[{"data":1,"prerenderedAt":342},["ShallowReactive",2],{"post-\u002Fdo-you-know-how-dependent-your-company-is-on-ai":3},{"page":4,"translation":263,"nav":265,"related":332,"random":335},{"id":5,"title":6,"body":7,"categories":236,"category":240,"date":241,"description":242,"draft":243,"extension":244,"image":245,"kind":246,"lang":247,"meta":248,"navigation":249,"path":250,"publishedAt":240,"readingTime":251,"seo":252,"slug":253,"stem":253,"tags":254,"translationKey":260,"type":261,"updated":240,"__hash__":262},"posts\u002Fdo-you-know-how-dependent-your-company-is-on-ai.md","Do You Know How Dependent Your Company Is on AI?",{"type":8,"value":9,"toc":225},"minimark",[10,49,52,55,58,61,64,69,72,75,78,81,90,94,97,100,143,146,149,153,156,159,162,167,175,181,184,187,191,194,197,200,204,207,220],[11,12,13,21],"blockquote",{},[14,15,16,17],"p",{},"💡 ",[18,19,20],"strong",{},"TL;DR: Key Takeaways",[22,23,24,31,37,43],"ul",{},[25,26,27,30],"li",{},[18,28,29],{},"AI sometimes does not merely accelerate work; it makes previously uneconomic work possible."," That is a new operational capability, not just a productivity gain.",[25,32,33,36],{},[18,34,35],{},"The central risk is not that AI must become more expensive."," It is attaching critical quality, capacity, access, and unit economics to inputs the company does not control.",[25,38,39,42],{},[18,40,41],{},"Changing an API is not the same as preserving a business outcome."," Models differ in behaviour, tool use, structured output, and edge cases; a replacement needs evaluation.",[25,44,45,48],{},[18,46,47],{},"The answer is not to slow adoption."," A temporary trade-off for speed can be reasonable in an MVP; testing alternatives, moving simple work to cheaper options, and exercising the exit path must be part of that speed.",[14,50,51],{},"My company is going through a serious AI transformation. In systems I help steer, AI is not treated as a small feature bolted onto existing software. We are building new systems around AI from the start for oversight and evaluation work that people cannot perform individually; the software is the layer that integrates, directs, and governs those systems.",[14,53,54],{},"The more important change is this: cognitive work that would be impractical to perform with human labour at the same scale can now be done at a quality that is close enough to human quality, even if it does not fully reach it. Creating content, sorting and classifying it, improving it, reviewing it, evaluating it, summarising it, enriching information, directing work to the right queue, and prioritising it can all now run faster and at a scale that was not previously possible.",[14,56,57],{},"Take a manager with a team of thirty or forty people. Imagine that they regularly read, classify, quality-check, and prioritise every piece of work the team produces. Under normal circumstances, doing that personally would be extremely difficult and perhaps impossible. At best, they sample, review intermittently, delegate, or perhaps do none of it.",[14,59,60],{},"An AI system can perform much more of that control continuously and very quickly. Its output does not have to be perfect human judgement. Two people need not reach the same conclusion when they classify, evaluate, or interpret something. It is entirely natural that one person finds an AI judgement sound while another disagrees. What matters is not that a model returns one absolute answer in every case, but that we are confident about where it is consistent and sufficient. If it produces an output that is close enough to human quality across a volume no person could sustain, it takes our management capacity to an entirely different level.",[14,62,63],{},"That is why calling this merely a productivity gain understates what is happening.",[65,66,68],"h2",{"id":67},"not-every-kind-of-ai-work-creates-the-same-dependency","Not every kind of AI work creates the same dependency",[14,70,71],{},"In the first kind of work, AI accelerates something people were already doing. If a person writes a draft, AI can write it too. If access to AI disappears, the work slows down, becomes more expensive, or returns to people; but the activity continues as it did before.",[14,73,74],{},"In the second kind, AI creates a capacity that did not previously exist: continuously evaluating every output from forty people, enriching every customer interaction, or classifying each case immediately and sending it to the right queue. These are not activities that can simply return to an older process, because they were not being done before.",[14,76,77],{},"That distinction changes the nature of the dependency.",[14,79,80],{},"At first, this new capacity may be only a useful addition. Over time, a manager starts operating with greater control capacity. Quality expectations, SLAs, production volumes, customer promises, and even staffing levels are reset around the assumption that this capacity exists. If the AI disappears, the company is not merely twenty percent slower. A part of its operating model has disappeared.",[14,82,83,84,89],{},"I wrote recently about how quickly AI expands the set of ideas a company can economically test in ",[85,86,88],"a",{"href":87},"\u002Fthe-ai-productivity-baseline-is-moving-faster-than-we-remember","The AI Productivity Baseline Is Moving Faster Than We Remember",". The companion question is: once that capacity is embedded in the way the company works, how reliably can it be carried through a change in model, provider, price, or access?",[65,91,93],{"id":92},"subscription-economics-are-not-production-economics","Subscription economics are not production economics",[14,95,96],{},"One concrete observation that made me think about this was hearing that some systems use AI companies’ subscription products rather than direct API access. Not every such use violates terms of service, but many appear to operate in grey—or potentially more dangerous—territory. I could not find reliable measurement showing how common this is.",[14,98,99],{},"But the observation exposes a broader risk. If a workflow rests on an economics that depends on violating terms of service, its real production cost may never have been tested.",[14,101,102,103,114,115,121,122,128,129,135,136,142],{},"Providers do not define subscriptions and API use as the same product surface. ",[85,104,113],{"href":105,"className":106,"rel":108,"target":112},"https:\u002F\u002Fsupport.anthropic.com\u002Fen\u002Farticles\u002F9876003-i-subscribe-to-a-paid-claude-ai-plan-why-do-i-have-to-pay-separately-for-api-usage-on-console",[107],"dofollow",[109,110,111],"nofollow","noopener","noreferrer","_blank","Anthropic"," says that a Claude subscription does not include API usage, and its ",[85,116,120],{"href":117,"className":118,"rel":119,"target":112},"https:\u002F\u002Fwww.anthropic.com\u002Flegal\u002Fconsumer-terms",[107],[109,110,111],"consumer terms"," prohibit access by bots, scripts, or other automated means except through an Anthropic API key. ",[85,123,127],{"href":124,"className":125,"rel":126,"target":112},"https:\u002F\u002Fopenai.com\u002Fpolicies\u002Fterms-of-use\u002F",[107],[109,110,111],"OpenAI’s individual terms"," prohibit automatically or programmatically extracting output, while its ",[85,130,134],{"href":131,"className":132,"rel":133,"target":112},"https:\u002F\u002Fopenai.com\u002Fpolicies\u002Fservices-agreement\u002F",[107],[109,110,111],"business agreement"," treats API-powered customer applications as a separate surface. Google’s rules differ in scope, but ",[85,137,141],{"href":138,"className":139,"rel":140,"target":112},"https:\u002F\u002Fai.google.dev\u002Fgemini-api\u002Fdocs\u002Fpricing",[107],[109,110,111],"Gemini API pricing"," similarly describes separate, usage-priced tiers for production applications.",[14,144,145],{},"These sources do not support a blanket claim that every automation using a subscription is prohibited. The narrower conclusion is enough: access designed around a person using a chat interface is not a safe assumption for a production line with predictable capacity and cost.",[14,147,148],{},"There is no need to claim that subscriptions are necessarily subsidised; providers do not disclose their marginal cost. The problem is that a flat per-seat price and a token- or usage-based price are not measuring the same thing. If providers become more restrictive in their policies, usage limits, or enforcement of terms, AI workflows that look almost free today can become far more expensive.",[65,150,152],{"id":151},"falling-inference-prices-do-not-make-the-dependency-disappear","Falling inference prices do not make the dependency disappear",[14,154,155],{},"A serious counterargument deserves full weight: inference costs may continue to fall. Hardware advances, model efficiency, and competition may make today’s unit cost a poor long-term reference point. This essay does not claim that AI must become expensive.",[14,157,158],{},"Nor does every workflow need a frontier model. Smaller models may be good enough for classification, extraction, initial drafts, structured transformations, or selected quality checks. Caching, batching, tighter context, model routing, lower quality thresholds on lower-stakes work, fine-tuning, and open-weight models can lower cost further. In some cases, local inference can become a genuine second option.",[14,160,161],{},"The risk is therefore not “a provider increases prices fivefold and the transformation collapses.” The better question is:",[11,163,164],{},[14,165,166],{},"How much of the company’s critical operation depends on an input whose price, rate limits, model behaviour, product packaging, and availability it does not determine?",[65,168,170,174],{"id":169},"generateprompt-does-not-create-portability",[171,172,173],"code",{},"generate(prompt)"," does not create portability",[14,176,177,178,180],{},"Putting several providers behind one abstraction is useful. It can speed up a response to an outage or a change in cost. But simply replacing the model behind ",[171,179,173],{}," does not prove that the workflow will produce the same business result.",[14,182,183],{},"Models differ in instruction following, reasoning, hallucination tendencies, context handling, structured output, tool calling, latency, throughput, and edge-case behaviour. The same instruction can produce reliable JSON on one model and an explanatory paragraph on another; the right routing decision in one case and a quietly wrong priority in another.",[14,185,186],{},"API portability and behavioural portability are not the same thing. This becomes more important when AI created the second kind of capacity. If there is no credible manual alternative, migration time is direct operational capacity loss.",[65,188,190],{"id":189},"a-speed-trade-off-should-not-become-a-permanent-assumption","A speed trade-off should not become a permanent assumption",[14,192,193],{},"I am not arguing that violating terms of service is acceptable. But an authorised temporary solution can be a rational trade-off when testing one workflow as an MVP, getting AI into the work sooner, or moving towards an AI-centred way of operating in two months rather than twelve. Early speed has real value.",[14,195,196],{},"That choice must not become an assumption that the system can run this way indefinitely. Before subscription limits tighten, the access method changes, or the actual unit cost becomes visible, simple work should move to simpler models, open-weight alternatives, or lower-cost options. Caching, batching, tighter context, and model routing are not optimisations to postpone until “someday.”",[14,198,199],{},"The mistake to avoid is to write the name of a second provider into a document and think that creates resilience. An alternative is only a fallback if it has been exercised regularly against real work and known acceptance criteria.",[65,201,203],{"id":202},"do-not-use-less-ai-attach-it-more-deliberately","Do not use less AI. Attach it more deliberately.",[14,205,206],{},"This is not an argument to slow AI adoption. On the contrary: because I have seen that AI does not merely reduce cost but changes the boundary of what a company can do, declining to experiment aggressively may be the larger competitive risk.",[14,208,209,210,214,215,219],{},"AI may make code cheaper, but ",[85,211,213],{"href":212},"\u002Fai-made-code-cheap-verification-is-still-expensive","verification and ownership are still not cheap",". It can also change an organisation through ",[85,216,218],{"href":217},"\u002Fthe-job-ai-wont-take-and-the-five-it-prevents","the jobs it prevents from being created",", not only through the jobs it replaces.",[14,221,222],{},[18,223,224],{},"The point is not to avoid risk. It is to know the duration, boundary, and exit path of the risk you take for speed. Otherwise, the company that moved early on AI can become the company that is suddenly left without it.",{"title":226,"searchDepth":227,"depth":227,"links":228},"",2,[229,230,231,232,234,235],{"id":67,"depth":227,"text":68},{"id":92,"depth":227,"text":93},{"id":151,"depth":227,"text":152},{"id":169,"depth":227,"text":233},"generate(prompt) does not create portability",{"id":189,"depth":227,"text":190},{"id":202,"depth":227,"text":203},[237,238,239],"ai","business","engineering",null,"2026-08-22","AI is not only accelerating work people already did. It is creating operational capacity companies could never afford to have — and that capacity needs a continuity plan before it becomes structural.",false,"md","\u002Fimages\u002Fhero\u002Fai-operational-resilience-v2.avif","Essay","en",{},true,"\u002Fdo-you-know-how-dependent-your-company-is-on-ai",8,{"title":6,"description":242},"do-you-know-how-dependent-your-company-is-on-ai",[255,256,257,258,259],"ai-transformation","ai-economics","model-portability","business-continuity","operational-resilience","ai-operational-infrastructure-resilience","post","P3G5obftMq30upIft-rmNL51xvvppKUeYbxz9OxuWSk",{"path":264},"\u002Ftr\u002Fsirketinizin-ai-bagimliliginin-farkinda-misiniz",{"prev":240,"next":266,"others":268,"lucky":331,"readingTime":251},{"path":87,"title":267},"AI Wasn’t Always This Good. We Just Got Used to It.",[269,270,272,275,278,281,284,287,289,292,295,298,301,304,307,310,313,316,319,322,325,328],{"path":87,"title":267},{"path":212,"title":271},"AI Made Code Cheap. Verification Is Still Expensive.",{"path":273,"title":274},"\u002Fthe-threshold-collapsed-to-zero","The Threshold Collapsed to Zero: Building in Public in the Age of AI Cloning",{"path":276,"title":277},"\u002Fbuild-in-public-2-0","Build in Public 2.0: The New Playbook for Sharing Openly in the Age of AI Cloning",{"path":279,"title":280},"\u002Fgoogle-generative-ai-data-ai-citation-timing","AI Citations Run on a Separate Clock: What Google's Generative AI Data Revealed",{"path":282,"title":283},"\u002Fthe-era-of-the-previous-vibe-coder-begins","The Era of the \"Previous Vibe Coder\" Begins: The Invisibility of Clean Code and the Technical Debt Bill of AI",{"path":285,"title":286},"\u002Fone-victory-several-defeats","One Victory, Several Defeats",{"path":217,"title":288},"The Job AI Won't Take — and the Five It Quietly Prevents",{"path":290,"title":291},"\u002Fproductlog-the-platform-i-built-for-myself-first","ProductLog: The Platform I Built for Myself First",{"path":293,"title":294},"\u002Fcomprehension-debt-the-bill-comes-due-alone","Comprehension Debt: The Bill Comes Due Alone",{"path":296,"title":297},"\u002Fthe-new-evrenbal-com-a-new-era-and-blogging-philosophy","The New Evrenbal.com: A New Era and Blogging Philosophy",{"path":299,"title":300},"\u002Fai-visibility-illusion-bing-citation-share","The AI Visibility Illusion: What Bing's Citation Share Data Actually Reveals",{"path":302,"title":303},"\u002F1m-impressions-per-month-0-revenue-a-programmatic-seo-post-mortem","1M Impressions per Month, $0 Revenue: A Programmatic SEO Post-Mortem",{"path":305,"title":306},"\u002Fthe-end-of-coding-or-a-new-renaissance-the-invisible-crisis-of-ai","The End of Coding or a New Renaissance? The Invisible Crisis of AI",{"path":308,"title":309},"\u002Fraising-children-in-the-age-of-artificial-intelligence","Raising Children in the Age of Artificial Intelligence",{"path":311,"title":312},"\u002Fredar-ai-powered-summaries-for-kap-disclosures-and-open-sources","Redar: AI-Powered Summaries for KAP Disclosures and Open Sources",{"path":314,"title":315},"\u002Fpesintaksit-cash-vs-installments-a-turkish-inflation-aware-payment-comparison-tool","Cash or Installments? – The Story Behind PeşinTaksit",{"path":317,"title":318},"\u002Fbeyond-the-bot-lessons-from-building-a-chat-system-for-global-patients","Beyond the bot: lessons from building a chat system for global patients",{"path":320,"title":321},"\u002Fhow-i-built-a-modern-infrastructure-using-open-source-tools-and-the-power-of-cloudflare","How I Built a Modern Infrastructure Using Open Source Tools and the Power of Cloudflare",{"path":323,"title":324},"\u002Ffrom-rules-to-decisions-the-real-time-sales-intelligence-platform-we-built-at-vanity","From Rules to Decisions: The Real-Time Sales Intelligence Platform We Built at Vanity",{"path":326,"title":327},"\u002Fwhy-im-building-rankextension-making-google-search-console-actually-make-sense","Why I’m Building RankExtension: Making Google Search Console Actually Make Sense",{"path":329,"title":330},"\u002Fan-seo-experiment-in-a-low-competition-serp-with-google-maps-and-openai","An SEO Experiment in a Low-Competition SERP with Google Maps and OpenAI",{"path":290,"title":291},[333],{"path":87,"title":267,"date":334},"2026-08-20",[336,338,340],{"path":326,"title":327,"date":337},"2025-06-05",{"path":302,"title":303,"date":339},"2026-06-21",{"path":311,"title":312,"date":341},"2025-10-30",1787362131411]