-30T11:01:01.624Z</atom:updated>
            <content:encoded><![CDATA[<p>Coding is getting easier. Describing the work is not.</p><p>Give me an hour beside a production line, and I can usually spot unnecessary steps, repeated work, and avoidable delays. A worker reaches for the same part hundreds of times. A machine sits idle because the next batch of material has not arrived on time. Inspectors keep choosing samples that are easy to collect instead of selecting them randomly, making some defects difficult to detect.</p><p>On a factory floor, inefficient work is often visible. You can follow the movement, measure the delay, and ask the people doing the work where they see problems in the process.</p><p>I spent more than a decade working this way: observe the process, identify the friction, and find a more efficient way to handle the work. When automation could make that improved process faster, more consistent, or easier to manage, I automated it. This approach guided years of smart factory projects at my company and eventually became the basis of a Korean-language book I wrote about practical factory automation.</p><p>Then I tried to apply the same approach to office work.</p><p><strong>The approach stopped working in the office</strong></p><p>I began with accounting, purchasing, and quality documentation. The logic seemed straightforward. If observation could reveal waste on a production line, it should also reveal repetitive work in the office.</p><p>So I studied how office employees worked, where delays occurred, and which tasks they had to repeat.</p><p>Mostly, I saw people looking at screens.</p><p>They moved between Excel, the ERP system, email, and internal documents to collect data, perform calculations, and copy results from one place to another. The activity was visible, but the reasoning behind each step was not, which made the real source of inefficiency difficult to identify.</p><p>The most important steps were often not on the screen at all. The work often depended on unwritten standards, such as how to review a report before submission or how to reflect unusual circumstances in a monthly calculation.</p><p>Factory work often leaves a physical trail. Office work leaves a trail of clicks, files, and decisions, but the logic connecting them may exist only in the employee’s head. Improving or automating that work therefore begins by drawing out that hidden logic and turning it into explicit steps, rules, and decision criteria.</p><p><strong>Simply asking people was not enough</strong></p><p>I changed my approach. Instead of observing the work, I asked employees to explain it.</p><p>I asked what they did repeatedly, what took too long, and where errors were most likely to occur. Since they performed these tasks every day, I assumed they would know which parts of the work needed improvement and which could be automated.</p><p>It was harder than I expected to identify specific improvements from their answers.</p><p>People could explain individual tasks, but they often left out steps that seemed too obvious to mention. Because I did not perform the work myself, those missing details made it difficult to follow the process from beginning to end.</p><p>One example stayed with me. I asked a purchasing employee to describe how a routine order gets placed. His answer took three sentences: check the stock level, compare the quotes on file, issue the purchase order. When we walked through an actual order together, those three sentences turned into twelve decisions. Which stock figure to trust when the system count and the shelf count disagree. Whether a quote from four months ago still holds, which depends less on its age than on how much the commodity price has moved since. Whether to increase the quantity, reduce it, or hold it, knowing that material ordered today will not arrive for three months. None of it was written anywhere. He had made those calls hundreds of times without ever treating them as decisions</p><p>This was not because they lacked experience. The work had become so familiar that much of its logic was difficult to recognize and explain. When asked to describe every detail, people often found it easier to summarize the problem with a broad solution: “We need a new system,” or “The ERP should handle this.”</p><p>The employees’ expertise was exactly what made the work difficult to explain. Years of repetition had turned many steps, checks, and judgments into habit, so they no longer felt like separate pieces of knowledge that needed to be stated. This is how tacit knowledge works: people can apply it consistently without being able to describe the detailed steps and decisions that someone else would need to follow. This creates few problems while an experienced employee continues performing the same task. It becomes a serious issue when the company tries to improve or automate the process, train a new employee, or preserve the knowledge for future use.</p><p><strong>AI makes the missing explanation more important</strong></p><p>AI-assisted development has reduced the time and technical knowledge required to turn a clear description into a first working tool. A domain expert can now ask AI to organize data, calculate familiar indicators, generate a report, or build a small interface without first becoming a professional programmer. They no longer need to learn in advance which libraries or packages to use, how to install and connect them, or how to set up the environment and run the program. AI can make those technical choices, prepare the necessary instructions, and help resolve errors as they appear.</p><p>AI is increasingly capable of deciding how to turn a clearly defined task into working software. It can organize data, apply calculations, build interfaces, and generate reports, but it cannot determine what the work should accomplish, which process is appropriate, or which results are acceptable. AI does not know which readings are invalid, which exceptions require review, why one calculation is trusted, or what information a manager needs before making a decision. Those judgments must come from the people who understand the work. This is why many AI projects stall before technology becomes the main challenge: the organization may have capable AI and experienced IT staff, yet still lack a clear description of the process it wants to improve.</p><p>My employees had the practical knowledge and experience needed to build useful business tools. However, they were not accustomed to describing that knowledge in enough detail for another person or AI to follow. Many of the rules, exceptions, and judgments behind the work remained in their heads.</p><p><strong>Three questions before you ask AI anything</strong></p><p>Palantir’s forward-deployed engineering approach is built around working closely with users, breaking down ambiguous problems, and understanding the relevant data before building a solution. I found the underlying problem-solving discipline useful, but I needed a simpler version that a non-engineer could use without a technical team in the room.</p><p>I reduced that discipline to three questions, a method I call 3Q: Problem, Result, and Approach.</p><p>The clearest example I have is not from accounting but from quality control — the same invisible logic, in a process I could verify from beginning to end. To show how the three questions work together, I will revisit the quality dashboard from my earlier article, <a href="https://medium.com/@workinthenow/i-run-a-factory-ai-built-my-quality-dashboard-0ad2b752d099">“I Run a Factory. AI Built My Quality Dashboard,” </a>and use it as a running example.</p><p><strong>1. What is the problem?</strong></p><p>Before each shift meeting, the quality engineer must collect measurement data from CSV files, organize it, and calculate the necessary quality indicators. This preparation takes about thirty minutes, delaying the information the line leader needs to understand the current process condition and decide whether intervention is necessary.</p><p>At the 3Q stage, the problem should be stated in direct, simple language. The purpose is to make the current work and its difficulty easy to understand. Detailed rules and requirements will be developed later through TCREI.</p><p><strong>A useful problem statement identifies who performs the work, when it happens, what causes the difficulty, and why it matters.</strong></p><p><strong>2. What result do you want?</strong></p><p>People often confuse the desired outcome with the process or tool they expect to use. “We need a dashboard” proposes a tool without explaining what the work should achieve. Once the desired outcome is clear, AI can help identify and build an appropriate approach or tool.</p><p>In this example, the desired outcome is: “The line leader needs to see whether the process is drifting before the next shift meeting.” This statement gives both a human colleague and AI a clear target. It identifies who needs the information, what they need to know, and when the information must be available.</p><p><strong>A useful result statement identifies who needs the information, what they need to know, when they need it, and what decision or action it should support.</strong></p><p><strong>3. What approach will you use?</strong></p><p>“Approach” explains how the desired outcome should be produced in terms of the work itself. In this example, the line leader needs to recognize process drift quickly. Plotting the measurements over time and marking the expected range on a graph would allow the line leader to recognize the drift at a glance, without having to interpret rows of numbers. The graph turns a large volume of measurements into a visual pattern that supports a timely decision.</p><p>AI can generate many types of charts and interfaces. The domain expert must explain which presentation makes the situation easiest to recognize, which changes matter, and what decision the information should support. The detailed approach may also include which data should be used, which readings are valid, which calculations the team trusts, and how the result should be verified.</p><p><strong>A useful approach statement explains how the desired outcome should be produced using the data, rules, calculations, and presentation that reflect how the work is actually performed.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3_SlZhkV_B6mlTEsj4rRmg.png" /></figure><p><strong>Why the prompt is not the first step</strong></p><p>James Phoenix and Mike Taylor describe five principles in <em>Prompt Engineering for Generative AI</em>: give direction, specify format, provide examples, evaluate quality, and divide labor. These principles help people communicate more effectively with AI.</p><p>Every one of them assumes you already know the work. Before writing those instructions, the user must clarify the problem, define the desired result, and explain the business approach that AI should follow. This is the role of 3Q.</p><p>The challenge behind it is much older than generative AI. Michael Polanyi described the core idea decades ago: people know more than they can put into words. Ikujiro Nonaka built on that observation to explain how organizations convert tacit knowledge into explicit knowledge that others can share and reuse, and he developed the theory by studying Japanese manufacturers. I came at the same wall from the opposite direction. I started on the factory floor, where the knowledge was visible, and hit it in the office.</p><p>3Q is a practical way to apply that long-standing idea to AI-assisted work. It helps non-developers organize what they know into three parts: the problem, the desired result, and the business approach. It does not capture every detail, but it creates a clear starting point for describing the work to another person or to AI.</p><p>TCREI develops that starting point by adding the references, rules, evaluation criteria, and interaction needed to build and verify a working tool. Prompt engineering becomes useful after this foundation has been established.</p><p><strong>What changed once the work was described</strong></p><p>I began using 3Q with employees who found it difficult to explain work they had performed for years. Instead of asking them to describe the entire process at once, I asked them to identify the problem, the desired result, and the business approach separately. Three questions, answered one at a time, gave them a structure for examining work they had mostly performed out of habit.</p><p>Once the broad outline was visible, the conversation became more specific. Employees began to identify steps they had left out, exceptions that required judgment, calculations they trusted, and signs that indicated something was wrong. They could explain not only what they did, but also why particular steps and decisions mattered.</p><p>3Q did not capture every detail of the process. It made the work clear enough for those missing details to emerge. Asking for a complete description produces a summary. Asking three separate questions produces a conversation.</p><p>The purchasing conversation shows what happens next. Twelve decisions came out of it, and we did not try to solve twelve. One carried more weight than the rest: material ordered today arrives three months later, so the buyer has to decide now whether to increase the quantity, reduce it, or hold. Getting that one wrong costs real money in either direction. That single decision became its own 3Q worksheet, and the tool built from it <strong>comes later in this series.</strong></p><p>I learned that describing work is a skill of its own, and that most people have never been required to practice it.</p><p>The inefficiency I set out to remove was never really hidden on the screens. It sat in the rules, exceptions, and judgments that had never been written down. Once those were described clearly, AI could help turn them into a process that could be tested, improved, and eventually automated. Making the work explicit became the real starting point for using AI.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=063c4b9b6d1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Domain Experts Should Build Business Tools with AI]]></title>
            <link>https://medium.com/@workinthenow/why-domain-experts-should-build-business-tools-with-ai-1e586369a0cf?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e586369a0cf</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[business-tools]]></category>
            <category><![CDATA[ai-transformation]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Wed, 22 Jul 2026 11:26:01 GMT</pubDate>
            <atom:updated>2026-07-22T11:26:01.533Z</atom:updated>
            <content:encoded><![CDATA[<p>Small tools improve individual productivity. Tested prototypes help IT scale the right solution and preserve how the company actually works.</p><p>In the first article in this series, I described how AI helped us build a quality dashboard for my factory. The code was written by AI, but the important decisions came from the people who understood the production process.</p><p>They knew which measurements mattered, which readings had to be excluded, what the dashboard needed to show, and which known results the calculations had to reproduce before anyone could trust it.</p><p>That experience clarified where building business tools with AI creates value.</p><p>The first benefit is personal: domain experts can turn recurring work into small tools that improve their own productivity.</p><p>The second benefit appears when the tool becomes important enough to expand: the IT team receives a working, tested prototype instead of a vague request.</p><p>The third benefit belongs to the company: a process that once existed only in employees’ experience becomes explicit and reusable in the form of rules, tests, data structures, and software.</p><p>These three benefits matter more than whether the person who started the project can read every line of code.</p><h3><strong>Domain experts can solve the small problems IT never reaches</strong></h3><p>Most companies have a long list of small operational problems that are too specific to become formal IT projects.</p><p>A quality engineer prepares the same indicators before every shift meeting. A purchasing manager visits several websites and rebuilds the same price analysis each month. A finance employee copies data from an ERP export into last month’s workbook. Each task consumes time, but none may be large enough to compete with company-wide systems, security upgrades, or customer-facing projects.</p><p>The people doing this work already understand the problem. They know where time is lost, which exceptions matter, and what a useful result should look like.</p><p>AI gives them a practical way to turn that knowledge into a focused tool.</p><p>This is not about building a complete enterprise system. It is about creating a dashboard, report generator, reconciliation assistant, price tracker, or other small solution for a clearly defined workflow.</p><p>The domain expert still has to explain the work accurately. AI cannot guess that a zero sensor reading is startup noise, that one customer requires a different approval path, or that a report is useless if it arrives after the morning meeting.</p><p>The tool becomes useful only when the expert describes the sequence, rules, exceptions, and expected result clearly enough for AI to implement them.</p><p>Verification is equally important. At our factory, we did not trust the quality dashboard just because the charts appeared. We trusted it only after it reproduced values we had already calculated separately in a spreadsheet and correctly excluded readings that should not enter the analysis.</p><p>AI can generate a first version quickly. The domain expert decides whether that version reflects the real work.</p><p>When this process succeeds, the result is direct productivity improvement. The person who used to assemble the answer manually now has time to review it, investigate exceptions, and make decisions.</p><h3><strong>A tested tool gives IT a better starting point</strong></h3><p>A small internal tool should not automatically become company infrastructure.</p><p>Once it handles more users, sensitive data, important decisions, or connections to existing systems, professional IT review becomes necessary. Security, access control, reliability, maintainability, monitoring, and integration require technical expertise.</p><p>However, the work completed by the domain expert is not discarded at that point. It gives the IT team a much better starting point.</p><p>Traditionally, the handoff began with natural-language explanations. The domain expert described the process in the language of the business. The IT team interpreted that description through the language of systems, databases, and software development.</p><p>Both sides spoke natural language, but they did not always attach the same meaning to the same words.</p><p>A term that seems precise inside a department may be ambiguous outside it. A step that feels obvious to an experienced employee may never be mentioned. An exception handled automatically through habit may disappear entirely from the requirements document.</p><p>The IT team then had to translate an incomplete description into a working system. Misunderstandings and omissions were almost inevitable.</p><p>A tested business tool changes this handoff.</p><p>The domain expert can show the actual inputs, processing steps, outputs, error cases, and known correct results. The IT team does not receive only a sentence such as “We need a quality dashboard.” It receives a working example of what that dashboard does.</p><p>The specific programming language does not matter. The value comes from having the process expressed in an artificial language that forces greater precision than conversation alone.</p><p>Software has to specify what happens first, what happens next, which condition changes the path, which records are excluded, and what output each step produces. Those details may remain hidden in ordinary discussion, but they cannot remain hidden for long once the tool is run against real data.</p><p>The prototype therefore becomes an executable specification.</p><p>The domain expert can confirm that the business behavior is correct. The IT team can inspect the implementation, redesign the architecture, improve security, and connect the tool to enterprise systems. Both groups work from the same tested representation of the process.</p><p>This reduces the risk that the final system is technically complete but operationally wrong. It also helps IT include the small but essential functions that are often lost during a conventional requirements handoff.</p><h3><strong>The company preserves knowledge that used to live only in people</strong></h3><p>The third benefit is larger than individual productivity or faster IT delivery.</p><p>Every company depends on tacit knowledge.</p><p>An experienced employee knows which values are unreliable, which customer request needs escalation, which inventory mismatch is harmless, and which one signals a serious problem. Much of this knowledge is not written in procedures. It survives through habit, explanation, and apprenticeship.</p><p>That makes it fragile.</p><p>When the employee changes roles or leaves the company, the organization may keep the data but lose the reasoning used to interpret it.</p><p>Building business tools with AI creates an opportunity to preserve that reasoning.</p><p>To make a tool work, the domain expert has to state the rules. The person must explain which input is valid, what sequence should be followed, which exception changes the decision, and how the result should be checked.</p><p>Those explanations become more than notes. They become processing rules, test cases, sample data, expected outputs, and program behavior.</p><p>The company is not merely automating a task. It is capturing part of the way the work is actually performed.</p><p>This does not mean every judgment can or should be converted into software. Some decisions still require experience, negotiation, or human responsibility. The goal is not to remove judgment. The goal is to preserve the repeatable logic surrounding that judgment.</p><p>A well-built tool can show what to check, which case deserves attention, and which evidence supports the next decision. The employee remains responsible for the conclusion.</p><p>Over time, this creates a more durable form of organizational knowledge. Processes become easier to teach, review, improve, and transfer because they no longer exist only in employees’ memory.</p><h3><strong>A practical division of responsibility</strong></h3><p>The strongest model is not domain experts replacing IT.</p><p>It is a division of responsibility.</p><p>Domain experts identify narrow operational problems, build and test the first useful tools, and prove that those tools improve the work.</p><p>IT professionals review and expand the tools that deserve broader use. They make them secure, reliable, maintainable, and compatible with company systems.</p><p>The company benefits from both. Employees solve more of the small problems that slow them down. IT receives clearer requirements and tested examples. Valuable work methods become explicit instead of remaining hidden in individual experience.</p><p>That is the real promise of building business tools with AI.</p><p>The first result may be a simple dashboard used by one person. But behind it is a larger change: domain knowledge can move directly into a working form, technical teams can expand it with less misunderstanding, and the organization can retain more of what its experienced people know.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WHvxQE480xSXS0RSZ4DpOw.png" /></figure><p><strong>The first article documents the quality dashboard build, including the 3Q worksheet and verification process:</strong> [<a href="https://medium.com/@workinthenow/i-run-a-factory-ai-built-my-quality-dashboard-0ad2b752d099">I Run a Factory. AI Built My Quality Dashboard.]</a></p><p><strong>The prompt structure and sample database are in the companion repository:</strong> [<a href="https://github.com/Munjae-Song/medium_Build_Business_Tools_with_AI">Build Business Tools with AI on GitHub</a>]</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e586369a0cf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Run a Factory. AI Built My Quality Dashboard.]]></title>
            <link>https://medium.com/@workinthenow/i-run-a-factory-ai-built-my-quality-dashboard-0ad2b752d099?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ad2b752d099</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[manufacturing]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 12:06:00 GMT</pubDate>
            <atom:updated>2026-07-15T12:06:00.978Z</atom:updated>
            <content:encoded><![CDATA[<p>Before every shift meeting, our quality engineer spent thirty minutes preparing the quality indicators manually. A dashboard built in half a day reduced the process to a few clicks.</p><p>Every eight-hour shift, one inspection machine on our line generates about 9,600 measurements. We make flat panel components, and two numbers decide whether each unit is acceptable: outer dimension and thickness. Automated inspection equipment measures every single unit, roughly one reading every three seconds, and records every result. The data can then be exported as CSV files. Collecting the data has never been hard.</p><p>Before each shift, we hold a short meeting to confirm the previous shift ran clean. Preparing for it took our quality engineer thirty minutes, every shift, without exception. The work was repetitive: pull raw data, organize it across sources, then run each calculation separately for daily means, standard deviations, and process capability scores. By the time the meeting started, the numbers were ready, and the person who prepared them was already behind on everything else.</p><p>The line leader faced the more urgent consequence. When measurements begin moving gradually toward the edge of the acceptable range during a shift, someone has to decide whether production can continue or should be stopped. That decision cannot wait until the next meeting. It depends on figures that take the quality engineer thirty minutes to prepare. The factory already had thousands of measurements. The delay came from turning them into information the line leader could use.</p><h3><strong>Why we didn’t file a ticket</strong></h3><p>The measurement data comes out of the inspection equipment as CSV files. As the inspection equipment improved and began producing more detailed data, our existing internal systems did not keep pace. Updating those systems would require time, money, and coordination across the company. Until that work is justified and completed, the quality engineer has to collect the files, organize the data, and run the necessary calculations manually. That workflow is manageable for a scheduled meeting, but when the line leader must decide whether to continue or halt production, thirty minutes is too long.</p><p>I run this factory, and I am not a software engineer. What I had instead was a method.</p><h3><strong>Three questions on paper, before any AI</strong></h3><p>Before opening any AI tool, I answered three questions on paper. I now use this same worksheet for every internal build.</p><p><strong>What is the problem?</strong> One machine produces about 9,600 data points per shift — manageable to store, but time-consuming to organize and analyze manually. Preparing quality indicators for the pre-shift meeting takes thirty minutes of repetitive work, every shift. Without timely numbers, the line leader cannot make an informed decision when a problem appears, and every minute of delay increases the risk of defective units reaching the next stage.</p><p><strong>What result do you want?</strong> An interactive dashboard. The quality engineer selects a date and a product, then sees twenty-one days of data: daily mean and standard deviation, plus the process capability indicators Cpk and Ppk, shown through trend charts and distribution charts. A twenty-one-day view makes gradual changes visible that may be difficult to recognize from individual measurements.</p><p>For readers unfamiliar with these terms, Cpk and Ppk summarize how consistently a manufacturing process stays within its allowed limits.</p><p>The dashboard focuses on two measurements, dimension and thickness, and a small set of indicators already familiar to the quality team. I kept the scope narrow so the article could demonstrate the building method without becoming a lesson in statistical quality control.</p><p><strong>What approach will you use?</strong> The statistical indicators our quality team already trusts, delivered through a browser interface that users can open without installing software on their own computers.</p><p>That completed the 3Q worksheet. The three answers described the problem, the desired result, and the practical approach in the language of the work itself. They did not specify Python, software libraries, a database engine, or a charting framework. Those technical choices would come later.</p><p>The technical choices came afterward, through a question. I gave AI the constraints from the worksheet and asked which tools would fit them. It recommended Gradio, a Python library that turns code into an interface with dropdown menus, buttons, and charts, running inside Google Colab, a free browser-based workspace where code and results appear together. For storage, it recommended SQLite, a database stored as a single file in the same Google Drive folder as the Colab notebook.</p><p>One boundary matters here. The repository and public demo linked in this article use a prepared sample database, not live production data. Readers can adapt the dashboard to test their own data, but company data should be used only in an environment approved by their organization. The public demo is intended to show how the method works; it is not a place to upload or expose confidential production records.</p><h3><strong>The two prompt sections that decide everything</strong></h3><p>To turn the worksheet into an actual request, I used a five-part prompt structure called TCREI: Task, Context, References, Evaluate, Interact. The skeleton is adapted from a prompting framework Google teaches; I reworked its final step for this kind of building, a change I will explain properly in a later article. Each part has one job. Task states what to build and in what environment. Context explains who will use the tool, when, and for what decision. References carries the domain rules and facts that AI cannot guess from the data alone. Evaluate defines the tests the result must pass before anyone trusts it. Interact covers the follow-up conversation that turns the first working version into the finished tool.</p><p>The full TCREI prompt used for this build is available in the repository linked at the end of the article. Here, I want to focus on two sections, References and Evaluate, because they determine whether the resulting tool is accurate and reliable enough for real work.</p><p><strong>References is where your domain knowledge goes.</strong> In the References section, I included details that AI could not determine from the data alone. Our product tables encode the nominal dimension in their names: a table called DS_PRDT_805_20 represents a nominal dimension of 8.0 millimeters. I also specified a filtering rule: readings where dimension or thickness is zero or below are equipment initialization noise and must be excluded before any calculation.</p><p>Without that rule, the dashboard would still run and produce charts, but the results would be wrong. These details come from understanding how the data is structured, which values are valid, and what rules the analysis must follow. AI can apply that knowledge accurately, but only after the domain expert makes it explicit.</p><p><strong>Evaluate is where trust starts to get built.</strong> I deliberately kept this part simple. For product DS_PRDT_805_20 on 2020–06–15, the dimension Cpk had to equal 1.919 and the thickness Cpk had to equal 1.325. We had already calculated those values manually before building the dashboard, so they provided useful test cases for checking whether the Python calculations were correct. The analysis period also had to cover exactly twenty-one days, the nonpositive initialization readings had to be excluded, and every chart had to display a proper legend.</p><p>The dashboard produced the expected Cpk values of 1.919 and 1.325. More detailed validation would be required for a full production system, but for this example, matching those known values was enough to confirm that the calculation was working correctly.</p><h3><strong>The first version is not the last version</strong></h3><p>The first version then moved into the Interact step: run the dashboard, identify what needed improvement, and ask AI to revise it. In this case, the calculations were correct, but the layout was not practical for a meeting. The dimension and thickness charts appeared side by side, making each one too small to read comfortably. I described the problem to AI in plain language: “Display each chart in a single column so it takes up the full width of the screen.” AI revised the layout accordingly.</p><p>I also learned something about the tools themselves: the identical prompt, given to different AI models, returns different designs. One model showed every chart at once; another used a tabbed layout with one large chart per tab. Both versions worked. Choosing between them, based on how the team actually uses the dashboard, became part of the builder’s role. The quality engineer could now spend time judging which design supported the meeting best and refining the tool around the team’s workflow.</p><h3><strong>What changed on the floor</strong></h3><p>The thirty minutes of pre-shift preparation became the time it takes to open the tool, pick a date and product, and run the analysis. The indicators are ready for the meeting without rebuilding every calculation. When the quality team needs to examine another product or date, it can rerun the same verified workflow instead of starting the spreadsheet work again. And our quality engineer’s expertise now goes into interpreting the indicators rather than assembling them.</p><p>Total build time, including the paper worksheet, the prompt, the verification gate, and the layout fix: about half a day.</p><h3><strong>The part that transfers</strong></h3><p>The same approach can be applied in other industries. Replace Cpk with conversion rate and the inspection data with a CRM export, and the same structure can produce a sales dashboard. What transfers is the method: define the problem before the tool, hand AI the domain rules it cannot know unless you record them, and give it a known result it must reproduce before you rely on what it produces.</p><p>If you can answer the three questions, state the rules, and recognize a correct result, you can begin building this class of internal tool. That is not the same as completing a security review or production deployment. It is a practical way to turn domain knowledge into a testable first version.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vGN6WgXu8NWQdCAy13MjKw.png" /></figure><p><a href="https://github.com/Munjae-Song/medium_Build_Business_Tools_with_AI"><strong>Build it:</strong> The full TCREI prompt and prepared sample database (quality_data_small.db) are in the repository: [[GitHub repo link]]</a></p><p><a href="https://quality-dashboard-demo.streamlit.app/"><strong>Try the sample-data dashboard:</strong> [[demo link]]</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ad2b752d099" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Deleted OpenClaw and Built a Smaller Tool I Actually Use]]></title>
            <link>https://medium.com/@workinthenow/i-deleted-openclaw-and-built-a-smaller-tool-i-actually-use-ee1431d0d080?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/ee1431d0d080</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[telegram-bot]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Wed, 22 Apr 2026 09:40:29 GMT</pubDate>
            <atom:updated>2026-04-22T09:40:29.943Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The most valuable tool is not the one with the longest feature list.</em></p><p>When OpenClaw, an AI tool for personal workflows, was having a moment, I installed it too. Not long after, I deleted it.</p><p>Not because it was bad. If anything, it struck me as a well-made tool: lots of features, a substantial structure, and clear room to grow. The problem came after that. I could not imagine myself managing something that big for long.</p><p>Just because a tool is good does not mean you will keep using it. A long feature list does not guarantee that you will keep reaching for it. The tools I use most are usually simple. They have a clear purpose, respond quickly, and slip into my routine without much effort. That was the real reason I deleted OpenClaw. At some point, I decided that frequency of use mattered more than total functionality.</p><h4><strong>Instead of using a big tool, I built something smaller that fit me better</strong></h4><p>That does not mean OpenClaw gave me nothing. The part that stayed with me most was its approach to “memory.” The idea of saving a web page and later finding it again from only a rough clue, even when I could not remember the exact title, was genuinely appealing.</p><p>So instead of adopting the whole thing, I kept only the part I knew I would actually use.</p><p>The result was simple: rather than keep using OpenClaw, I built a small Telegram bot that fit the way I work.</p><p>A few years ago, that would not have been an easy choice. Building something yourself meant taking on the design, implementation, fixes, and maintenance all at once. For someone like me, programming was more of a hobby than a profession, so that burden felt especially heavy.</p><p>Now it feels different. Thanks to language models, and to what people call vibe coding, the cost of building small personal tools has dropped noticeably. You describe the feature you need in natural language, check the result, and keep refining it. You do not have to design the perfect structure from day one. You do not have to document every requirement in advance. You can start with the workflow you use most often and improve the weak spots as they show up.</p><p>I am not saying everyone can become a developer now. But even if programming is not your profession, it is now much easier to build simple tools for your own use.</p><h4><strong>I added only what I needed to an environment I already had</strong></h4><p>This choice was possible because I already had the environment in place. I had a free Oracle Cloud Infrastructure server running 24/7, and I was already using it to collect and store different kinds of information. In that situation, it was far more realistic to add only the features I needed than to start a large new system from scratch.</p><p>For personal tools, low operational burden matters more than flashy technology. That is one reason Telegram made sense. It is always open, I can issue commands as naturally as sending a message, and it was already an interface I used every day. I chose Python for the same reason. It was the fastest way to build something, and later it would be the easiest way to change.</p><p>What mattered here was not whether the stack was objectively superior. What mattered was whether it was familiar, immediately within reach, and cheap to maintain. With personal tools, it is easier to start something impressive than to keep something alive.</p><h4><strong>Once I kept only what I needed, I started using it more often</strong></h4><p>The bot I built is not ambitious. What matters more to me is that I keep using it. I built it around the way I actually live and work, so I kept only the functions I knew I would actually use.</p><p>The feature I use most is web page summarization and storage. When I find an interesting page while researching, I send the URL to Telegram. The bot summarizes it and saves it. Later, even if I cannot remember the title, I can find it again with only a rough description of the content. I tried bookmarks, spreadsheets, and separate database apps before, but none of them lasted. Saving was easy. Finding things again was not. Now I just forward a link. A language model summarizes it, the summary goes into a vector database, and later I can find it again with only a rough description.</p><p>The second feature is access to economic information. I was already collecting price data and other indicators, so I made them easy to check through my Telegram bot. That turned out to be much faster than opening financial websites and clicking through several screens. I picked the indicators I actually follow and used a language model to analyze them. That made it much lighter and easier for me to understand than a professional analysis tool with dozens of features. The point was to bring the information I check most often onto the shortest possible path.</p><p>The third feature is news curation. There is too much news and too little time. So every morning, the bot pulls out the topics that were mentioned most often among the subjects I care about, summarizes the related pages, and sends them to me first. Instead of reading every full article from the beginning, I use the summaries to grasp the flow and open the originals only when I need more detail. The language model also makes the summaries much more compact, and because they arrive through Telegram, they are easier for me to check than an email newsletter. That makes the whole workflow feel better suited to how I actually consume news.</p><p>When needed, I also added a simple web page. My Telegram bot is for quick checks. The web is for more detailed charts and granular data.</p><p>That setup turned out to be much more useful than checking financial websites, because the information was already arranged in the way that made the most sense to me. Once I split those roles, the whole system became more convenient. It was more usable to design around the flows I return to most often than to cram every function onto a single screen.</p><h4><strong>Building your own tool is now a realistic option</strong></h4><p>This experience made one thing clearer to me. Instead of forcing myself to adapt to a massive system, it has become a realistic alternative to build a small system that fits the way I actually work.</p><p>Of course, not everyone needs to build their own tools. Big tools still have real advantages. Well-made open source projects are still a great starting point for many people. I got ideas from OpenClaw too. But for me, having a small tool I open every day matters more than holding on to a large, polished tool over the long term.</p><p>That is why I deleted OpenClaw. Not because it was bad. If anything, it showed me very clearly what was good about it, and that made it easier to keep only the parts I would actually use.</p><p>In the end, what matters is not how many features a tool has, but how often you actually use it. And if existing tools do not fit, it is now much easier to build a simple one for yourself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MVA6jA9zi0BlSflp43cZOQ.png" /><figcaption>Deleted OpenClaw and Built a Smaller Tool I Actually Use</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ee1431d0d080" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Stopped Emailing Excel Reports and Built an AI Dashboard Instead]]></title>
            <link>https://medium.com/@workinthenow/i-stopped-emailing-excel-reports-and-built-an-ai-dashboard-instead-afb3986ac018?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/afb3986ac018</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[google-colab]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Mon, 06 Apr 2026 04:12:49 GMT</pubDate>
            <atom:updated>2026-04-06T04:12:49.180Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How a structured AI workflow turned a daily operational bottleneck into a real internal tool.</em></p><p>At 8:30 every morning, the report was already useless.</p><p>We were still exporting data from the database, pasting it into Excel, fixing the columns, updating the chart, and emailing the file around. By the time the numbers reached the quality engineer, the line had already moved on. It looked like monitoring. It was not. It was a delayed summary of a problem that was still happening.</p><p>That was the real problem. Not lack of data. We had plenty of it. Every unit on the line already produced dimension and thickness measurements. The factory was collecting thousands of records every shift. The problem was that turning raw data into usable charts and indicators required technical skill most people on the floor did not have — and even those who could do it were doing it too late.</p><p>So I replaced the daily Excel report with a small dashboard built in Google Colab. It pulled 21 days of inspection data, filtered out bad records, calculated quality indicators, and displayed the result in charts anyone on the team could read in seconds.</p><p>AI wrote the code. I defined the problem.</p><p>To get there, I had to do two things in the right order. First, define the work clearly enough that AI could build it. Second, make the result usable enough that other people would actually rely on it. This was not a practice exercise. It was a fix for a daily operational bottleneck.</p><p>And it was not one magic prompt. It was a workflow: define the problem, map the process, generate the first version, test it against a known case, and refine it with real users.</p><h3><strong>Start with the workflow, not the hype</strong></h3><p>When people say they want to build something with AI, they often start in the wrong place. “<em>I should use Python.</em>” “<em>I should build an app.</em>” “<em>I should automate this report.</em>” Those are not problem statements. They are jumping to answers.</p><p>I started with three questions:</p><ul><li><strong>What is the problem?</strong> Inspection data exists, but people cannot see it in time to act on it.</li><li><strong>What result do I want? </strong>An interactive dashboard that shows trends, variation, and capability at a glance.</li><li><strong>What approach will I use? </strong>Use Google Colab and Gradio, let AI generate the code, and iterate by testing the output.</li></ul><p>That was enough to get moving. I did not need to choose libraries first. I did not need to design the database logic first. I did not need to think like a developer first. I needed to think like the person closest to the problem.</p><p>That is the shift this article is really about. The technical work still matters, but it no longer has to be the first thing you solve. In practice, that means the useful unit is not “a prompt.” It is a workflow that helps you move from ambiguity to something testable.</p><h3>Then I mapped the work step by step</h3><p>Before asking AI to build anything, I wrote down the sequence:</p><ol><li>Select a date and a product.</li><li>Retrieve 21 days of inspection data.</li><li>Remove invalid records.</li><li>Calculate the quality indicators.</li><li>Display charts and a summary.</li><li>Wait for the next request.</li></ol><p>This looks simple. That is the point. Most operational work feels obvious until you try to describe it clearly. Then the hidden assumptions appear. What counts as invalid data? How many days matter? Which indicators matter? What should the user see first?</p><p>AI gets better when the process gets clearer. Without structure, AI amplifies vagueness. With structure, AI amplifies judgment.</p><h3>The workflow that built the dashboard</h3><p>I use a five-part structure called TCREI, adapted from a prompting framework originally proposed by Google. The original version used <em>Iterate</em> for the final step. I changed it to <em>Interact</em> because the real work is not repeating the same request — it is exchanging feedback with AI to improve the result.</p><ul><li><strong>Task</strong>: what the solution should do</li><li><strong>Context</strong>: why it is needed and who will use it</li><li><strong>References</strong>: the information AI needs to do the work</li><li><strong>Evaluate</strong>: how to verify the result</li><li><strong>Interact</strong>: how to refine it through feedback</li></ul><p>Here is how I applied it to this project:</p><p><strong>Task:</strong> Build an interactive quality dashboard in Google Colab with a date picker, a product dropdown, and a run button. When the user clicks the button, retrieve 21 days of inspection data, filter out invalid records, and display trend charts, capability metrics, histograms, and a summary table. Lay out dimension on the left and thickness on the right.</p><p><strong>Context:</strong> This dashboard is for quality engineers and line leaders who need to review process trends during daily meetings — without waiting for a manually prepared report. Anyone on the production floor should be able to check quality status without data handling skills.</p><p><strong>References:</strong> A SQLite database file containing dimension and thickness inspection records. Three product specifications, each with annual tables. Key fields: inspection date, dimension, and thickness. Records where dimension or thickness is zero or negative are equipment noise — exclude them. Dimension specification is derived from the product name. Thickness upper limit: 0.7.</p><p><strong>Evaluate:</strong> Test with a known product and date. Verify that all charts render correctly, exactly 21 days of data are retrieved, and capability calculations (Cpk and Ppk — measures of whether the process stays within specification limits) produce numbers that match what you already know to be correct.</p><p><strong>Interact:</strong> Run the dashboard and test each function. If charts are too small, ask AI to adjust the layout. If calculations seem wrong, describe the error and ask AI to debug. If the interface is confusing, describe what should change. Continue until the dashboard works as intended.</p><p>I pasted Task through Evaluate into Gemini. AI returned working Python code. I copied it into a Colab notebook and ran it. A few seconds later, the dashboard appeared in the browser. Today, this step is even simpler — Google Colab now has Gemini built in, so you can paste the request and generate the code without leaving the notebook.</p><p>No studying programming languages. No researching which tools to use. No spending the whole afternoon fixing errors. One clear request. One working result.</p><p>But ‘working’ was only the first threshold. So I checked it with a known product and date, and compared the numbers against what I knew to be correct. That step is easy to skip when people talk about prompts. It is the step that makes the output trustworthy.</p><h3>What the dashboard actually showed</h3><figure><img alt="The quality dashboard in Google Colab — trend charts, capability indicators, and histograms in a single view." src="https://cdn-images-1.medium.com/max/1024/1*tuSExt4l99Kel5Zh_XGG2w.png" /><figcaption>The quality dashboard in Google Colab — trend charts, capability indicators, and histograms in a single view.</figcaption></figure><p>The dashboard gave us three useful views.</p><ul><li><strong>Trend charts</strong> showed daily mean and standard deviation across the 21-day window. Is the process drifting?</li><li><strong>Capability charts</strong> showed Cpk and Ppk — how well the process stays within specification limits, and whether it does so consistently over time.</li><li><strong>Histograms</strong> showed the distribution for the selected date. What does today’s process actually look like?</li></ul><p>In quality control work, one view is never enough. You need to see the trend over time and the shape of the distribution on a single day — and you need to move freely between dates when something looks wrong.</p><p>That is where the Excel report hit its limits. In theory, you can add macros to make a spreadsheet interactive — but the file grows heavy with large datasets, security warnings pop up every time someone opens it, and building the macros requires the same technical skill the team did not have. A report answers the question it was designed to answer. A dashboard lets people ask the next question immediately.</p><p>Instead of waiting for tomorrow’s file, we could look at drift, variation, and capability while the line was still running.</p><h3>One URL changed the meeting</h3><p>At first, the dashboard was only visible in my Colab session. Then I noticed something useful: when Gradio launches inside Colab, it can generate a shareable public URL. No extra development. No server setup. No deployment project. I sent that link to the team before the next quality review meeting.</p><p>The meeting changed immediately. One person checked the latest date and reviewed current quality metrics. Another compared the line against a previous month on a phone. A third opened a different product line on a laptop.</p><p>The old meeting had a single driver. One person shared a screen, clicked through tabs, changed filters, and answered requests one by one. The new meeting had parallel analysis. Everyone was looking at the same solution, but each person controlled their own view. Nobody was waiting for one operator to drive the screen.</p><p>The dashboard did not become smarter. The data did not change. The formulas did not change. Access changed. Analysis that used to happen in sequence started happening in parallel. Questions that used to become follow-up tasks got answered in the room. Less time retrieving numbers. More time deciding what they meant.</p><h3>Generation was only the middle of the process</h3><p>AI gave me a working solution, not a finished one. That is where the last part of TCREI matters: Interact.</p><p>I made three simple revisions after people actually touched it:</p><ul><li>The first layout stacked everything vertically, so I asked AI to place dimension charts on the left and thickness charts on the right.</li><li>The first summary table showed too many decimals, so I asked AI to round values to four places.</li><li>The first mobile view was hard to read, so I asked AI to stack charts vertically on narrow screens.</li></ul><p>None of those fixes required me to understand the code deeply. They required me to notice what was wrong and describe it clearly.</p><p>That is the real working pattern with AI. Not magic. Not one-shot perfection. Clear direction, visible output, short feedback loops. Many AI demos stop too early. They show that something can be generated. They do not show whether someone would trust it enough to use it in the next meeting.</p><h3>The bigger lesson</h3><p>This article is about a quality dashboard. But the approach is not limited to quality control. The same workflow applies anywhere: define the problem clearly, choose the approach, map the process, give AI a structured request, and refine the output through use.</p><p>One question that always comes up: is it safe to use company data in Google Colab? For this project, the data was operational manufacturing measurements — no personal information, no regulated records. The files lived in Google Drive, the processing ran in the Colab session, and the session disappeared when the work was done. For highly regulated data, check with IT and legal first. But for ordinary operational metrics, the barrier is often lower than people assume.</p><p>If you are still updating spreadsheets by hand every morning, that does not automatically mean you need a bigger system. You may just need a clearer problem definition and a better way to work with AI. That was the lesson for me. The Excel file was not the real issue. The real issue was that the people who understood the problem best were too far from the solutions that could solve it.</p><p>AI shortened that distance. And once that happened, the daily report stopped being a ritual and started becoming a live decision tool.</p><p>The prompt, notebook, and project files are on GitHub:</p><p>[<a href="https://github.com/Munjae-Song/Build_Business_Tools_with_AI">GitHub repository link</a>]</p><p>The sample database file is available on Google Drive:</p><p>[<a href="https://drive.google.com/file/d/1e-g1MSGoQbX4jXOWskjazFH31fxdloCR/view?usp=sharing">Google Drive database link</a>]</p><p><strong>About the book</strong></p><p><em>Source code, notebooks, and sample data for Build Business Tools with AI: A Problem-First Playbook for Domain Experts, Analysts, and Managers (coming soon).</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=afb3986ac018" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Market Got KVTC and TurboQuant Wrong]]></title>
            <link>https://medium.com/@workinthenow/the-market-got-kvtc-and-turboquant-wrong-8298b12e5e7c?source=rss-22231eb9f058------2</link>
            <guid isPermaLink="false">https://medium.com/p/8298b12e5e7c</guid>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[semiconductors]]></category>
            <dc:creator><![CDATA[Munjae Song]]></dc:creator>
            <pubDate>Sun, 29 Mar 2026 03:14:53 GMT</pubDate>
            <atom:updated>2026-03-29T03:14:53.971Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>What these KV-cache breakthroughs actually show — and why efficiency is not the same thing as lower memory demand</strong></h4><p>When Nvidia and Google published KV-cache compression papers at ICLR 2026 in late March — <a href="https://openreview.net/forum?id=aNVKROYpLB">KVTC from Nvidia</a> and <a href="https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/">TurboQuant from Google</a> — memory semiconductor stocks sold off almost immediately. <a href="https://www.investing.com/news/stock-market-news/samsung-sk-hynix-slide-as-google-touts-ai-memory-compression-tech-turboquant-4581363">Samsung fell about 5% and SK Hynix about 6% in Seoul</a>, while <a href="https://www.marketwatch.com/story/microns-stock-is-dropping-is-google-partly-to-blame-d4cad72e">Micron dropped about 3.4% in U.S. trading</a>.</p><p>The market’s logic was simple: if AI can use less memory during inference, memory demand must weaken.</p><p>That conclusion moves too fast.</p><p>What these papers show directly is narrower. They show that certain inference workloads can be served more efficiently, especially on the KV-cache side. My argument begins after that point: in an early market with unsaturated demand, better efficiency does not necessarily reduce total resource consumption. It can do the opposite by lowering the cost of new use cases, expanding feasible workloads, and increasing total throughput.</p><p>That distinction matters. It is the difference between “AI needs less memory” and “the same hardware can now do more work.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jxLvsW9Y9nNkxZLg4vH0lQ.png" /><figcaption>More Work, Same Memory</figcaption></figure><h4><strong>The Bottleneck These Papers Actually Address</strong></h4><p>To see why the distinction matters, it helps to start with the problem the papers are solving.</p><p>During inference, large language models do not simply process the latest prompt in isolation. They keep a running record of prior tokens, the units of text a model processes, so the model can maintain context across a conversation or task. That running record is the KV cache, short for Key-Value cache. It helps the model answer the current question based on the prior context.</p><p>The longer the interaction becomes, the larger this cache grows. That creates pressure on expensive GPU memory. Once memory pressure rises too far, systems have to make trade-offs: shorten the usable context window, lower batch sizes, push data into slower memory tiers, or recompute work that has already been done. In practice, the KV cache becomes one of the important constraints on long-context inference.</p><p>That is the problem both papers are targeting. They are not broad claims that memory no longer matters. They are engineering attempts to reduce one specific source of inference overhead.</p><h4><strong>Two Different Ways to Compress the Same Burden</strong></h4><p>A car dealership analogy makes the difference between these two techniques intuitive.</p><p><strong>Nvidia’s KVTC — compressing by importance.</strong> Imagine parents bringing their college-bound child to buy a first car. The parents, the student, and the salesperson spend an hour talking. Over that hour, fuel economy and long-distance highway comfort come up again and again — the student has a 90-minute commute to campus. These topics dominate because they matter most. So the notes from that meeting reflect that: fuel economy, maintenance costs, and tire wear get detailed entries; long-distance comfort and driver-assist features get a full paragraph. Exterior styling gets a brief note — the student has no strong preferences, but wants something clean and unremarkable on campus.</p><p>KVTC works the same way. It uses PCA (Principal Component Analysis) — a statistical technique that identifies which elements in a dataset carry the most information — to figure out which tokens in the AI’s context carry the most weight. Topics that came up repeatedly and at length are recorded in detail. Topics that barely registered are compressed aggressively. The AI’s context record shrinks dramatically, but the important parts are preserved in full. The result: up to 20x compression in certain scenarios, with less than 1 score point drop in accuracy, across models from 1.5B to 70B parameters.</p><p><strong>Google’s TurboQuant — compressing by notation.</strong> Now imagine that same hour-long dealership conversation, but instead of deciding what to keep and what to cut, you change how you write everything down. Conventionally, a salesperson would write paragraph-length notes for each category. TurboQuant replaces that with a compact rating card:</p><p>- Fuel Economy: <strong>Excellent</strong> (32 mpg, all consumables recently replaced)</p><p>- Comfort: <strong>Medium</strong> (smooth suspension, no driver-assist features)</p><p>- Appearance: <strong>Medium</strong> (clean and presentable, no strong preferences)</p><p>Every item stays on the card. Nothing is thrown away. But long descriptions become short grades, and only items that need clarification get a parenthetical note.</p><p>That is exactly what Google’s compression technique — called quantization — does. Conventionally, every value in the KV cache is stored at full precision — 16 bits per number, like a paragraph of detailed notes. TurboQuant compresses each value down to just 3 bits — a short grade. The information is still there, just expressed more efficiently.</p><p>TurboQuant compresses the KV cache by at least 6x and speeds up attention computation by up to 8x on H100 GPUs (Nvidia’s widely used data center chips), with zero measured accuracy loss.</p><p><strong>The engineering trade-off.</strong> KVTC requires a one-time training step per model — like a salesperson learning a methodology for how to classify and prioritize information from any consultation, before ever meeting a customer. Once that methodology is learned, it applies to every conversation that follows. TurboQuant takes a different approach entirely. Instead of learning which information matters, it compresses precise numbers into categories — no model-specific training, no calibration needed. Think of it like changing the unit of measurement: instead of writing “32.4 mpg,” just check “Category 4.” Any salesperson applies it instantly, for any car, any customer.</p><p>The difference comes down to one trade-off: setup cost versus compression power.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/909/1*W2wu7L0dbZy9qn0dA5D46A.png" /><figcaption>Comparison table</figcaption></figure><p>Higher compression, more upfront investment. Instant deployment, moderate compression. Different contexts call for different tools — but both represent a generational leap over what came before.</p><h4><strong>What the Papers Show Directly — and What They Do Not</strong></h4><p>What the papers show directly is that inference can be served more efficiently per workload, especially in long-context or memory-constrained settings. That is a systems-level result.</p><p>What they do not show directly is that AI infrastructure memory demand will fall in aggregate. Reaching that conclusion requires a separate market-level argument about adoption, workload growth, and total throughput.</p><p>My view diverges from the market’s first reaction at exactly that point. These papers matter for memory demand, but they should be interpreted in the context of the current AI market. If this were a mature and saturated market, efficiency gains could reduce total resource consumption. But that is not the market AI is in today. In the current AI market, which is still early and still constrained by unmet demand, greater efficiency is more likely to expand usage than to shrink it. That is the core investment implication — and it is my interpretation, not a claim the papers themselves make.</p><h4><strong>Why Efficiency Often Expands Demand Instead of Shrinking It</strong></h4><p>This pattern is older than AI.</p><p>William Stanley Jevons made the classic observation in the 19th century: improvements in steam-engine efficiency did not reduce coal consumption. They made coal economical in more places, for more uses, at larger scale. Lower cost per unit of work expanded the market faster than efficiency reduced usage per task.</p><p>The same logic appears repeatedly in technology adoption. When computers became cheaper, firms did not buy fewer of them. They put computers into more functions, more departments, and more workflows.</p><p>I saw this firsthand. I spent a decade in the electronic components industry. When a new manufacturing process sharply improved automation rates, unit costs fell fast. Margins collapsed. A brutal price war followed, and some manufacturers did not survive. But the longer-term effect was market expansion. Products that had previously been too expensive became viable at scale. New applications opened. New buyers entered. Total volume grew far beyond what the old cost structure would have supported, and end users got better technology at a fraction of the original cost. The efficiency gain did not shrink the market. It created a bigger one.</p><p>That is the lens I would apply here. When inference becomes meaningfully cheaper, providers do not simply reduce infrastructure investment. In a fast-growing market, they are more likely to reinvest those efficiency gains into better service: longer context, higher throughput, lower latency, or more capable workflows. As long as user expectations keep rising, efficiency is more likely to expand demand than to shrink it.</p><h4><strong>Why AI Still Looks Like an Early Market</strong></h4><p>That reinvestment dynamic matters because AI still looks much more like an early market than a saturated one.</p><p><a href="https://openai.com/index/scaling-ai-for-everyone/">OpenAI said in February 2026</a> that ChatGPT had more than 900 million weekly active users and more than 50 million consumer subscribers. Just a year earlier, Reuters reported that OpenAI had <a href="https://www.reuters.com/technology/artificial-intelligence/openais-weekly-active-users-surpass-400-million-2025-02-20/">surpassed 400 million weekly active users</a>. That is extraordinary scale — but it also suggests that monetized usage is still a minority of total usage.</p><p>Workplace adoption tells a similar story. Gallup reported in late 2025 that [<a href="https://www.gallup.com/workplace/699689/ai-use-at-work-rises.aspx">45% of U.S. employees used AI at least a few times a year, while only about 10% used it daily</a>]. Awareness is broad. Habit is still uneven.</p><p>At the same time, the workload itself is changing. The market is no longer centered only on short question-and-answer chat. <a href="https://hbr.org/2025/04/how-people-are-really-using-gen-ai-in-2025">Harvard Business Review’s 2025 analysis</a> pointed to growing use in personal coaching, life organization, and goal-related support, while enterprise use is also expanding into coding, creative work, and other tasks that require more context and more sustained interaction.</p><p>Long-context usage is also getting easier to ship. Anthropic’s API docs said on March 13, 2026 that the <a href="https://docs.anthropic.com/en/release-notes/api">1M-token context window became generally available for Claude Opus 4.6 and Sonnet 4.6 at standard pricing</a>. That kind of move does not prove broad market demand by itself, but it does show that long-context capability is moving closer to mainstream product economics.</p><p>Each of these shifts tends to increase token consumption, context-management needs, or concurrency demands.</p><p>That matters because agentic usage is structurally heavier than simple chat. An agent often maintains state across multiple steps, calls tools, evaluates outputs, retries, and chains inference into longer workflows. Even if compression makes each step more efficient, the overall workload can still increase as usage moves from isolated prompts to extended tasks.</p><p>These papers do not mean memory matters less. They mean the same hardware can support more use. In the current AI market, that is more likely to expand demand than reduce it.</p><h4><strong>How to Read Breakthroughs Like This</strong></h4><p>This is why I do not think papers like KVTC and TurboQuant should be read in isolation.</p><p>In a market as young and fast-moving as AI, a technical improvement in one part of the stack does not automatically overturn the direction of the whole industry. AI is still expanding rapidly. Adoption is still spreading. Product expectations are still rising. And many parts of the market are still constrained less by lack of demand than by cost, latency, and system limits.</p><p>That is why I would be careful about reacting too quickly to a single technical result. The better approach is tounderstand what the technology actually does, identify which constraint it actually relaxes, and then ask how that change interacts with the broader market.</p><p><em>*Munjae Song is a manufacturing founder with 10+ years of operating experience and an MBA. He is the author of*</em> Build Business Tools with AI <em>*and writes about how domain expertise and AI tools come together.*</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8298b12e5e7c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>