[{"data":1,"prerenderedAt":2303},["ShallowReactive",2],{"writing-index":3,"profile":2292},[4,127,179,234,285,371,454,521,612,703,795,861,928,997,1093,1191,1260,1327,1396,1461,1530,1596,1671,1748,1853,2102],{"id":5,"title":6,"body":7,"date":113,"description":13,"draft":114,"excerpt":115,"extension":116,"meta":117,"navigation":114,"path":118,"seo":119,"slug":120,"status":115,"stem":121,"tags":122,"__hash__":126},"writing\u002Fwriting\u002Fvibe-coding-is-not-a-strategy.md","Vibe Coding Is Not a Strategy",{"type":8,"value":9,"toc":104},"minimark",[10,14,17,22,25,60,64,67,70,74,77,91,94,98,101],[11,12,13],"p",{},"Every software team, especially in startups, is now building with AI at full\nspeed. Prompt, generate, ship. Output is up and timelines are down, and that looks\nlike an unqualified win. For a CXO, it should also be a warning.",[11,15,16],{},"Because speed of code is not soundness of system. AI is very good at producing\ncode that is fast, plausible, and works on the happy path. What it quietly skips is\neverything that does not show up in a demo. That skipped layer is where your real\nrisk now lives.",[18,19,21],"h2",{"id":20},"what-vibe-coding-leaves-behind","What vibe coding leaves behind",[11,23,24],{},"When a team lets AI drive without a frame, five things go missing, and none of them\nare visible until later.",[26,27,28,36,42,48,54],"ul",{},[29,30,31,35],"li",{},[32,33,34],"strong",{},"No design standards."," Every feature is built its own way. The codebase has no\ncoherent shape, because there was never a shape to hold it to.",[29,37,38,41],{},[32,39,40],{},"Fragile stabilization."," It works in the demo, then wobbles under real load,\nreal data, and the edge cases nobody prompted for.",[29,43,44,47],{},[32,45,46],{},"Lost maintainability."," Nobody, not even the AI, fully understands the code.\nChanging it becomes risky, and every change gets slower, not faster.",[29,49,50,53],{},[32,51,52],{},"Component mismatch."," Pieces generated in isolation do not fit together. The\nintegration seams are where it falls apart.",[29,55,56,59],{},[32,57,58],{},"Missing architecture."," Security, deployment, scale, observability. The things\nthe AI was never asked about are absent by default, and absent is the dangerous state.",[18,61,63],{"id":62},"speed-without-a-frame-is-debt","Speed without a frame is debt",[11,65,66],{},"This is not an argument against AI in software. AI is a real accelerator, and I use\nit every day. It is an argument that acceleration without a framework is not speed.\nIt is debt. You are borrowing velocity today and paying it back with interest in\nyear two: rework, incidents, a security exposure you did not know you shipped, and a\nsystem nobody can safely change.",[11,68,69],{},"The irony is sharp. The more code AI writes, the more the architecture matters, not\nless. The generation got cheap. The judgment about what to generate, and how it fits\ntogether, got more valuable.",[18,71,73],{"id":72},"the-frame-cxos-need-to-insist-on","The frame CXOs need to insist on",[11,75,76],{},"Embracing AI in development is right. Doing it without a stronger frame is the\nmistake. What that frame looks like:",[26,78,79,82,85,88],{},[29,80,81],{},"Clear design standards and a reference architecture the AI must build within.",[29,83,84],{},"Security, deployment, and scale defined up front, not discovered in production.",[29,86,87],{},"Human architectural ownership over what the AI produces. The AI writes; a person\nowns the system.",[29,89,90],{},"Reviews that check the system, not just the feature.",[11,92,93],{},"None of this slows a good team down. It is the guardrail that lets them run fast\nwithout driving off a cliff.",[18,95,97],{"id":96},"the-part-that-does-not-build-itself","The part that does not build itself",[11,99,100],{},"The frame does not appear on its own. Someone has to set the standards, own the\narchitecture, and hold the line on the parts that never demo well: security, scale,\nand the shape of the whole. That is the work I have done for twenty years, and it is\nexactly the work AI makes more important, not less.",[11,102,103],{},"Your team is shipping faster than it ever has. The question worth asking, before\nyear two asks it for you: do you know what it is quietly skipping?",{"title":105,"searchDepth":106,"depth":106,"links":107},"",3,[108,110,111,112],{"id":20,"depth":109,"text":21},2,{"id":62,"depth":109,"text":63},{"id":72,"depth":109,"text":73},{"id":96,"depth":109,"text":97},"2026-08-05",true,null,"md",{},"\u002Fwriting\u002Fvibe-coding-is-not-a-strategy",{"title":6,"description":13},"vibe-coding-is-not-a-strategy","writing\u002Fvibe-coding-is-not-a-strategy",[123,124,125],"ai","software","cxo","-6jrptd9K98o6VUwtZru27jrsvzvf8a6pZ_6L4m3uB8",{"id":128,"title":129,"body":130,"date":170,"description":134,"draft":114,"excerpt":115,"extension":116,"meta":171,"navigation":114,"path":172,"seo":173,"slug":174,"status":115,"stem":175,"tags":176,"__hash__":178},"writing\u002Fwriting\u002Fthe-coming-ai-write-down.md","The Coming AI Write-Down",{"type":8,"value":131,"toc":165},[132,135,139,142,145,149,152,155,159,162],[11,133,134],{},"For two days I have argued that AI is moving faster than we can absorb. Here is\nwhere that ends if we are not careful. A large share of today's AI spend will never\nshow up as value. Some of it will quietly be written off. The only question worth\nasking is whether any of it is yours.",[18,136,138],{"id":137},"hype-funds-the-future-ahead-of-its-time","Hype funds the future ahead of its time",[11,140,141],{},"Hype is not lying. It is just early. It pulls investment forward, ahead of the value\nthe technology can actually realize this year. That is not always wrong, some of it\nis how the future gets built. But when the gap between what you spent and what you\ncan show persists, it does not disappear. It gets written down: the abandoned\nplatform, the unused licenses, the reorganization that produced slides instead of\noutcomes.",[11,143,144],{},"We have seen this pattern before. The dot-com era was right about the internet and\ncatastrophically wrong about the timing and the price. The technology was real. The\nvaluations were not. A lot of correct bets still lost money, because they were made\ntoo early and too expensively.",[18,146,148],{"id":147},"this-is-not-doom-it-is-discipline","This is not doom, it is discipline",[11,150,151],{},"Let me be clear, because I build this stuff and believe in it. AI is real, and the\nvalue is real. But value realized is not the same as value promised, and the\ndifference is a number on a balance sheet. The firms that come through this well are\nnot the ones that spent the most or the least. They are the ones that treated AI as\ncapital allocation with discipline, not FOMO with a budget.",[11,153,154],{},"That means naming the outcome in a real unit, funding the ability to absorb as hard\nas the ability to adopt, and being honest about which bets are grounded and which\nare faith.",[18,156,158],{"id":157},"the-question-to-sit-with","The question to sit with",[11,160,161],{},"So before your next AI budget, one exercise. Split your AI investments into two\nlists: the ones you could defend in a downturn, with a number, and the ones you are\ncarrying on faith. If the second list is longer than the first, you are not\ninvesting in AI. You are buying hype, and hype is exactly the thing that gets\nwritten down.",[11,163,164],{},"The technology is not the risk. Getting drunk on it is. Think again.",{"title":105,"searchDepth":106,"depth":106,"links":166},[167,168,169],{"id":137,"depth":109,"text":138},{"id":147,"depth":109,"text":148},{"id":157,"depth":109,"text":158},"2026-08-04",{},"\u002Fwriting\u002Fthe-coming-ai-write-down",{"title":129,"description":134},"the-coming-ai-write-down","writing\u002Fthe-coming-ai-write-down",[123,177,125],"roi","7p0XtAtJ5_gCyJPx0VPCpMfodv_MD5E19nYQaCwjJQI",{"id":180,"title":181,"body":182,"date":225,"description":186,"draft":114,"excerpt":115,"extension":116,"meta":226,"navigation":114,"path":227,"seo":228,"slug":229,"status":115,"stem":230,"tags":231,"__hash__":233},"writing\u002Fwriting\u002Fgulping-more-than-we-can-digest.md","Are We Gulping More AI Than We Can Digest?",{"type":8,"value":183,"toc":220},[184,187,191,194,197,201,204,207,211,214,217],[11,185,186],{},"Yesterday I wrote that AI is running ahead of even the analysts who forecast it.\nToday, the harder question. If the technology is this far ahead, can our\norganizations actually keep up? I do not think most can, and I do not think that is\na failure. It is physics.",[18,188,190],{"id":189},"adoption-is-not-absorption","Adoption is not absorption",[11,192,193],{},"You can buy frontier intelligence in an afternoon. You cannot buy the things that\nturn it into value at the same speed: clean and ready data, redesigned processes, an\nevaluation discipline, change management, and the plain human trust it takes for\npeople to rely on a machine. Those move at the pace of an organization, not the pace\nof a model release.",[11,195,196],{},"Pour more capability into a company that cannot digest it and you do not get more\nvalue. You get indigestion. A wall of pilots. Shelfware. Half-built systems no one\nowns. Budgets that rise while the results stay flat. Eating faster does not make you\nbetter nourished.",[18,198,200],{"id":199},"the-bottleneck-moved-and-nobody-moved-with-it","The bottleneck moved, and nobody moved with it",[11,202,203],{},"For two years the constraint was the model. Now the model is the easy part. The\nconstraint has quietly shifted to absorption, the unglamorous layer of data,\nevaluation, governance, and people. That layer did not get faster just because the\nmodels did. If anything, the faster the models move, the more exposed that layer becomes.",[11,205,206],{},"This is why so much AI spend produces motion without progress. The organization is\nbeing asked to swallow capability faster than it can chew, and the gap shows up as\npilots that never ship and value that never lands.",[18,208,210],{"id":209},"what-a-serious-leader-asks-instead","What a serious leader asks instead",[11,212,213],{},"The question is not how fast can we adopt AI. Everyone can adopt fast, that is the\neasy part now. The question is how fast can we absorb it. That is set by your data\nreadiness, your platform, your eval loop, and your people, and you cannot shortcut\nit with a bigger model or a bigger budget.",[11,215,216],{},"So before the next wave of AI you bring in: what is your absorption rate, honestly,\nand are you funding capability far faster than you are funding the ability to digest it?",[11,218,219],{},"Tomorrow, the consequence of getting this wrong: the AI write-down.",{"title":105,"searchDepth":106,"depth":106,"links":221},[222,223,224],{"id":189,"depth":109,"text":190},{"id":199,"depth":109,"text":200},{"id":209,"depth":109,"text":210},"2026-08-03",{},"\u002Fwriting\u002Fgulping-more-than-we-can-digest",{"title":181,"description":186},"gulping-more-than-we-can-digest","writing\u002Fgulping-more-than-we-can-digest",[123,232,125],"adoption","_y_NV3m-x1-YD-bJcbQAm51k3roYh9RYRP7Z7GQfbOY",{"id":235,"title":236,"body":237,"date":275,"description":241,"draft":276,"excerpt":115,"extension":116,"meta":277,"navigation":114,"path":278,"seo":279,"slug":280,"status":115,"stem":281,"tags":282,"__hash__":284},"writing\u002Fwriting\u002Fgartner-forecast-already-behind.md","Gartner's AI Forecast Is Already Behind",{"type":8,"value":238,"toc":271},[239,242,245,248,252,255,258,262,265,268],[11,240,241],{},"A few days ago a leadership team showed me a Gartner report to back their case for\nwhere AI is heading. Solid report. Also eighteen months old.",[11,243,244],{},"In most fields, an eighteen-month-old analyst report is still current. In AI,\neighteen months is an era. So I read it again, end to end. Not to catch Gartner\nout, the analysts did careful, honest work. I read old forecasts on purpose,\nbecause the gap between what we predicted and what actually arrived is the most\nhonest teacher this field has.",[11,246,247],{},"The report is the Emerging Tech Impact Radar for Generative AI, February 2025\n(ID G00809486). What struck me was not where it was wrong. It was how far we have\nalready run past it. Things it placed three to six years out are here now: agentic\nAI, reasoning models, cheap inference. The forecast did not age. Reality overtook it.",[18,249,251],{"id":250},"why-that-is-a-warning-not-a-win","Why that is a warning, not a win",[11,253,254],{},"It is tempting to feel clever for beating the forecast. I would resist that. When\nthe technology outruns the people whose job is to see it coming, it is also\noutrunning the slower things: your governance, your data, your people, your ability\nto prove any of this creates value.",[11,256,257],{},"Speed of capability is not speed of value. They run on different clocks. And here\nis the part that should worry a CXO: the faster capability moves, the wider that gap\ngrows, not the narrower. We keep celebrating adoption velocity and mistaking it for\nvalue velocity. They are not the same, and the difference is where money quietly burns.",[18,259,261],{"id":260},"the-question-under-the-question","The question under the question",[11,263,264],{},"An out-of-date map is not dangerous when you know it is old. It is dangerous when\nyou are still steering by it. Plenty of AI roadmaps in boardrooms right now are\nbuilt on last year's assumptions about cost, capability, and timing, all of which\nhave already moved.",[11,266,267],{},"So over the next three days I want to sit with three questions I cannot shake. Are\nwe moving faster than we can absorb? Are we creating value, or just spending? And if\nwe get this wrong, how much of today's AI investment gets written off as hype we\ncould not digest?",[11,269,270],{},"Start here: if the technology is eighteen months ahead of the forecast, is your\nvalue capture eighteen months ahead too, or eighteen months behind?",{"title":105,"searchDepth":106,"depth":106,"links":272},[273,274],{"id":250,"depth":109,"text":251},{"id":260,"depth":109,"text":261},"2026-08-02",false,{},"\u002Fwriting\u002Fgartner-forecast-already-behind",{"title":236,"description":241},"gartner-forecast-already-behind","writing\u002Fgartner-forecast-already-behind",[123,283,125],"strategy","MWKeu-83Y0h84xgW9iYZdVK-HkajK9c_xIFI8Dj2Zy0",{"id":286,"title":287,"body":288,"date":362,"description":292,"draft":276,"excerpt":115,"extension":116,"meta":363,"navigation":114,"path":364,"seo":365,"slug":366,"status":115,"stem":367,"tags":368,"__hash__":370},"writing\u002Fwriting\u002Fagents-are-the-new-apps.md","Agents Are the New Apps",{"type":8,"value":289,"toc":356},[290,293,297,300,304,307,328,332,335,339,350,353],[11,291,292],{},"For thirty years, software meant buttons. You clicked, it responded. The user\nsupplied the intent, and the app supplied the mechanism. Agentic AI inverts\nthat. You state the goal, and the software works out the steps. That sounds like\na feature. It is actually a new category.",[18,294,296],{"id":295},"from-mechanism-to-intent","From mechanism to intent",[11,298,299],{},"An app waits to be told what to do, one step at a time. An agent is handed an\noutcome and decides the steps itself, calling tools, reading data, and retrying\nwhen it fails. The interface stops being a screen full of controls and becomes a\nconversation about what you want. That is why \"agents are the new apps\" is not a\nslogan. It is a shift in where the intelligence sits, from the user's head into\nthe software.",[18,301,303],{"id":302},"what-actually-changes","What actually changes",[11,305,306],{},"Three things break the old model.",[308,309,310,316,322],"ol",{},[29,311,312,315],{},[32,313,314],{},"The unit of work changes."," You no longer design screens, you design goals,\ntools and guardrails. The screen becomes optional.",[29,317,318,321],{},[32,319,320],{},"The failure mode changes."," A button either works or it does not. An agent\ncan do the wrong thing, confidently. Reliability becomes a design problem, not\na testing step at the end.",[29,323,324,327],{},[32,325,326],{},"The org changes."," If a capable agent can carry a task end to end, the\nquestion is no longer \"which app does the team use,\" it is \"which work does the\nteam still own.\" That is an operating-model question, and most companies have\nnot asked it yet.",[18,329,331],{"id":330},"the-trap","The trap",[11,333,334],{},"The easy mistake is to bolt an agent onto an existing app and call it\ntransformation. That is a chatbot standing in front of the same buttons. The real\nmove is to redesign the workflow around the goal: decide what the agent owns, what\nthe human owns, and where the hard stop lives. An agent without a guardrail is not\nautonomy. It is liability.",[18,336,338],{"id":337},"what-to-do-now","What to do now",[26,340,341,344,347],{},[29,342,343],{},"Pick one bounded, high-volume workflow and design it goal-first, not screen-first.",[29,345,346],{},"Give the agent a narrow tool surface and an explicit point where it must stop and ask.",[29,348,349],{},"Measure it like a colleague, on outcomes and error rate, not on usage.",[11,351,352],{},"The companies that win the next decade will not be the ones with the most AI\nfeatures. They will be the ones that worked out which work to hand to an agent,\nand which to keep.",[11,354,355],{},"So here is the question for your product and your org: which of your buttons\nshould have been a goal?",{"title":105,"searchDepth":106,"depth":106,"links":357},[358,359,360,361],{"id":295,"depth":109,"text":296},{"id":302,"depth":109,"text":303},{"id":330,"depth":109,"text":331},{"id":337,"depth":109,"text":338},"2026-08-01",{},"\u002Fwriting\u002Fagents-are-the-new-apps",{"title":287,"description":292},"agents-are-the-new-apps","writing\u002Fagents-are-the-new-apps",[123,369,283],"agentic","ax8yn5jOcclSvlUVqiFDRpqvbPJiJZhrg1gC5mjp0Ek",{"id":372,"title":373,"body":374,"date":362,"description":378,"draft":276,"excerpt":115,"extension":116,"meta":446,"navigation":114,"path":447,"seo":448,"slug":449,"status":115,"stem":450,"tags":451,"__hash__":453},"writing\u002Fwriting\u002Fbuild-buy-or-rent-intelligence.md","Build, Buy, or Rent Intelligence",{"type":8,"value":375,"toc":441},[376,379,383,386,406,409,413,416,418,438],[11,377,378],{},"Every enterprise faces the same question with AI, and most answer it badly: do we\nbuild it, buy it, or rent it. It is not one decision, it is a decision per layer,\nand getting it right is the difference between owning a compounding asset and paying\na rising bill for a capability that never becomes yours.",[18,380,382],{"id":381},"the-three-layers","The three layers",[11,384,385],{},"Think of your AI stack in three layers, and decide each separately.",[26,387,388,394,400],{},[29,389,390,393],{},[32,391,392],{},"The model."," Almost always rent. The frontier moves too fast and costs too much\nto build your own, and a general model is not where your advantage lives anyway.\nRent it, and keep the freedom to swap it.",[29,395,396,399],{},[32,397,398],{},"The workflow and product."," Almost always build. This is the part specific to\nyou: your process, your integrations, your edge cases. It is where the value and\nthe defensibility live. Do not outsource your moat.",[29,401,402,405],{},[32,403,404],{},"The data and evals."," Always own. Your proprietary data, and the evaluation loop\nthat measures and improves your system, are the compounding asset. Whoever owns\nthis owns the advantage in three years. Never rent it, never give it away inside a\nvendor's platform.",[11,407,408],{},"The most common and most expensive mistake is inverting this: renting the workflow\n(a generic tool everyone has) while giving away the data (into the vendor's system).\nThat is renting your moat and selling your advantage, in one move.",[18,410,412],{"id":411},"the-test-for-each-layer","The test for each layer",[11,414,415],{},"Ask: does this layer compound, and is it specific to us? If yes to both, own it. If\nit is fast-moving, capital-intensive, and generic, rent it. If it is specific but not\ncore, buy it and integrate. Run every AI decision through that filter.",[18,417,338],{"id":337},[308,419,420,426,432],{},[29,421,422,425],{},[32,423,424],{},"Rent the model, keep it swappable."," Avoid architectures that lock you to one\nvendor's weights.",[29,427,428,431],{},[32,429,430],{},"Build the workflow and own the data."," That is where durable return lives.",[29,433,434,437],{},[32,435,436],{},"Watch for the inversion."," If a vendor's pitch has you renting the workflow and\nfeeding them your data, you are on the wrong side of every layer.",[11,439,440],{},"Build vs buy was always really a question about where to own the compounding asset.\nRent made it three-sided, and easier to get wrong. So layer by layer: what are you\nbuilding, what are you renting, and are you sure you did not swap the two?",{"title":105,"searchDepth":106,"depth":106,"links":442},[443,444,445],{"id":381,"depth":109,"text":382},{"id":411,"depth":109,"text":412},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fbuild-buy-or-rent-intelligence",{"title":373,"description":378},"build-buy-or-rent-intelligence","writing\u002Fbuild-buy-or-rent-intelligence",[125,283,452],"architecture","fTIL6309fvNlj00qWvUoQCtXAkFUX_iQ6Z-ln3QeCec",{"id":455,"title":456,"body":457,"date":362,"description":461,"draft":276,"excerpt":115,"extension":116,"meta":512,"navigation":114,"path":513,"seo":514,"slug":515,"status":115,"stem":516,"tags":517,"__hash__":520},"writing\u002Fwriting\u002Fbuild-for-bharat-first.md","Build for Bharat First",{"type":8,"value":458,"toc":507},[459,462,466,469,472,476,479,481,501,504],[11,460,461],{},"The default instinct for an ambitious founder here is to build for the West, chase\ndollar customers, and hope it eventually helps at home. I would flip it. The\nlarger, more durable opportunity is to solve a hard Bharat problem first. Our\nproblems are large, specific, and under-served, and a solution that works at our\nscale and diversity tends to travel to the rest of the world. Build for Bharat first.",[18,463,465],{"id":464},"why-bharat-is-the-harder-better-training-ground","Why Bharat is the harder, better training ground",[11,467,468],{},"Our problems come with constraints that make for stronger products. Many\nlanguages, not one. Low bandwidth and shared devices, not unlimited compute in\nevery pocket. Users who speak rather than type. Price points measured in rupees,\nnot dollars. Informal systems where clean data does not exist. If your AI works\nunder these constraints, it is genuinely robust. A product that assumes English,\nfast internet, high willingness to pay, and tidy data is a fragile product that\nhappens to work in easy conditions.",[11,470,471],{},"And the market is not small. It is a billion-plus people entering the digital\neconomy at once, on a digital public infrastructure of identity, payments and data\nrails that most countries would envy. That is not a charity case. It is one of the\nlargest greenfields on earth.",[18,473,475],{"id":474},"the-reverse-trickle","The reverse trickle",[11,477,478],{},"Here is the part founders miss. Solve for the hardest version of a problem and the\neasy version comes for free. A voice agent that handles ten Indian languages and\nnoisy audio on a cheap phone will handle English on a good phone trivially. Frugal,\nrobust, multilingual AI built for Bharat is exactly what much of the Global South,\nand plenty of the West, will need next. The solution travels outward. It rarely\ntravels the other way.",[18,480,338],{"id":337},[308,482,483,489,495],{},[29,484,485,488],{},[32,486,487],{},"Pick a problem painful specifically here."," Languages, agriculture, credit for\nthe thin-file borrower, healthcare access, small-business compliance.",[29,490,491,494],{},[32,492,493],{},"Design for the constraint, not around it."," Voice-first, low-bandwidth,\nlow-cost, offline-tolerant. The constraint is the moat.",[29,496,497,500],{},[32,498,499],{},"Build on the rails we already have."," Identity, payments and consented data\ninfrastructure are a head start most markets do not have.",[11,502,503],{},"A developed Bharat by 2047 will not be built by importing solutions. It will be\nbuilt by founders who treated our hardest problems as their best opportunities.",[11,505,506],{},"So what is the Bharat problem you understand better than anyone in Silicon Valley\never will, and why not build that?",{"title":105,"searchDepth":106,"depth":106,"links":508},[509,510,511],{"id":464,"depth":109,"text":465},{"id":474,"depth":109,"text":475},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fbuild-for-bharat-first",{"title":456,"description":461},"build-for-bharat-first","writing\u002Fbuild-for-bharat-first",[518,123,519],"startup","bharat","qYKU3-BMb9gSjD__SOOrsNQVmleaP62xfVpg3k3Boa4",{"id":522,"title":523,"body":524,"date":362,"description":528,"draft":276,"excerpt":115,"extension":116,"meta":602,"navigation":114,"path":603,"seo":604,"slug":605,"status":115,"stem":606,"tags":607,"__hash__":611},"writing\u002Fwriting\u002Fcompliance-is-the-killer-app.md","The Compliance Workflow Is the Killer App",{"type":8,"value":525,"toc":597},[526,529,533,536,562,565,569,572,574,594],[11,527,528],{},"If you want to find the highest-return AI project in a large enterprise, do not\nlook at the innovation lab. Look at the compliance department. KYC checks, claims\nadjudication, audit prep, contract review, regulatory reporting. These\ndocument-heavy, rule-bound processes are the least glamorous work in the building,\nand that is exactly why they are the killer app for AI.",[18,530,532],{"id":531},"why-compliance-is-the-sweet-spot","Why compliance is the sweet spot",[11,534,535],{},"These workflows have every property that makes AI pay.",[26,537,538,544,550,556],{},[29,539,540,543],{},[32,541,542],{},"Volume."," They run constantly, at scale, across every customer or transaction.",[29,545,546,549],{},[32,547,548],{},"Cost."," They are staffed by expensive, skilled people doing repetitive reading and cross-checking.",[29,551,552,555],{},[32,553,554],{},"Structure."," The rules are written down. There is a right answer, and a defensible reason for it.",[29,557,558,561],{},[32,559,560],{},"Pain."," They are slow, they bottleneck the business, and getting them wrong carries real regulatory risk.",[11,563,564],{},"AI that reads a document, extracts the facts, checks them against the rules, and\nflags the exceptions turns a slow, costly, manual bottleneck into a fast,\nconsistent, auditable one. The value is immediate, and measurable in a unit the CFO\nand the risk committee both recognize.",[18,566,568],{"id":567},"the-catch-that-is-also-the-point","The catch that is also the point",[11,570,571],{},"Compliance is high stakes, so it demands the discipline AI needs anyway:\ntraceability, citations, a clear audit trail, and a human on the exceptions. That\nfeels like friction. It is actually why compliance is a good first home for\nenterprise AI. It forces you to build AI you can trust and explain, which is the\ncapability you will need everywhere else. You get a high-ROI win and build the\ngovernance muscle at the same time.",[18,573,338],{"id":337},[308,575,576,582,588],{},[29,577,578,581],{},[32,579,580],{},"Map your regulated, document-heavy processes."," That map is a ranked list of\nAI opportunities hiding in plain sight.",[29,583,584,587],{},[32,585,586],{},"Demand traceability by design."," Every AI decision should cite the document\nand the rule it came from. In compliance that is not optional, and that is a feature.",[29,589,590,593],{},[32,591,592],{},"Keep the human on the exceptions, not the routine."," Let AI clear the ninety\npercent that is clear, and route the hard ten percent to your experts.",[11,595,596],{},"The most boring department in your company may be sitting on your best AI business\ncase. So which of your compliance bottlenecks is quietly the most expensive, and\nwhy is it not first in line?",{"title":105,"searchDepth":106,"depth":106,"links":598},[599,600,601],{"id":531,"depth":109,"text":532},{"id":567,"depth":109,"text":568},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fcompliance-is-the-killer-app",{"title":523,"description":528},"compliance-is-the-killer-app","writing\u002Fcompliance-is-the-killer-app",[608,609,610],"enterprise","compliance","value","i7BOzo1usVgiub680AczvVFD4-HtmZqiz9jcTVkadvk",{"id":613,"title":614,"body":615,"date":362,"description":619,"draft":276,"excerpt":115,"extension":116,"meta":696,"navigation":114,"path":697,"seo":698,"slug":699,"status":115,"stem":700,"tags":701,"__hash__":702},"writing\u002Fwriting\u002Fcopilot-to-autopilot.md","From Copilot to Autopilot",{"type":8,"value":616,"toc":691},[617,620,624,650,653,657,660,663,665,685,688],[11,618,619],{},"Most AI in the enterprise today is a copilot. It suggests, drafts, recommends, and\na human decides. That is the safe starting point, and it is also a waypoint, not a\ndestination. Every use case travels a curve from copilot to autopilot, and the\nskill that separates good AI programs from stuck ones is knowing where each use\ncase sits, and when it has earned the right to move.",[18,621,623],{"id":622},"the-curve","The curve",[26,625,626,632,638,644],{},[29,627,628,631],{},[32,629,630],{},"Suggest."," The AI proposes, the human does everything. Maximum safety, minimum leverage.",[29,633,634,637],{},[32,635,636],{},"Draft."," The AI produces the work, the human reviews and edits. Most enterprise value today lives here.",[29,639,640,643],{},[32,641,642],{},"Act with approval."," The AI does the task and executes once a human clicks yes. The human becomes an approver, not a doer.",[29,645,646,649],{},[32,647,648],{},"Autopilot."," The AI acts on its own within defined bounds, and the human is pulled in only on exceptions.",[11,651,652],{},"Value climbs at every stage. So does risk, if you move before you have earned it.",[18,654,656],{"id":655},"when-to-move-up","When to move up",[11,658,659],{},"You earn the next stage with evidence, not enthusiasm. The signal is a boring one:\na track record. When the AI's output has been right often enough, for long enough,\nthat the human review has become a rubber stamp, you are paying a person to approve\nwhat the machine already gets right. That is the moment to move up, and to spend\nthe freed attention on the exceptions instead.",[11,661,662],{},"The mistake runs in both directions. Move too fast and you automate an error at\nscale. Stay too slow and you cap the value at \"a slightly faster human,\" and you\npay for a reviewer forever.",[18,664,338],{"id":337},[308,666,667,673,679],{},[29,668,669,672],{},[32,670,671],{},"Place every use case on the curve."," Most teams cannot say what stage they\nare at, which means they are drifting, not deciding.",[29,674,675,678],{},[32,676,677],{},"Define promotion criteria in advance."," What accuracy, over what period, on\nwhat metric, earns the next stage. Make it a gate, not a vibe.",[29,680,681,684],{},[32,682,683],{},"Design the exception path first."," Autopilot is only safe when the handoff to\na human on the hard cases is clean and fast.",[11,686,687],{},"The goal is not to keep a human in every loop forever. It is to move the human from\ndoing the work to governing it, one earned step at a time.",[11,689,690],{},"So for your top AI use case: what stage is it at, and what exactly would earn it\nthe next one?",{"title":105,"searchDepth":106,"depth":106,"links":692},[693,694,695],{"id":622,"depth":109,"text":623},{"id":655,"depth":109,"text":656},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fcopilot-to-autopilot",{"title":614,"description":619},"copilot-to-autopilot","writing\u002Fcopilot-to-autopilot",[123,369,232],"8HHsVfTRqRMkFN6W2Dh4vaBBALMW1BZGVpIi3pVXncY",{"id":704,"title":705,"body":706,"date":362,"description":710,"draft":276,"excerpt":115,"extension":116,"meta":787,"navigation":114,"path":788,"seo":789,"slug":790,"status":115,"stem":791,"tags":792,"__hash__":794},"writing\u002Fwriting\u002Ffind-the-three-workflows.md","Find the Three Workflows",{"type":8,"value":707,"toc":782},[708,711,715,718,736,739,743,746,750,776,779],[11,709,710],{},"Most AI strategies fail by being too ambitious. Forty initiatives, a committee, a\nplatform, and no clear first win. The teams that succeed do the opposite. They\nfind three workflows where AI moves a number the business already cares about, and\nthey do those, well, before anything else. Here is how to find your three.",[18,712,714],{"id":713},"the-filter","The filter",[11,716,717],{},"A workflow is worth doing when it scores on three things at once.",[26,719,720,725,730],{},[29,721,722,724],{},[32,723,542],{}," It happens a lot. Small improvements compound only when the base is large.",[29,726,727,729],{},[32,728,560],{}," It is manual, slow, or error-prone today, so the improvement is felt.",[29,731,732,735],{},[32,733,734],{},"Boundedness."," The task is defined enough that a machine can do it reliably,\nwith a clear right answer or a clear place to stop and ask.",[11,737,738],{},"Miss any one and the case gets shaky. High volume but unbounded is a reliability\nnightmare. Bounded and painful but rare is not worth the build. You want all three.",[18,740,742],{"id":741},"the-number-test","The number test",[11,744,745],{},"Then apply the number test: can you name the outcome in a unit the CFO recognizes?\nHours saved, dollars, error rate, cycle time, revenue. If the best you can say is\n\"efficiency\" or \"better experience,\" it is not one of your three. Not yet.",[18,747,749],{"id":748},"the-method","The method",[308,751,752,758,764,770],{},[29,753,754,757],{},[32,755,756],{},"List your highest-volume operational processes."," Start from where the people\nand the tickets are, not from where the excitement is.",[29,759,760,763],{},[32,761,762],{},"Score each on volume, pain, and boundedness."," Be honest about boundedness.\nIt is the one teams consistently overrate.",[29,765,766,769],{},[32,767,768],{},"Attach a number to the top few."," If you cannot, dig until you can, or drop it.",[29,771,772,775],{},[32,773,774],{},"Pick three."," Not thirty. Three you can ship and prove within two quarters.",[11,777,778],{},"The point of three is focus. Three shipped, proven workflows build the\ncredibility, the platform, and the appetite for the next ten. Forty parallel\npilots build a graveyard.",[11,780,781],{},"Strategy is choosing. So which three workflows would you bet the whole AI program\non, and why not start there?",{"title":105,"searchDepth":106,"depth":106,"links":783},[784,785,786],{"id":713,"depth":109,"text":714},{"id":741,"depth":109,"text":742},{"id":748,"depth":109,"text":749},{},"\u002Fwriting\u002Ffind-the-three-workflows",{"title":705,"description":710},"find-the-three-workflows","writing\u002Ffind-the-three-workflows",[608,232,793],"method","vW31VvOZsp4RYA25kkzCbwfJ9L7llYiJcIuEnhiKRfQ",{"id":796,"title":797,"body":798,"date":362,"description":802,"draft":276,"excerpt":115,"extension":116,"meta":853,"navigation":114,"path":854,"seo":855,"slug":856,"status":115,"stem":857,"tags":858,"__hash__":860},"writing\u002Fwriting\u002Ffive-questions-before-the-ai-budget.md","The Five Questions Your Board Should Ask Before the Next AI Budget",{"type":8,"value":799,"toc":849},[800,803,807,839,843,846],[11,801,802],{},"Most AI budgets get approved on a mix of faith, fear of missing out, and an\nimpressive demo. A year later, nobody can quite say what happened to the money. The\nfix is not more oversight. It is five sharp questions, asked before the budget is\napproved, not after it is spent. Any CEO should welcome them, and any program that\ncannot answer them is not ready.",[18,804,806],{"id":805},"the-five-questions","The five questions",[308,808,809,815,821,827,833],{},[29,810,811,814],{},[32,812,813],{},"What number does this move, and in what unit?"," If the answer is \"efficiency\"\nor \"innovation,\" it is not a business case. Demand dollars, hours, risk, revenue,\nor cycle time. A number you did not name is a number you cannot defend.",[29,816,817,820],{},[32,818,819],{},"Are we building an asset we own, or renting a capability everyone rents?"," If\nthe advantage comes from a general model on a generic workflow, every competitor\nbuys the same thing. Ask what here is ours, the data, the workflow, the judgment,\nthat a rival cannot buy tomorrow.",[29,822,823,826],{},[32,824,825],{},"What is the plan for year two?"," AI value compounds in the second year, through\nthe platform, the data flywheel, and the eval loop. A program funded for a quarter\nand expected to pay off for a decade is mis-designed. Ask what makes this compound,\nnot just launch.",[29,828,829,832],{},[32,830,831],{},"Who owns it when the pilot ends?"," Most AI dies in the gap between a successful\npilot and a production owner. Ask which team runs this, on which platform, with\nwhat budget, after the excitement fades. If nobody owns year two, there will not be one.",[29,834,835,838],{},[32,836,837],{},"What could go wrong, and who is accountable?"," Governance, data, model error,\nvendor lock-in. Ask what the failure modes are and who carries the risk. An AI\nprogram without a named owner of its downside is an incident waiting to be scheduled.",[18,840,842],{"id":841},"why-these-five","Why these five",[11,844,845],{},"Each question kills a specific way AI programs waste money: unnamed outcomes, rented\nadvantage, quarter-thinking, orphaned pilots, and unowned risk. A program that\nanswers all five is not guaranteed to succeed, but it is honest, and it is fundable.\nA program that cannot answer them will absorb budget and produce slides.",[11,847,848],{},"So before the next AI budget lands on the table: which of these five can your\nprogram not yet answer?",{"title":105,"searchDepth":106,"depth":106,"links":850},[851,852],{"id":805,"depth":109,"text":806},{"id":841,"depth":109,"text":842},{},"\u002Fwriting\u002Ffive-questions-before-the-ai-budget",{"title":797,"description":802},"five-questions-before-the-ai-budget","writing\u002Ffive-questions-before-the-ai-budget",[125,859,283],"governance","K05WYGfiuRJf-0uva1wWPD-VWyDoBLAlCwfshd8oiUQ",{"id":862,"title":863,"body":864,"date":362,"description":868,"draft":276,"excerpt":115,"extension":116,"meta":921,"navigation":114,"path":922,"seo":923,"slug":924,"status":115,"stem":925,"tags":926,"__hash__":927},"writing\u002Fwriting\u002Fsell-the-workflow-not-the-model.md","Sell the Workflow, Not the Model",{"type":8,"value":865,"toc":915},[866,869,873,876,880,883,887,890,892,912],[11,867,868],{},"There are two kinds of AI startup. One sells access to intelligence: a thin\nwrapper that passes your request to a model and passes the answer back. The other\nsells a workflow: it owns an entire job end to end, and the model is just one part\ninside. The first gets commoditized the day the next model ships. The second keeps\nthe value. If you are building, sell the workflow, not the model.",[18,870,872],{"id":871},"why-the-wrapper-loses","Why the wrapper loses",[11,874,875],{},"A wrapper's advantage is the model, and the model is not yours. When a better or\ncheaper one arrives, your differentiation evaporates, and your customer can switch\nor rebuild it in a weekend. You are renting your entire value proposition from a\nvendor whose price and roadmap you do not control. That is not a startup. It is a\nfeature waiting to be absorbed.",[18,877,879],{"id":878},"why-the-workflow-wins","Why the workflow wins",[11,881,882],{},"A workflow product owns the parts a model cannot: the integrations into the\ncustomer's systems, the domain-specific steps, the data it accumulates, the\nguarantees on the outcome, the handling of the ugly edge cases. The model can\nimprove or be swapped underneath, and the product only gets better. The customer is\nnot buying intelligence, they are buying a job done. That is defensible, and it\ncompounds, because every customer's usage makes the workflow smarter.",[18,884,886],{"id":885},"the-test","The test",[11,888,889],{},"Ask one question: if the model I use became free and perfect tomorrow, do I still\nhave a business? If the answer is no, you are selling the model. If the answer is\nyes, because you own the workflow, the data and the outcome, you have something.\nBuild toward yes.",[18,891,338],{"id":337},[308,893,894,900,906],{},[29,895,896,899],{},[32,897,898],{},"Own the end-to-end job, not a step."," Wrap the boring integration and edge\ncases the model will not do.",[29,901,902,905],{},[32,903,904],{},"Accumulate proprietary data as a byproduct."," Every run should make your\nproduct harder to copy.",[29,907,908,911],{},[32,909,910],{},"Sell the outcome, priced on value, not tokens."," Customers buy results, not\ninference.",[11,913,914],{},"The model is a commodity input, and it is getting cheaper. The workflow around it\nis the product. So if your model became free tomorrow, would you still have a company?",{"title":105,"searchDepth":106,"depth":106,"links":916},[917,918,919,920],{"id":871,"depth":109,"text":872},{"id":878,"depth":109,"text":879},{"id":885,"depth":109,"text":886},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fsell-the-workflow-not-the-model",{"title":863,"description":868},"sell-the-workflow-not-the-model","writing\u002Fsell-the-workflow-not-the-model",[518,123,283],"eQiAIBSsiQPWbcMjHtgGr_oPEF8-mlXn7-l-6owxXnk",{"id":929,"title":930,"body":931,"date":362,"description":935,"draft":276,"excerpt":115,"extension":116,"meta":989,"navigation":114,"path":990,"seo":991,"slug":992,"status":115,"stem":993,"tags":994,"__hash__":996},"writing\u002Fwriting\u002Fsmall-models-big-deployments.md","Small Models, Big Deployments",{"type":8,"value":932,"toc":984},[933,936,940,943,946,950,953,956,958,978,981],[11,934,935],{},"There is a reflex in most AI projects: reach for the biggest, most capable model\navailable, and route everything through it. It feels safe. It is often the wrong\ncall. A large share of real enterprise work does not need frontier intelligence.\nIt needs a competent model that is small, fast, cheap, and running somewhere you\ncontrol.",[18,937,939],{"id":938},"the-case-nobody-makes-in-the-demo","The case nobody makes in the demo",[11,941,942],{},"Demos are built to impress, so they use the largest model. Production is built to\nsurvive, and production has different priorities: latency, cost per call, data\nresidency, and reliability under load. On every one of those, a well-chosen small\nmodel, sometimes running on your own hardware or even on the device, quietly wins.",[11,944,945],{},"Classifying a document, extracting fields, routing a ticket, drafting a\ntemplated reply, checking a form: these are high-volume, bounded tasks. A small\nmodel fine-tuned on your data does them faster and cheaper than a giant\ngeneral-purpose model, and it does not send your sensitive data to someone\nelse's cloud to do it.",[18,947,949],{"id":948},"why-this-matters-more-now","Why this matters more now",[11,951,952],{},"Two forces make small models a serious strategy, not a compromise. Small models\nhave become genuinely capable, closing much of the gap on narrow tasks. And the\ncost and privacy pressures of running everything through a frontier API have\nbecome real line items on the P&L and real questions from your risk committee.",[11,954,955],{},"The mature architecture is not \"one big model for everything.\" It is a portfolio:\na small model for the routine high-volume work, a large model reserved for the\ngenuinely hard reasoning, and a router that sends each request to the cheapest\nmodel that can do the job.",[18,957,338],{"id":337},[308,959,960,966,972],{},[29,961,962,965],{},[32,963,964],{},"Audit your traffic."," Most requests are routine. Measure what share truly\nneeds frontier reasoning. It is usually smaller than the team assumes.",[29,967,968,971],{},[32,969,970],{},"Default small, escalate to large."," Make the big model the exception you\nreach for, not the default you pay for on every call.",[29,973,974,977],{},[32,975,976],{},"Treat privacy as architecture."," For sensitive data, a model you run is not\na nice-to-have. It is the difference between a yes and a no from compliance.",[11,979,980],{},"Bigger is not a strategy. It is a default, and defaults are where money and trust\nquietly leak.",[11,982,983],{},"So before your next model decision: how much of your workload is paying frontier\nprices for routine work?",{"title":105,"searchDepth":106,"depth":106,"links":985},[986,987,988],{"id":938,"depth":109,"text":939},{"id":948,"depth":109,"text":949},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fsmall-models-big-deployments",{"title":930,"description":935},"small-models-big-deployments","writing\u002Fsmall-models-big-deployments",[123,452,995],"cost","7pN_-FqeRav1zu11LQj0Ya4WnteHNdZfpd3zR-Dquj0",{"id":998,"title":999,"body":1000,"date":362,"description":1004,"draft":276,"excerpt":115,"extension":116,"meta":1085,"navigation":114,"path":1086,"seo":1087,"slug":1088,"status":115,"stem":1089,"tags":1090,"__hash__":1092},"writing\u002Fwriting\u002Ften-ai-startups-at-25.md","Ten AI Startups I Would Build If I Were 25",{"type":8,"value":1001,"toc":1081},[1002,1005,1009,1071,1075,1078],[11,1003,1004],{},"People collect AI startup ideas like trading cards. Ideas are cheap. What is\nscarce is a wedge, a specific reason a small team can win against incumbents and\nagainst general models. So here are ten I would build if I were twenty-five again,\neach chosen for a real wedge, and biased toward Bharat, where the problems are\nlarge and the tools are finally good enough.",[18,1006,1008],{"id":1007},"ten-i-would-build","Ten I would build",[308,1010,1011,1017,1023,1029,1035,1041,1047,1053,1059,1065],{},[29,1012,1013,1016],{},[32,1014,1015],{},"Vernacular voice agents for the frontline."," Millions of workers and\ncustomers speak rather than type, in a dozen languages. Voice-first agents for\nfield sales, support and public services, in Indian languages, are a wedge\nglobal players will not prioritize.",[29,1018,1019,1022],{},[32,1020,1021],{},"Compliance back office for regulated SMEs."," GST, filings, audits: painful,\nrule-bound, high volume. A vertical agent that reads, checks and files is dull\nand durable.",[29,1024,1025,1028],{},[32,1026,1027],{},"Agentic leverage for solo legal and accounting practices."," Not to replace\nthem, to give a two-person firm the reach of twenty.",[29,1030,1031,1034],{},[32,1032,1033],{},"Diagnostic vision for under-served clinics."," Screening from images where\nspecialists are scarce, built with the humility and traceability healthcare demands.",[29,1036,1037,1040],{},[32,1038,1039],{},"An AI layer over informal credit."," Underwriting the thin-file borrower from\nalternative, consented data. The data moat is real and local.",[29,1042,1043,1046],{},[32,1044,1045],{},"Agri intelligence per plot, not per region."," Sowing, disease and market\ntiming specific to a farmer's field, delivered by voice.",[29,1048,1049,1052],{},[32,1050,1051],{},"A workflow-owning agent for one boring vertical"," (freight, claims, logistics\ndocumentation). Sell the outcome, not the model.",[29,1054,1055,1058],{},[32,1056,1057],{},"On-device AI for privacy-sensitive work."," Where data cannot leave the\nbuilding, a small-model product wins by default.",[29,1060,1061,1064],{},[32,1062,1063],{},"An eval and reliability platform for AI teams."," When everyone builds agents,\nthe picks-and-shovels play is helping them trust the agents.",[29,1066,1067,1070],{},[32,1068,1069],{},"Talent infrastructure for the AI era."," Mapping real skills to real work, and\nclosing the gap, which happens to be what I am building with GuildTrek.",[18,1072,1074],{"id":1073},"the-pattern-behind-the-list","The pattern behind the list",[11,1076,1077],{},"Notice what these share. Each owns a workflow, not a model. Each sits on\nproprietary or local data a general model cannot reach. Each solves a painful,\nhigh-volume problem for a specific user. None of them is \"a better chatbot.\" That\nis the wedge. Ideas are free. Wedges are earned.",[11,1079,1080],{},"So of these ten, which one is sitting in your industry, waiting for someone who\nactually understands the domain to build it?",{"title":105,"searchDepth":106,"depth":106,"links":1082},[1083,1084],{"id":1007,"depth":109,"text":1008},{"id":1073,"depth":109,"text":1074},{},"\u002Fwriting\u002Ften-ai-startups-at-25",{"title":999,"description":1004},"ten-ai-startups-at-25","writing\u002Ften-ai-startups-at-25",[518,123,1091],"ideas","y7Hn6R-aoX13AeZZ8WTflSLWsOeK1Z93lvtupca-8Ko",{"id":1094,"title":1095,"body":1096,"date":362,"description":1100,"draft":276,"excerpt":115,"extension":116,"meta":1182,"navigation":114,"path":1183,"seo":1184,"slug":1185,"status":115,"stem":1186,"tags":1187,"__hash__":1190},"writing\u002Fwriting\u002Fthe-ai-native-org-chart.md","The AI-Native Org Chart",{"type":8,"value":1097,"toc":1176},[1098,1101,1105,1108,1112,1115,1141,1144,1148,1151,1153,1173],[11,1099,1100],{},"Most AI transformations fail not at the model but at the org chart. You cannot bolt\nAI onto a structure designed entirely around human labor and expect AI-native\nresults. But the answer is not a dramatic reorganization either. Those consume a year\nand produce mostly anxiety. The change that works is small, specific, and structural.\nHere it is.",[18,1102,1104],{"id":1103},"why-the-current-org-resists-ai","Why the current org resists AI",[11,1106,1107],{},"A traditional org is built around people doing tasks, grouped into functions. AI cuts\nacross that. It needs a shared platform, shared data, and shared evals that no single\nfunction owns. Left to the default structure, every function builds its own AI in\nisolation, re-solving the same problems badly, and nothing reaches production. The org\nis not hostile to AI. It is simply shaped for a different kind of work.",[18,1109,1111],{"id":1110},"the-four-roles-that-make-it-work","The four roles that make it work",[11,1113,1114],{},"You do not need to blow up the org. You need to add and connect four things.",[308,1116,1117,1123,1129,1135],{},[29,1118,1119,1122],{},[32,1120,1121],{},"A small central AI platform team."," It owns the shared substrate: the model\ngateway, the eval harness, observability, and governance tooling. Small, senior, and\nproduct-minded, not a committee.",[29,1124,1125,1128],{},[32,1126,1127],{},"Embedded AI engineers in the business lines."," They build the actual use cases,\nclose to the domain, consuming the platform instead of rebuilding it.",[29,1130,1131,1134],{},[32,1132,1133],{},"Domain experts paired with engineers, not parked on an advisory board."," The\nperson who knows the workflow sits with the person who builds it. That pairing is\nwhere good use cases come from.",[29,1136,1137,1140],{},[32,1138,1139],{},"A governance function with real teeth."," It owns risk, data, and the standards,\nand it can say no. Not a blocker, a guardrail that lets everyone else move faster,\nsafely.",[11,1142,1143],{},"That is the whole structure. A central platform, embedded builders, paired experts,\nand real governance. It fits inside your existing org without a reorg circus.",[18,1145,1147],{"id":1146},"what-does-not-work","What does not work",[11,1149,1150],{},"Two failure patterns are worth naming. A Center of Excellence that owns no product is a\nbudget line, not a structure, it produces frameworks and no shipped value. And fully\ndecentralized AI, where every team does its own thing, produces forty half-built\nsystems and no platform. The winning shape is central substrate, distributed building.",[18,1152,338],{"id":337},[308,1154,1155,1161,1167],{},[29,1156,1157,1160],{},[32,1158,1159],{},"Stand up a small platform team before the use cases, not after."," The substrate\nis what turns pilots into production.",[29,1162,1163,1166],{},[32,1164,1165],{},"Embed builders in the business, do not centralize all AI work."," Value is built\nclose to the domain.",[29,1168,1169,1172],{},[32,1170,1171],{},"Pair experts with engineers, and give governance real authority."," Those two\nmoves quietly decide whether any of this compounds.",[11,1174,1175],{},"You do not need to redraw the whole org chart. You need to add four boxes and connect\nthem well. So in your company: does the AI platform exist as a team, or is every\nfunction quietly rebuilding the same thing alone?",{"title":105,"searchDepth":106,"depth":106,"links":1177},[1178,1179,1180,1181],{"id":1103,"depth":109,"text":1104},{"id":1110,"depth":109,"text":1111},{"id":1146,"depth":109,"text":1147},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fthe-ai-native-org-chart",{"title":1095,"description":1100},"the-ai-native-org-chart","writing\u002Fthe-ai-native-org-chart",[125,1188,1189],"organization","transformation","leMn8VZHSmZ2r7ZTM_D2h1YXruddXhghfvzozSY7GxM",{"id":1192,"title":1193,"body":1194,"date":362,"description":1198,"draft":276,"excerpt":115,"extension":116,"meta":1252,"navigation":114,"path":1253,"seo":1254,"slug":1255,"status":115,"stem":1256,"tags":1257,"__hash__":1259},"writing\u002Fwriting\u002Fthe-boring-standard.md","The Boring Standard That Beats the Next Model",{"type":8,"value":1195,"toc":1247},[1196,1199,1203,1206,1209,1213,1216,1219,1221,1241,1244],[11,1197,1198],{},"Every few weeks a new model launches and the industry holds its breath.\nMeanwhile, the thing that actually decides whether AI works inside your company\nis boring, unglamorous plumbing: the standard way a model connects to your tools,\nyour data, and your systems. That plumbing just grew up, and it matters more than\nthe next benchmark.",[18,1200,1202],{"id":1201},"the-real-bottleneck-was-never-the-model","The real bottleneck was never the model",[11,1204,1205],{},"For most enterprises, the model was never the constraint. Integration was. Every\nAI feature re-solved, from scratch, how to let a model read a database, call an\ninternal API, or use a tool, each with its own bespoke, brittle glue. You did not\nhave one AI capability. You had forty half-built connectors.",[11,1207,1208],{},"Standard protocols for tool use (Model Context Protocol and its cousins) change\nthat. They give models a common, reusable way to discover and call tools and data\nsources. Build the connector once, and every model and every agent can use it.\nThat is the shift the USB port and the API each brought in their turn: not\nexciting, but the thing that made everything after it possible.",[18,1210,1212],{"id":1211},"why-this-beats-the-next-model","Why this beats the next model",[11,1214,1215],{},"A better model buys you a few points on a benchmark. A standard integration layer\nbuys you leverage on everything. Every workflow reuses the same tools, every new\nmodel plugs into the same estate, and you stop rewriting plumbing each time the\nfrontier moves. The compounding value is in the connective tissue, not the model\nof the month.",[11,1217,1218],{},"It also protects you from lock-in. If your tools speak a standard, swapping the\nmodel underneath becomes a config change, not a rebuild. That is real negotiating\npower with your vendor.",[18,1220,338],{"id":337},[308,1222,1223,1229,1235],{},[29,1224,1225,1228],{},[32,1226,1227],{},"Stop building bespoke connectors."," Adopt a standard tool-use layer so\nintegrations are built once and reused everywhere.",[29,1230,1231,1234],{},[32,1232,1233],{},"Treat your tools and data as a product."," A clean, well-described tool\nsurface is now a strategic asset, because every agent you deploy will use it.",[29,1236,1237,1240],{},[32,1238,1239],{},"Judge platforms on interoperability, not just model quality."," The question\nis not only \"how smart is the model,\" it is \"how easily does it plug into my\nestate, and how easily can I replace it.\"",[11,1242,1243],{},"Watch the model launches if you enjoy them. But fund the plumbing. The next model\nwill be old news in a month. The integration layer you build will still be paying\noff in three years.",[11,1245,1246],{},"So how many times has your team re-solved the same connection between a model and\nyour own systems?",{"title":105,"searchDepth":106,"depth":106,"links":1248},[1249,1250,1251],{"id":1201,"depth":109,"text":1202},{"id":1211,"depth":109,"text":1212},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fthe-boring-standard",{"title":1193,"description":1198},"the-boring-standard","writing\u002Fthe-boring-standard",[123,452,1258],"integration","8R3Dl_tD_EJixULJkva61IkhfTO3Bu2ngGdZ7PTLctE",{"id":1261,"title":1262,"body":1263,"date":362,"description":1267,"draft":276,"excerpt":115,"extension":116,"meta":1320,"navigation":114,"path":1321,"seo":1322,"slug":1323,"status":115,"stem":1324,"tags":1325,"__hash__":1326},"writing\u002Fwriting\u002Fthe-boring-startup-that-prints-money.md","The Boring Startup That Prints Money",{"type":8,"value":1264,"toc":1314},[1265,1268,1272,1275,1279,1282,1286,1289,1291,1311],[11,1266,1267],{},"Founders chase the exciting startup: the consumer app, the frontier model, the\ndemo that goes viral. The startup that quietly prints money is usually the\nopposite: a services business, in a dull vertical, disguised as a product, where AI\ndoes most of the work behind the scenes. Boring is a feature.",[18,1269,1271],{"id":1270},"the-shape-of-it","The shape of it",[11,1273,1274],{},"Pick an unglamorous industry with expensive manual work: bookkeeping for small\nfirms, insurance claims processing, medical billing, legal document review, freight\npaperwork. Today it is done by people, slowly. You wrap the whole job as a product,\nand you use AI to do the bulk of the labor under the hood. The customer sees an\noutcome and a price. You keep the margin between what they pay for the result and\nwhat it costs you to produce it with AI doing most of the work.",[18,1276,1278],{"id":1277},"why-it-works","Why it works",[11,1280,1281],{},"Three reasons this beats the flashy play. The pain is proven, so you do not have to\ncreate demand, it already exists as a line item. The buyer is unglamorous and\nunderserved, so incumbents are weak and general tools do not bother. And AI\ngenuinely changes the unit economics, so a job that used to need ten people now\nneeds two plus software, and that gap is your margin. It will not trend on social\nmedia. It will pay salaries.",[18,1283,1285],{"id":1284},"the-honest-caveat","The honest caveat",[11,1287,1288],{},"This is not passive. Owning the outcome means owning the edge cases, the errors, and\nthe customer relationship. It is a real operating business, not a magic API. But\nthat operating difficulty is also the moat. It is exactly what a wrapper cannot be\nbothered to do, and exactly what keeps your margin from being competed away.",[18,1290,338],{"id":337},[308,1292,1293,1299,1305],{},[29,1294,1295,1298],{},[32,1296,1297],{},"Find a dull vertical with expensive manual work."," The more boring, the less\ncompetition.",[29,1300,1301,1304],{},[32,1302,1303],{},"Sell the outcome, run AI underneath."," The customer buys the result, not the technology.",[29,1306,1307,1310],{},[32,1308,1309],{},"Keep the human on the last mile."," Own the edge cases. That is where trust and\nmargin both live.",[11,1312,1313],{},"The best AI business you can start this year is probably one you would be slightly\nembarrassed to pitch at a party. So which boring, expensive, manual job would you\nturn into a product?",{"title":105,"searchDepth":106,"depth":106,"links":1315},[1316,1317,1318,1319],{"id":1270,"depth":109,"text":1271},{"id":1277,"depth":109,"text":1278},{"id":1284,"depth":109,"text":1285},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fthe-boring-startup-that-prints-money",{"title":1262,"description":1267},"the-boring-startup-that-prints-money","writing\u002Fthe-boring-startup-that-prints-money",[518,123,283],"R0O6pE0n2kMVCXopxSYQ3sT6-TmjwMN5KdeC_5t9GF4",{"id":1328,"title":1329,"body":1330,"date":362,"description":1334,"draft":276,"excerpt":115,"extension":116,"meta":1388,"navigation":114,"path":1389,"seo":1390,"slug":1391,"status":115,"stem":1392,"tags":1393,"__hash__":1395},"writing\u002Fwriting\u002Fthe-year-inference-got-cheap.md","The Year Inference Got Cheap",{"type":8,"value":1331,"toc":1383},[1332,1335,1339,1342,1345,1349,1352,1355,1357,1377,1380],[11,1333,1334],{},"Quietly, the most important number in AI has been falling off a cliff. The cost\nof a unit of machine reasoning, a token, has dropped by more than an order of\nmagnitude in a short span, and it is still falling. Reasoning models that think\nbefore they answer arrived at the same time. Thinking, in other words, got both\nbetter and cheaper at once. Most enterprise AI strategies were written when it\nwas still expensive, and they have quietly gone out of date.",[18,1336,1338],{"id":1337},"why-this-changes-the-plan","Why this changes the plan",[11,1340,1341],{},"When intelligence is expensive, you ration it. You use AI only on the few\nhigh-value calls, you keep prompts short, you avoid letting a model loop. Every\none of those instincts becomes wrong when the price collapses.",[11,1343,1344],{},"Cheap inference means you can afford to let a system think longer, check its own\nwork, try several approaches, and call a model many times inside a single task.\nThe behaviours that were too expensive last year, verification, self-correction,\nmulti-step reasoning, are now just line items. The question shifts from \"can we\nafford to use AI here\" to \"why are we not using it everywhere it removes toil.\"",[18,1346,1348],{"id":1347},"the-trap-on-the-other-side","The trap on the other side",[11,1350,1351],{},"Cheap does not mean free, and cheap per call does not mean cheap at scale. When a\nsingle user action quietly triggers fifty model calls, your bill can grow faster\nthan your value. Falling prices reward volume, and volume without discipline is\nhow AI programs blow their budget while looking productive.",[11,1353,1354],{},"So the discipline is not rationing anymore. It is instrumentation. Know your cost\nper outcome, not your cost per token, and watch it like a unit economic.",[18,1356,338],{"id":337},[308,1358,1359,1365,1371],{},[29,1360,1361,1364],{},[32,1362,1363],{},"Revisit the use cases you killed for being too expensive."," Many are now\nviable. Your no-list from last year is your opportunity list this year.",[29,1366,1367,1370],{},[32,1368,1369],{},"Design for thinking, not just answering."," Let systems verify and retry.\nThe extra calls are cheap, the wrong answer is not.",[29,1372,1373,1376],{},[32,1374,1375],{},"Track cost per outcome."," Falling unit prices hide rising totals. Measure\nthe thing the business actually pays for.",[11,1378,1379],{},"The teams that win are not the ones who spent the least on tokens. They are the\nones who noticed the price moved, and rebuilt the plan around it.",[11,1381,1382],{},"So before your next AI roadmap: which of your assumptions were priced at last\nyear's cost of thinking?",{"title":105,"searchDepth":106,"depth":106,"links":1384},[1385,1386,1387],{"id":1337,"depth":109,"text":1338},{"id":1347,"depth":109,"text":1348},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fthe-year-inference-got-cheap",{"title":1329,"description":1334},"the-year-inference-got-cheap","writing\u002Fthe-year-inference-got-cheap",[123,1394,283],"economics","eV4zupH4Q7_O-gzI6bXgUrK3WHJNV1CesdSMRatoQ9c",{"id":1397,"title":1398,"body":1399,"date":362,"description":1403,"draft":276,"excerpt":115,"extension":116,"meta":1454,"navigation":114,"path":1455,"seo":1456,"slug":1457,"status":115,"stem":1458,"tags":1459,"__hash__":1460},"writing\u002Fwriting\u002Fuse-case-embarrassed-by.md","The Best AI Use Case Is the One You Are Embarrassed By",{"type":8,"value":1400,"toc":1449},[1401,1404,1408,1411,1414,1418,1421,1423,1443,1446],[11,1402,1403],{},"Ask a leadership team for their flagship AI use case and you will hear something\nimpressive: a customer-facing assistant, a strategy copilot, a moonshot. Ask\nwhich AI use case actually paid for itself, and if they are honest, it is\nsomething they are almost embarrassed to mention. Reconciling invoices. Reading\nclaims. Tagging tickets. The boring stuff is where the money is.",[18,1405,1407],{"id":1406},"why-the-boring-wins","Why the boring wins",[11,1409,1410],{},"The value of an AI use case is roughly volume times friction times repeatability.\nThe flashy use cases are usually low volume, high stakes, and full of edge cases,\nwhich is exactly where AI is hardest to trust. The unglamorous ones are the\nopposite: enormous volume, painful manual friction, and a task bounded enough\nthat a machine can do it reliably. That is the sweet spot.",[11,1412,1413],{},"A back-office process run ten thousand times a day, each time costing a person a\nfew minutes of tedium, is a bigger prize than a dazzling assistant used twice a\nweek. It just does not demo well.",[18,1415,1417],{"id":1416},"the-reason-it-gets-skipped","The reason it gets skipped",[11,1419,1420],{},"It gets skipped for a human reason, not a technical one. Nobody gets promoted for\nautomating the invoice queue. The prestige sits on the shiny front-end. So the\nhighest-ROI work stays unfunded while the demo-friendly work absorbs the budget.\nThat is a governance failure dressed up as a roadmap.",[18,1422,338],{"id":337},[308,1424,1425,1431,1437],{},[29,1426,1427,1430],{},[32,1428,1429],{},"Follow the toil, not the glamour."," List the tasks people do the most and\nenjoy the least. Your best use case is near the top.",[29,1432,1433,1436],{},[32,1434,1435],{},"Score by volume times friction, not by visibility."," Pick the one that moves\na real number, even if it never makes a slide.",[29,1438,1439,1442],{},[32,1440,1441],{},"Ship it quietly, then show the result."," A boring workflow in production\nteaches more, and saves more, than ten impressive demos.",[11,1444,1445],{},"The best AI use case in your company is probably one nobody wants to present.\nFund it anyway.",[11,1447,1448],{},"So which of your workflows is too boring to demo, and too expensive to ignore?",{"title":105,"searchDepth":106,"depth":106,"links":1450},[1451,1452,1453],{"id":1406,"depth":109,"text":1407},{"id":1416,"depth":109,"text":1417},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fuse-case-embarrassed-by",{"title":1398,"description":1403},"use-case-embarrassed-by","writing\u002Fuse-case-embarrassed-by",[608,232,610],"LiT7M9uxlxUFom_phRYW2I4y0fTt1OiZiB-7baUccLI",{"id":1462,"title":1463,"body":1464,"date":362,"description":1468,"draft":276,"excerpt":115,"extension":116,"meta":1522,"navigation":114,"path":1523,"seo":1524,"slug":1525,"status":115,"stem":1526,"tags":1527,"__hash__":1529},"writing\u002Fwriting\u002Fvertical-ai-eats-horizontal-ai.md","Vertical AI Eats Horizontal AI",{"type":8,"value":1465,"toc":1517},[1466,1469,1473,1476,1479,1483,1486,1489,1491,1511,1514],[11,1467,1468],{},"The first wave of AI products was horizontal: one assistant, for everyone, for\neverything. Impressive, and increasingly a commodity, because your competitor has\nthe same one. The durable value is moving the other way, into vertical AI:\nsystems that go deep into a single industry, carrying the data, the workflow, and\nthe judgment that only that domain has. Vertical eats horizontal.",[18,1470,1472],{"id":1471},"why-depth-wins","Why depth wins",[11,1474,1475],{},"A horizontal assistant knows a little about everything and owns nothing. A vertical\nsystem knows one domain cold: its documents, its regulations, its edge cases, the\nspecific decisions its people make all day. That depth is exactly what a general\nmodel cannot buy off the shelf, because it lives in proprietary data and hard-won\nworkflow, not in the model weights.",[11,1477,1478],{},"Depth also compounds. A vertical system that sees a domain's data and outcomes\nevery day gets sharper at that domain in a way a general tool never will. The moat\nis not the model. It is the domain loop wrapped around it.",[18,1480,1482],{"id":1481},"the-strategic-read","The strategic read",[11,1484,1485],{},"For a buyer, this reframes build vs buy. A horizontal copilot is a rent: useful,\nand identical to what everyone else rents. A vertical capability built on your\nproprietary data and workflow is an asset: harder to build, and yours. If you want\nadvantage rather than parity, you fund depth.",[11,1487,1488],{},"For a builder, it says the same thing from the other side. Do not build a thinner\nversion of a general assistant. Pick one industry, one painful workflow, and own it\nend to end. A narrow product a specialist cannot live without beats a broad product\nthey can take or leave.",[18,1490,338],{"id":337},[308,1492,1493,1499,1505],{},[29,1494,1495,1498],{},[32,1496,1497],{},"Ask what only you know."," Your proprietary data and domain workflow are the\nraw material of a moat. A general model is not.",[29,1500,1501,1504],{},[32,1502,1503],{},"Prefer deep over broad."," One workflow owned completely beats ten touched shallowly.",[29,1506,1507,1510],{},[32,1508,1509],{},"Judge AI products on domain fit, not feature lists."," The winner understands\nyour industry's edge cases, not the one with the longest capability sheet.",[11,1512,1513],{},"Horizontal AI made intelligence a utility. Vertical AI is where you turn that\nutility into an advantage.",[11,1515,1516],{},"So in your business: what does your industry know that a general model never will,\nand who is building on it, you or someone else?",{"title":105,"searchDepth":106,"depth":106,"links":1518},[1519,1520,1521],{"id":1471,"depth":109,"text":1472},{"id":1481,"depth":109,"text":1482},{"id":337,"depth":109,"text":338},{},"\u002Fwriting\u002Fvertical-ai-eats-horizontal-ai",{"title":1463,"description":1468},"vertical-ai-eats-horizontal-ai","writing\u002Fvertical-ai-eats-horizontal-ai",[123,1528,283],"product","3VUs_o04284WHkJoeoyTNJRl2ZMyCdrfG7ZO70txLa4",{"id":1531,"title":1532,"body":1533,"date":362,"description":1537,"draft":276,"excerpt":115,"extension":116,"meta":1588,"navigation":114,"path":1589,"seo":1590,"slug":1591,"status":115,"stem":1592,"tags":1593,"__hash__":1595},"writing\u002Fwriting\u002Fwhat-your-ai-vendor-wont-tell-you.md","What Your AI Vendor Will Not Tell You",{"type":8,"value":1534,"toc":1584},[1535,1538,1542,1574,1578,1581],[11,1536,1537],{},"Your AI vendor is not lying to you. They are just answering the questions you ask,\nand the important questions are the ones that do not come up in the demo. The demo\nis designed to impress. The contract, the cost curve, and the lock-in are where the\nreal story lives. Here are five things worth asking before you sign, because the\nvendor will not raise them first.",[18,1539,1541],{"id":1540},"five-things-to-ask-about","Five things to ask about",[308,1543,1544,1550,1556,1562,1568],{},[29,1545,1546,1549],{},[32,1547,1548],{},"The cost at scale, not the cost of the pilot."," The pilot is cheap by design.\nAsk for the fully loaded cost per outcome at your real volume, including the model\ncalls a single user action triggers. Pilots are priced to get you in. Production\nis priced differently.",[29,1551,1552,1555],{},[32,1553,1554],{},"What happens to your data."," Ask plainly: is our data used to train your models,\nis it isolated, can we get it back, and does your platform become the only place\nour accumulated data and evals live. If your moat ends up inside their system, that\nis the real price.",[29,1557,1558,1561],{},[32,1559,1560],{},"How hard it is to leave."," Ask what it takes to switch models or vendors in\neighteen months. If the architecture ties you to their weights, their format, or\ntheir proprietary layer, you are not buying a tool, you are buying a dependency,\nand they know it.",[29,1563,1564,1567],{},[32,1565,1566],{},"Who owns the accuracy."," When the model gets something wrong in production, whose\nproblem is it. Vendors happily own the demo and quietly disown the edge cases. Ask\nwhere their responsibility ends and yours begins, before it matters.",[29,1569,1570,1573],{},[32,1571,1572],{},"Where the price goes next."," You do not control the per-unit price of the\nintelligence you are renting. Ask what has happened to their pricing, and what\nleverage you have if it moves against you. The margin you save today can be\nrepriced away tomorrow.",[18,1575,1577],{"id":1576},"the-pattern","The pattern",[11,1579,1580],{},"Every one of these is about the same thing: who captures the value in three years.\nThe vendor's honest goal is to make themselves indispensable and your switching costs\nhigh. That is not villainy, it is strategy, and your job is to negotiate against it\nwith open eyes. Own your data, keep the model swappable, price on outcomes, and read\nthe exit clause as carefully as the feature list.",[11,1582,1583],{},"So before your next AI contract: which of these five did the vendor not bring up, and\nwhy do you think that is?",{"title":105,"searchDepth":106,"depth":106,"links":1585},[1586,1587],{"id":1540,"depth":109,"text":1541},{"id":1576,"depth":109,"text":1577},{},"\u002Fwriting\u002Fwhat-your-ai-vendor-wont-tell-you",{"title":1532,"description":1537},"what-your-ai-vendor-wont-tell-you","writing\u002Fwhat-your-ai-vendor-wont-tell-you",[125,1594,283],"procurement","sdhf0YhdP8kiUCuE1j0TLpwdMJAiB7RnQjvLruiynAc",{"id":1597,"title":1598,"body":1599,"date":1662,"description":1603,"draft":276,"excerpt":115,"extension":116,"meta":1663,"navigation":114,"path":1664,"seo":1665,"slug":1666,"status":115,"stem":1667,"tags":1668,"__hash__":1670},"writing\u002Fwriting\u002Fdata-problem-in-costume.md","Your AI Problem Is a Data Problem Wearing a Costume",{"type":8,"value":1600,"toc":1657},[1601,1604,1607,1611,1614,1617,1621,1624,1628,1631,1651,1654],[11,1602,1603],{},"Every stalled AI program I am asked to look at arrives with the same framing:\nwe need a better model. A different vendor. More fine-tuning. Almost none of\nthem have a model problem. They have a data problem wearing a model's costume.",[11,1605,1606],{},"The tell is consistent. The pilot worked beautifully on a clean, hand-picked\nslice of data. Then it met production, where the data is fragmented across a\ndozen systems, half of it is stale, the definitions disagree between\ndepartments, and nobody owns the pipeline that would keep it current. The\nmodel did not get worse. It finally saw the real inputs.",[18,1608,1610],{"id":1609},"why-the-data-layer-is-where-programs-die","Why the data layer is where programs die",[11,1612,1613],{},"A model is a function of its inputs. If the inputs are incomplete,\ninconsistent, or out of date, no amount of prompt engineering, fine-tuning, or\nvendor-switching recovers the output. You are not one model away from success.\nYou are one data platform away, and that is a much larger and much less\nglamorous project.",[11,1615,1616],{},"This is the part that never makes the board deck. \"We will stand up a modern\ndata foundation\" does not demo. \"Watch the AI answer this question\" does. So\nprograms get funded on the demo and starved on the foundation, and then\neveryone is surprised when the thing that worked in the room falls apart at\nscale.",[18,1618,1620],{"id":1619},"the-honest-sequencing","The honest sequencing",[11,1622,1623],{},"The uncomfortable truth is that data readiness is often 60 to 80 percent of\nthe real work in an enterprise AI program, and it has to come first. You\ncannot promise a model win on top of plumbing you have not built. When a\nleadership team asks me to accelerate an AI roadmap, the first question is\nrarely about models. It is this: is the data this depends on actually ready,\nowned, and current? If the answer is no, that is the roadmap, whether anyone\nwanted it to be or not.",[18,1625,1627],{"id":1626},"what-to-actually-do","What to actually do",[11,1629,1630],{},"Three moves before you approve another model-centric plan:",[308,1632,1633,1639,1645],{},[29,1634,1635,1638],{},[32,1636,1637],{},"Run a data-readiness check first."," For each use case, ask where the data\nlives, who owns it, how fresh it is, and whether the definitions are agreed.\nIf those answers are shaky, the use case is not ready, no matter how good\nthe model is.",[29,1640,1641,1644],{},[32,1642,1643],{},"Fund the foundation as its own line, not a footnote."," The data platform\nis not overhead on the AI project. For most enterprises it is the project.\nBudget it like one.",[29,1646,1647,1650],{},[32,1648,1649],{},"Sequence to where the data already is."," The fastest AI win is almost\nalways the use case sitting on data that is already clean, owned, and\ncurrent. Start there, earn the credibility, then fund the harder\nfoundations with the trust you built.",[11,1652,1653],{},"The teams that ship are not the ones with the best models. They are the ones\nthat were honest, early, about the state of their data, and did the\nunglamorous work before they promised the demo.",[11,1655,1656],{},"So before the next model debate: is this an AI problem, or a data problem\nwearing a costume?",{"title":105,"searchDepth":106,"depth":106,"links":1658},[1659,1660,1661],{"id":1609,"depth":109,"text":1610},{"id":1619,"depth":109,"text":1620},{"id":1626,"depth":109,"text":1627},"2026-07-19",{},"\u002Fwriting\u002Fdata-problem-in-costume",{"title":1598,"description":1603},"data-problem-in-costume","writing\u002Fdata-problem-in-costume",[608,1669,283],"data","mUpf1PV9hQ-OXnvSaFmPYJSCTDbKU2Z_vq3TCjyBPLc",{"id":1672,"title":1673,"body":1674,"date":1740,"description":1678,"draft":276,"excerpt":115,"extension":116,"meta":1741,"navigation":114,"path":1742,"seo":1743,"slug":1744,"status":115,"stem":1745,"tags":1746,"__hash__":1747},"writing\u002Fwriting\u002Fai-roi-treadmill.md","Your CFO Is About to Ask What the AI Budget Bought",{"type":8,"value":1675,"toc":1735},[1676,1679,1682,1686,1689,1692,1696,1699,1702,1706,1709,1729,1732],[11,1677,1678],{},"Somewhere in the next two quarters, your CFO is going to ask a simple\nquestion: what did the AI budget actually buy us? A lot of technology leaders\ndo not have a clean answer. Not because the projects failed, but because of\nhow they were funded.",[11,1680,1681],{},"Most enterprise AI is sold to the board as cost-out. Fewer hours, fewer\npeople, faster tickets. That is an easy business case to approve. It is also\nthe reason the return is so hard to point to a year later.",[18,1683,1685],{"id":1684},"the-efficiency-trap","The efficiency trap",[11,1687,1688],{},"Here is the uncomfortable part. If your AI advantage comes from a\ngeneral-purpose model applied to a generic workflow, every competitor can buy\nexactly the same advantage from exactly the same model. The productivity lift\nis real, but it is not yours. It becomes the new baseline for the whole\nindustry, and within a few quarters it competes away. You did not build a\nmoat. You joined a treadmill, and now you have to keep spending just to stay\nlevel.",[11,1690,1691],{},"Worse, the savings rarely stay with you. In a competitive market, an\nefficiency everyone shares gets passed to customers as lower prices, or gets\ncaptured upstream by the model vendor whose per-token price you do not\ncontrol. You cut cost, the customer or the vendor kept the value, and the P&L\nbarely moved.",[18,1693,1695],{"id":1694},"efficiency-is-table-stakes-advantage-is-something-else","Efficiency is table stakes. Advantage is something else.",[11,1697,1698],{},"Efficiency is the floor everyone reaches. Advantage is what is hard to copy:\nproprietary data, a workflow no competitor has, judgment encoded into the\nproduct, a feedback loop that makes your system better every week. Those\ncompound. A generic copilot does not.",[11,1700,1701],{},"The CFO question is really a strategy question wearing a finance costume.\n\"What did we buy\" is asking \"did we buy something defensible, or did we rent\nthe same tool as everyone else.\"",[18,1703,1705],{"id":1704},"what-to-fund-instead","What to fund instead",[11,1707,1708],{},"Three shifts before the next budget cycle:",[308,1710,1711,1717,1723],{},[29,1712,1713,1716],{},[32,1714,1715],{},"Name the outcome in a real unit."," Dollars, hours, risk, revenue. If the\nbusiness case says \"efficiency,\" it is not a business case. You cannot\ndefend a number you never named.",[29,1718,1719,1722],{},[32,1720,1721],{},"Fund at least one top-line or moat play, not only cost-out."," Cost-out is\nsymmetric and competes away. A capability built on your proprietary data or\na workflow only you own is asymmetric, and that is where durable return\nlives.",[29,1724,1725,1728],{},[32,1726,1727],{},"Own the compounding asset."," Rent the model, fine. But own the data, the\nevals, and the workflow around it. The margin in three years accrues to\nwhoever owns the compounding layer, not whoever rented the cleverest model\nthis quarter.",[11,1730,1731],{},"The programs that survive the CFO conversation are not the ones that saved the\nmost this year. They are the ones that can point to something a competitor\ncannot buy off the shelf.",[11,1733,1734],{},"So before the question comes: what did your AI spend buy that your competitor\ncannot buy tomorrow?",{"title":105,"searchDepth":106,"depth":106,"links":1736},[1737,1738,1739],{"id":1684,"depth":109,"text":1685},{"id":1694,"depth":109,"text":1695},{"id":1704,"depth":109,"text":1705},"2026-07-16",{},"\u002Fwriting\u002Fai-roi-treadmill",{"title":1673,"description":1678},"ai-roi-treadmill","writing\u002Fai-roi-treadmill",[608,177,283],"S6HgDnmljta_i-_FAUpFfz4e8-qDLNSR06vA5Gn1xn8",{"id":1749,"title":1750,"body":1751,"date":1845,"description":1755,"draft":276,"excerpt":115,"extension":116,"meta":1846,"navigation":114,"path":1847,"seo":1848,"slug":1849,"status":115,"stem":1850,"tags":1851,"__hash__":1852},"writing\u002Fwriting\u002Fpilot-purgatory.md","You Have 40 AI Pilots and Zero in Production",{"type":8,"value":1752,"toc":1840},[1753,1756,1762,1766,1778,1785,1788,1792,1799,1802,1805,1807,1810,1830,1837],[11,1754,1755],{},"Most large enterprises I walk into have the same dashboard: a dozen,\nsometimes forty, AI pilots. Impressive demos. Executive sponsors.\nSlideware. And almost none of it is in production, carrying real load,\nchanging a real number on the P&L.",[11,1757,1758,1759,1761],{},"The reflex is to call this an innovation problem: we need more experiments,\nbetter models, a bigger AI team. It isn't. It's an ",[32,1760,452],{}," problem,\nand until you name it as one, the next ten pilots die exactly where the last\nten did.",[18,1763,1765],{"id":1764},"why-pilots-stall","Why pilots stall",[11,1767,1768,1769,1773,1774,1777],{},"A pilot is optimized to ",[1770,1771,1772],"em",{},"demo",". Production is optimized to ",[1770,1775,1776],{},"survive",". Those\nare different engineering problems, and the gap between them is where every\nstalled program lives.",[11,1779,1780,1781,1784],{},"A demo has to work once, on the happy path, in front of an audience. A\nproduction system has to work on the unhappy paths: the malformed\ninput, the model that regresses after a version bump, the edge case that only\nshows up at ten thousand requests. The demo needs a good model. Production\nneeds evaluation, monitoring, rollback, retraining, governance, and a way to\nconsume the output ",[1770,1782,1783],{},"inside a workflow people already use",".",[11,1786,1787],{},"That second list is the unglamorous 60% of the work. It's also the 60% nobody\nfunds, because it doesn't demo.",[18,1789,1791],{"id":1790},"the-missing-layer","The missing layer",[11,1793,1794,1795,1798],{},"Across five AI builds and two Fortune 500 accounts, the pattern is identical:\npilots stall not because the models are bad, but because there's no\n",[32,1796,1797],{},"platform"," underneath them.",[11,1800,1801],{},"Each team stands up its own stack: its own prompt, its own eval (if\nany), its own deployment path, and every one of them re-solves the same\nhard problems, badly. You don't end up with one AI capability. You end up with\nforty half-built ones, none of which crosses into production, because crossing\nthat line requires infrastructure no single pilot can justify building alone.",[11,1803,1804],{},"A Center of Excellence doesn't fix this. A CoE that owns no product is a budget\nline, not a structure. What fixes it is a small platform team, five to\nten people, that owns the substrate every pilot needs: the eval harness,\nthe model gateway, observability, the deployment path, governance. Pilots stop\nrebuilding plumbing and start shipping.",[18,1806,1627],{"id":1626},[11,1808,1809],{},"If you're staring at a wall of stalled pilots, three moves:",[308,1811,1812,1818,1824],{},[29,1813,1814,1817],{},[32,1815,1816],{},"Kill most of them."," Forty pilots isn't a portfolio; it's a deferred\ndecision. Pick the two or three tied to a real, measurable outcome:\ndollars, hours, or risk, not \"efficiency\", and stop the rest today.",[29,1819,1820,1823],{},[32,1821,1822],{},"Fund the platform, not just the pilots."," The infrastructure that gets one\npilot to production gets all of them there. It's the highest-ROI line on the\nboard, and the one that never makes the slide.",[29,1825,1826,1829],{},[32,1827,1828],{},"Ship the thinnest end-to-end thing first."," One capability, working all the\nway through into a workflow people already open, not a standalone \"AI\ntool\" nobody does. A thin slice in production teaches you more than forty\ndemos.",[11,1831,1832,1833,1836],{},"The enterprises pulling ahead aren't the ones with the most pilots. They're the\nones that treated getting ",[1770,1834,1835],{},"one"," thing into production as the real work,\nand built the layer that made the second, third, and tenth almost free.",[11,1838,1839],{},"How many pilots does your organization have in flight right now, and how\nmany are carrying real load?",{"title":105,"searchDepth":106,"depth":106,"links":1841},[1842,1843,1844],{"id":1764,"depth":109,"text":1765},{"id":1790,"depth":109,"text":1791},{"id":1626,"depth":109,"text":1627},"2026-07-13",{},"\u002Fwriting\u002Fpilot-purgatory",{"title":1750,"description":1755},"pilot-purgatory","writing\u002Fpilot-purgatory",[608,232,283],"Aw6ZSk6V1tvMAaghC9dWpo7AHBxWi9r7A1ihlI0OUlo",{"id":1854,"title":1855,"body":1856,"date":2091,"description":105,"draft":276,"excerpt":115,"extension":116,"meta":2092,"navigation":114,"path":2093,"seo":2094,"slug":2095,"status":2096,"stem":2097,"tags":2098,"__hash__":2101},"writing\u002Fwriting\u002Fenterprise-ai-adoption-playbook.md","The Enterprise AI Adoption Playbook Nobody Writes (Because It's Boring)",{"type":8,"value":1857,"toc":2076},[1858,1867,1871,1881,1885,1900,1904,1909,1920,1924,1932,1936,1944,1948,1956,1960,1968,1972,1998,2001,2005,2016,2019,2023,2043,2047,2057,2060],[1859,1860,1861],"blockquote",{},[11,1862,1863,1866],{},[32,1864,1865],{},"Status:"," outline. Promote to draft after adding two real anecdotes\n(without naming clients) that anchor each section.",[18,1868,1870],{"id":1869},"why-this-post","Why this post",[11,1872,1873,1874,1877,1878,1880],{},"Most public writing on enterprise AI is about ",[1770,1875,1876],{},"capability",": what models can\ndo, what frameworks exist, what demos look like. Very little is about ",[1770,1879,232],{},",\nthe unglamorous shape of an organization that actually consumes AI well.\nAfter twenty years inside enterprise transformation programs, the shape is\nconsistent, the failures are consistent, and almost none of it is about the\nmodel.",[18,1882,1884],{"id":1883},"the-argument-in-one-paragraph","The argument in one paragraph",[11,1886,1887,1888,1891,1892,1895,1896,1899],{},"Enterprise AI succeeds when three things are present at the same time:\n",[32,1889,1890],{},"a real problem with a measurable outcome",", ",[32,1893,1894],{},"a small platform team that\nowns the AI substrate",", and ",[32,1897,1898],{},"a feedback loop tying production output back\nto model improvement",". Everything else, the model choice, the vendor\ndeal, the framework debate, is downstream and reversible. Most programs\nget the first wrong, skip the second, and never build the third. Then they\nrestart with a different model.",[18,1901,1903],{"id":1902},"the-five-layer-reality-check","The five-layer reality check",[1905,1906,1908],"h3",{"id":1907},"_1-problem-layer","1. Problem layer",[26,1910,1911,1914,1917],{},[29,1912,1913],{},"Real problem or board-deck problem?",[29,1915,1916],{},"Measurable in dollars, hours, or risk, not \"efficiency.\"",[29,1918,1919],{},"Reversible if the experiment fails.",[1905,1921,1923],{"id":1922},"_2-data-layer","2. Data layer",[26,1925,1926,1929],{},[29,1927,1928],{},"Where the program actually dies. If the data isn't ready, no model\nrecovers from it.",[29,1930,1931],{},"Don't promise model wins on top of unbuilt data plumbing.",[1905,1933,1935],{"id":1934},"_3-platform-layer","3. Platform layer",[26,1937,1938,1941],{},[29,1939,1940],{},"A small (5 to 10 person) AI platform team owning evals, model gateway,\nobservability, governance, fine-tune pipelines, and the prompt registry.",[29,1942,1943],{},"Without this, every team builds its own stack and the org has 14\nhalf-working AI projects.",[1905,1945,1947],{"id":1946},"_4-product-layer","4. Product layer",[26,1949,1950,1953],{},[29,1951,1952],{},"AI inside an existing product workflow beats a \"new AI tool.\"",[29,1954,1955],{},"The mistake: standing up a separate UX for the AI capability. Adoption\ncollapses.",[1905,1957,1959],{"id":1958},"_5-feedback-layer","5. Feedback layer",[26,1961,1962,1965],{},[29,1963,1964],{},"Production output -> eval -> next version of the model or prompt.",[29,1966,1967],{},"Skip this and the system gets quietly worse as the world changes around it.",[18,1969,1971],{"id":1970},"the-org-design-that-actually-works","The org design that actually works",[26,1973,1974,1980,1986,1992],{},[29,1975,1976,1979],{},[32,1977,1978],{},"AI platform team"," (central, small, owns the substrate)",[29,1981,1982,1985],{},[32,1983,1984],{},"Embedded AI engineers"," per product line (consume the platform)",[29,1987,1988,1991],{},[32,1989,1990],{},"Domain SMEs"," paired with engineers, not on advisory committees",[29,1993,1994,1997],{},[32,1995,1996],{},"A governance function"," with veto power and explicit risk taxonomy",[11,1999,2000],{},"What does NOT work: a Center of Excellence that owns no product. A Center\nof Excellence is a budget line, not a structure.",[18,2002,2004],{"id":2003},"the-18-month-operating-rhythm","The 18-month operating rhythm",[26,2006,2007,2010,2013],{},[29,2008,2009],{},"Months 0 to 3: pick one real problem, build the data layer, stand up the\nplatform team.",[29,2011,2012],{},"Months 3 to 9: ship the first capability inside an existing product. Build\nthe eval harness in parallel.",[29,2014,2015],{},"Months 9 to 18: second and third capabilities. Begin retiring tools the AI\ncapabilities replace.",[11,2017,2018],{},"The shape that fails: an \"AI strategy\" that's really a vendor evaluation,\nfollowed by a six-month pilot, followed by re-organizing.",[18,2020,2022],{"id":2021},"what-cxos-underestimate","What CxOs underestimate",[26,2024,2025,2031,2037],{},[29,2026,2027,2030],{},[32,2028,2029],{},"The unsexy work",", data, evals, observability, governance, is\n60% of the budget.",[29,2032,2033,2036],{},[32,2034,2035],{},"Talent dynamics",": AI engineers want to work on AI, not on data\npipelines. Architect around that or you'll lose them.",[29,2038,2039,2042],{},[32,2040,2041],{},"Compounding",": the second year of the program creates more value\nthan the first if (and only if) the platform layer exists.",[18,2044,2046],{"id":2045},"what-id-ask-a-board-to-commit-to-before-starting","What I'd ask a board to commit to before starting",[1859,2048,2049],{},[11,2050,2051],{},[1770,2052,2053],{},[2054,2055,2056],"span",{},"List of 5 to 7 commitments. To be filled.",[2058,2059],"hr",{},[11,2061,2062,2065,2066,2071,2072],{},[32,2063,2064],{},"Related:"," ",[2067,2068,2070],"a",{"href":2069},"\u002Fwork\u002Fdigihire-ai","DigiHire.ai, AI Recruitment Platform"," · ",[2067,2073,2075],{"href":2074},"\u002Fwork\u002Fenterprise-digital-twin","Enterprise Digital Twin",{"title":105,"searchDepth":106,"depth":106,"links":2077},[2078,2079,2080,2087,2088,2089,2090],{"id":1869,"depth":109,"text":1870},{"id":1883,"depth":109,"text":1884},{"id":1902,"depth":109,"text":1903,"children":2081},[2082,2083,2084,2085,2086],{"id":1907,"depth":106,"text":1908},{"id":1922,"depth":106,"text":1923},{"id":1934,"depth":106,"text":1935},{"id":1946,"depth":106,"text":1947},{"id":1958,"depth":106,"text":1959},{"id":1970,"depth":109,"text":1971},{"id":2003,"depth":109,"text":2004},{"id":2021,"depth":109,"text":2022},{"id":2045,"depth":109,"text":2046},"2026-06-25",{},"\u002Fwriting\u002Fenterprise-ai-adoption-playbook",{"title":1855,"description":105},"enterprise-ai-adoption-playbook","outline","writing\u002Fenterprise-ai-adoption-playbook",[2099,608,2100],"transform","leadership","cNVjobdAqpuQHX1czMjmJl3clL9Pugt10ghtFUZmbME",{"id":2103,"title":2104,"body":2105,"date":2283,"description":105,"draft":276,"excerpt":115,"extension":116,"meta":2284,"navigation":114,"path":2285,"seo":2286,"slug":2287,"status":2096,"stem":2288,"tags":2289,"__hash__":2291},"writing\u002Fwriting\u002Fagentic-architecture-patterns.md","Five Patterns I Keep Reaching For When Designing Agentic Systems",{"type":8,"value":2106,"toc":2272},[2107,2114,2116,2127,2129,2136,2140,2154,2158,2169,2173,2191,2195,2206,2210,2221,2225,2245,2249,2260,2262],[1859,2108,2109],{},[11,2110,2111,2113],{},[32,2112,1865],{}," outline. Promote to draft once each pattern has at least one\nconcrete code or diagram example pulled from a real build.",[18,2115,1870],{"id":1869},[11,2117,2118,2119,2122,2123,2126],{},"There's a lot of writing on what agents ",[1770,2120,2121],{},"are"," and very little on what\n",[1770,2124,2125],{},"shipping"," one looks like. I want this post to compress what I've learned\nacross the AI Interviewer, DigiHire, the no-code platform, and the digital\ntwin builds, the patterns I keep reaching for, why, and the ones that\nsounded good in theory but never made it past the second iteration.",[18,2128,1884],{"id":1883},[11,2130,2131,2132,2135],{},"Production agentic systems are not built around prompts. They're built around\n",[32,2133,2134],{},"five primitives",": a typed contract for the agent's output, a constrained\ntool surface, a deterministic orchestrator, a transparent eval harness, and a\ngraceful interruption model. The model you pick matters less than whether\nthese five exist. If they don't, scaling the prompt makes the system worse,\nnot better.",[18,2137,2139],{"id":2138},"pattern-1-output-as-schema-not-as-text","Pattern 1: Output as schema, not as text",[26,2141,2142,2145,2148,2151],{},[29,2143,2144],{},"Why: comparability, queryability, eval-ability all live in structured output.",[29,2146,2147],{},"How: schema-constrained decoding, structured outputs APIs, JSON-mode + validator + retry loop.",[29,2149,2150],{},"When NOT to: when the task is genuinely generative (writing, summarization),\nschema is harm.",[29,2152,2153],{},"Example from the AI Interviewer: rubric-aligned JSON per round, not transcripts.",[18,2155,2157],{"id":2156},"pattern-2-tools-are-the-api-surface-treat-them-like-one","Pattern 2: Tools are the API surface, treat them like one",[26,2159,2160,2163,2166],{},[29,2161,2162],{},"Why: an agent with too many tools is an agent with no tools.",[29,2164,2165],{},"How: 5 to 9 tools, named like product features, with explicit pre- and\npostconditions in their descriptions. Version them.",[29,2167,2168],{},"The mistake: exposing the database schema as 40 micro-tools.",[18,2170,2172],{"id":2171},"pattern-3-deterministic-orchestration-around-non-deterministic-reasoning","Pattern 3: Deterministic orchestration around non-deterministic reasoning",[26,2174,2175,2185,2188],{},[29,2176,2177,2178,2181,2182,1784],{},"Why: the LLM should choose ",[1770,2179,2180],{},"what"," to do, not ",[1770,2183,2184],{},"how the system runs",[29,2186,2187],{},"How: state machines, pipelines, or workflow engines on the outside;\nreasoning calls on the inside.",[29,2189,2190],{},"Reference: Pipecat-style pipelines in the AI Interviewer.",[18,2192,2194],{"id":2193},"pattern-4-eval-harness-before-scale-not-after","Pattern 4: Eval harness before scale, not after",[26,2196,2197,2200,2203],{},[29,2198,2199],{},"Why: every model swap, prompt edit, and tool change is a silent regression\nrisk. Without an eval, you find out from users.",[29,2201,2202],{},"How: a fixed eval set per agent capability, golden outputs, semantic-diff\nscoring, run-on-PR.",[29,2204,2205],{},"The honest cost: this takes longer to set up than the agent itself. Do it\nanyway.",[18,2207,2209],{"id":2208},"pattern-5-interruption-is-a-first-class-concern","Pattern 5: Interruption is a first-class concern",[26,2211,2212,2215,2218],{},[29,2213,2214],{},"Why: real users don't take turns politely. Especially in voice and chat.",[29,2216,2217],{},"How: barge-in handling, cancellable tool calls, partial-output replay,\nstate checkpoints.",[29,2219,2220],{},"Lesson from the AI Interviewer: this was the difference between \"demo\" and\n\"deploy.\"",[18,2222,2224],{"id":2223},"patterns-ive-stopped-reaching-for","Patterns I've stopped reaching for",[26,2226,2227,2233,2239],{},[29,2228,2229,2232],{},[32,2230,2231],{},"Pure ReAct loops with unlimited horizon",", impressive in benchmarks,\nunwieldy in production. Cap the steps; orchestrate the rest.",[29,2234,2235,2238],{},[32,2236,2237],{},"One mega-prompt with everything in it",", readable for a week, then a\nliability forever.",[29,2240,2241,2244],{},[32,2242,2243],{},"\"Let the agent decide\" as a UX choice",": users want predictability.\nAgents are an implementation detail of the product, not its surface.",[18,2246,2248],{"id":2247},"what-this-implies-for-teams","What this implies for teams",[26,2250,2251,2254,2257],{},[29,2252,2253],{},"Hire (or grow) a platform engineer before the second model swap.",[29,2255,2256],{},"Budget for the eval harness like you'd budget for tests, because that's\nwhat it is.",[29,2258,2259],{},"Pick a battle-tested orchestrator before you pick a clever model.",[2058,2261],{},[11,2263,2264,2065,2266,2071,2270],{},[32,2265,2064],{},[2067,2267,2269],{"href":2268},"\u002Fwork\u002Fai-interviewer","Real-time Multimodal AI Interviewer",[2067,2271,2075],{"href":2074},{"title":105,"searchDepth":106,"depth":106,"links":2273},[2274,2275,2276,2277,2278,2279,2280,2281,2282],{"id":1869,"depth":109,"text":1870},{"id":1883,"depth":109,"text":1884},{"id":2138,"depth":109,"text":2139},{"id":2156,"depth":109,"text":2157},{"id":2171,"depth":109,"text":2172},{"id":2193,"depth":109,"text":2194},{"id":2208,"depth":109,"text":2209},{"id":2223,"depth":109,"text":2224},{"id":2247,"depth":109,"text":2248},"2026-06-15",{},"\u002Fwriting\u002Fagentic-architecture-patterns",{"title":2104,"description":105},"agentic-architecture-patterns","writing\u002Fagentic-architecture-patterns",[2290,369,452],"build","gjjNI-1ljA8b9Dkc3JYgCLuC2_xXZGKeITIatMhk270",{"name":2293,"role":2294,"tagline":2295,"location":2296,"contact":2297},"Sankar Vema","AI Builder & Architect of Agentic Systems","I help enterprise leaders turn AI ambition into capability that actually ships.","India · open to global advisory engagements",{"email":2298,"linkedin":2299,"github":2300,"blog":2301,"twitter":2302},"sankar.vema@gmail.com","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Fsankarvema\u002F","http:\u002F\u002Fsankarvema.github.io\u002F","http:\u002F\u002Fsankarvema.blogspot.com\u002F","https:\u002F\u002Ftwitter.com\u002Fsansvema",1790601541553]