[{"data":1,"prerenderedAt":309},["ShallowReactive",2],{"locale-alternates:\u002Fai-tools-are-easier-to-build-transformation-still-takes-work":3,"post-\u002Fai-tools-are-easier-to-build-transformation-still-takes-work":8},{"path":4,"alternates":5},"\u002Fai-tools-are-easier-to-build-transformation-still-takes-work",{"en":4,"tr":6,"de":7},"\u002Ftr\u002Fyapay-zeka-donusumu-arac-gelistirmekten-fazlasi","\u002Fde\u002Fki-tools-lassen-sich-leichter-bauen-transformation-braucht-weiter-arbeit",{"page":9,"translations":170,"nav":176,"related":299,"random":300},{"id":10,"title":11,"body":12,"categories":140,"category":143,"changeHistory":143,"date":144,"description":145,"disclosures":146,"draft":149,"extension":150,"firstLiveAt":143,"image":151,"imageAlt":152,"kind":153,"lang":154,"meta":155,"navigation":156,"omitGermanLocalizationDisclosure":149,"path":4,"publishedAt":143,"readingTime":157,"rights":143,"seo":158,"seoTitle":159,"slug":160,"sources":143,"stem":160,"tags":161,"translationKey":167,"type":168,"updated":143,"__hash__":169},"posts\u002Fai-tools-are-easier-to-build-transformation-still-takes-work.md","Building an AI Tool Is Easy. Transforming a Company Is Not.",{"type":13,"value":14,"toc":132},"minimark",[15,19,32,35,40,43,46,49,52,55,62,66,69,72,75,78,81,85,88,97,105,108,111,115,123,126,129],[16,17,18],"p",{},"Having many software systems does not mean a company has eliminated repetitive work. Alongside its core systems, spreadsheets, email threads, and small solutions built by individual employees continue to carry work forward. AI makes it faster to create new tools for those jobs. It does not change the way the company works at the same speed.",[16,20,21,22,31],{},"In ",[23,24,30],"a",{"href":25,"rel":26,"target":29},"https:\u002F\u002Fwww.ben-evans.com\u002Fbenedictevans\u002F2026\u002F9\u002F3\u002Fai-tools-and-transformation",[27,28],"nofollow","noopener","_blank","“AI, tools and transformation”",", Benedict Evans examines that gap. Making a tool easier to build does not make it easier to identify the right problem, find a workable solution, or get other people to use it. Once a solution created for one person begins to affect other teams and their decisions, further responsibilities appear. Someone has to decide who maintains it, which data can be trusted, and who corrects an error.",[16,33,34],{},"That argument resonated with work I have done myself. When we tried to use AI to analyse conversations between prospective customers and the sales team, a task that initially looked simple turned out to depend on far more information and judgment.",[36,37,39],"h2",{"id":38},"reading-the-conversation-was-not-enough","Reading the conversation was not enough",[16,41,42],{},"At first, the task seemed clear: give the conversation transcripts to an AI model and ask it to evaluate them against a set of criteria. We soon found that assessing a conversation properly required information beyond the transcript. We needed data from several sources.",[16,44,45],{},"The data in the CRM system we use to manage customer relationships was not ready for that analysis either. Its records were structured to help an employee form a quick view of a customer or prospect. That arrangement was useful for its original purpose, but it became a problem when we wanted AI to perform detailed analysis.",[16,47,48],{},"We cannot expect a model to supply on its own the meaning that an experienced person fills in while looking at a record. The data we give it needs to make sufficiently clear what each piece of information represents. On that measure, our data was messy. The project made the difference visible between a record that supports daily work and data that is suitable for analysis.",[16,50,51],{},"We also found data that had not been updated for some time and was no longer in active use. Some of it might be useful again for the analysis AI now makes possible. First, though, we needed to understand how current and reliable it was. Data remaining in a system does not mean it can be used directly in an analysis.",[16,53,54],{},"We began by planning to examine conversation transcripts. Before long, we were reconsidering which information we needed in order to understand the customer and the conditions of the conversation.",[16,56,57],{},[58,59],"img",{"alt":60,"src":61},"Neat-looking CRM records become usable evaluation criteria only after context and exceptions are examined.","\u002Fimages\u002Finline-ai-sales-conversation-analysis.webp",[36,63,65],{"id":64},"there-is-no-ready-made-definition-of-a-good-conversation","There is no ready-made definition of a good conversation",[16,67,68],{},"Bringing the data together was not enough on its own. An evaluation based on general criteria did not always match the way managers viewed the same conversation. They took account of details and exceptions that the model did not see by default.",[16,70,71],{},"We therefore revisited the evaluation criteria and prompts repeatedly. We had to state more clearly what mattered in each situation and which details changed the conclusion.",[16,73,74],{},"From the outside, this may look like the task of writing better instructions for a model. But before we can improve an instruction, we need to explain the reasoning behind our own decision. Why do we consider one conversation successful? Why do we assess the same behaviour differently in another? An exception that is clear to people inside a company remains invisible to a model until it is made explicit.",[16,76,77],{},"Treating every manager’s assessment as unquestionably correct would not solve the problem either. We need to understand the source of the difference: did we give the model incomplete information, leave the criterion vague, or need to reconsider our own assessment?",[16,79,80],{},"The experience showed me how much operational knowledge is embedded in everyday judgment. Building the analysis tool meant bringing that knowledge into the open as well. The time-consuming part was making explicit what the company assumed it already knew.",[36,82,84],{"id":83},"we-also-decide-where-the-system-should-stop","We also decide where the system should stop",[16,86,87],{},"Some boundaries cannot be set simply by looking at the data or the model. Healthcare is a direct example.",[16,89,90,91,96],{},"A chatbot can help collect information from a patient. Give that same system the task of making a diagnosis, and the responsibility it takes on changes. For example, the ",[23,92,95],{"href":93,"rel":94,"target":29},"https:\u002F\u002Feur-lex.europa.eu\u002Feli\u002Freg\u002F2017\u002F745\u002Foj\u002Feng",[27,28],"EU Medical Device Regulation"," also covers software intended for particular medical purposes. The intended purpose matters. Software used in a healthcare organisation does not automatically become a medical device.",[16,98,99,100,104],{},"I discussed this distinction in more detail in my Turkish-language article on ",[23,101,103],{"href":102},"\u002Fhealthcare-ai-chatbot-triage-medical-device","the difference between chatbots, triage, and medical devices",".",[16,106,107],{},"We chose to leave diagnostic decisions to doctors. The limits set by the industry and regulation were part of that choice. Protecting the doctor’s decision was also the outcome we wanted from our own way of working. A model’s ability to produce an answer is not sufficient reason to delegate that decision to it.",[16,109,110],{},"Designing a system therefore includes deciding where it stops. Collecting information, preparing it for assessment, and making a medical decision may happen within the same conversation, but they carry different responsibilities.",[36,112,114],{"id":113},"a-personal-project-is-different-from-the-companys-work","A personal project is different from the company’s work",[16,116,117,118,122],{},"In personal projects, I have seen a more direct benefit from AI. When I ",[23,119,121],{"href":120},"\u002Fai-turned-a-refactor-i-wouldnt-do-into-a-one-hour-job","simplified PeşinTaksit’s infrastructure",", I completed in about an hour a task for which I normally would not have made time. It was my project: I knew the need and could make the change myself.",[16,124,125],{},"With sales-conversation analysis, the definition of the work, the data, and the evaluation criteria all had to be considered together. Applying the short development time from one case to the other would leave part of the necessary work out of the calculation.",[16,127,128],{},"That is why an AI project’s timeline or budget needs to include time for preparing data and reviewing the work with people who understand it. These are not unnecessary delays before the tool can be used. They are part of what makes the tool usable.",[16,130,131],{},"A working prototype shows that a project is moving. Whether a company is ready to make decisions on the basis of that prototype can only be learned by testing the data and criteria together.",{"title":133,"searchDepth":134,"depth":134,"links":135},"",2,[136,137,138,139],{"id":38,"depth":134,"text":39},{"id":64,"depth":134,"text":65},{"id":83,"depth":134,"text":84},{"id":113,"depth":134,"text":114},[141,142],"ai","business",null,"2026-09-28","What a sales-conversation analysis project revealed about the data, judgment, and responsibility required to turn an AI tool into a usable business process.",{"aiUse":147,"aiNote":148},"ai-assisted","The approach, personal experiences, and assessments in this article belong to Evren Bal. AI supported source checking, structuring, and draft writing.",false,"md","\u002Fimages\u002Fhero\u002Fai-tools-organizational-transformation.webp","Imperfect data sources flow from a personal prototype into a shared, accountable workflow.","Field note","en",{},true,5,{"title":11,"description":145},"AI Transformation: From Building Tools to Changing How Work Gets Done","ai-tools-are-easier-to-build-transformation-still-takes-work",[162,163,164,165,166],"AI transformation","business processes","data quality","decision criteria","enterprise AI","ai-tools-and-organizational-transformation","post","YenawkMu3X9ZEUKxFhog196K7uoJI2--TE0or0zst9E",{"en":171,"tr":172,"de":174},{"path":4,"title":11},{"path":6,"title":173},"Yapay Zekâ Aracı Kurmak Kolay, Şirketi Dönüştürmek Değil",{"path":7,"title":175},"Ein KI-Tool zu bauen ist leicht. Ein Unternehmen zu verändern nicht.",{"prev":177,"next":143,"others":180,"lucky":298,"readingTime":157},{"path":178,"title":179},"\u002Fthe-cost-of-a-hello-enterprise-ai","The Cost of a Hello: Managing Enterprise AI Use",[181,184,187,190,193,196,199,202,205,208,211,214,217,220,223,226,229,232,235,238,241,244,247,250,253,256,259,262,265,268,271,274,277,280,281,284,287,289,292,295],{"path":182,"title":183},"\u002Fkeeping-customers-happy-isnt-enough","Keeping Customers Happy Isn’t Enough. You Have to Follow Up.",{"path":185,"title":186},"\u002Fai-made-code-cheap-verification-is-still-expensive","AI Made Code Cheap. Verification Is Still Expensive.",{"path":188,"title":189},"\u002Fdoes-enterprise-ai-really-need-fine-tuning","Does Enterprise AI Really Need Fine-Tuning?",{"path":191,"title":192},"\u002Fai-visibility-is-also-a-state-concern","AI Visibility Is Also on States’ Agendas",{"path":194,"title":195},"\u002Fwhy-we-trust-ai-judgments","Why We Trust AI Judgments and What That Fails to Prove",{"path":197,"title":198},"\u002Fwhere-should-people-be-in-the-loop-while-ai-does-the-work","Where Should People Be in the Loop While AI Does the Work?",{"path":200,"title":201},"\u002Fai-assisted-migrations-deferred-work","AI Can Make Deferred Work Worth Doing",{"path":203,"title":204},"\u002Fis-buying-backlinks-worth-it","Would We Still Buy This Backlink If It Sent No Ranking Signal?",{"path":206,"title":207},"\u002Fmanaging-technology-and-transforming-the-business-are-not-the-same","Managing Technology and Transforming the Business Are Not the Same Thing",{"path":209,"title":210},"\u002Fllms-txt-was-never-the-point","llms.txt Was Never the Point",{"path":212,"title":213},"\u002Fcomprehension-debt-the-bill-comes-due-alone","Comprehension Debt: The Bill Comes Due Alone",{"path":215,"title":216},"\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":218,"title":219},"\u002Fwhat-ai-can-automate-in-cro-research","What Can AI Actually Automate in CRO Research?",{"path":221,"title":222},"\u002Fhow-can-companies-create-value-from-ai","How Can Companies Create Value From AI?",{"path":224,"title":225},"\u002Fthe-job-ai-wont-take-and-the-five-it-prevents","AI Is Reducing Hiring Without Layoffs",{"path":227,"title":228},"\u002Fvoice-agent-system-deployment-constraints","What to Test Before Investing in a Voice Agent",{"path":230,"title":231},"\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":233,"title":234},"\u002Fgoogle-generative-ai-data-ai-citation-timing","AI Visibility Dropped Before Search: The Google Data That Changed My Theory",{"path":236,"title":237},"\u002Fone-victory-several-defeats","One Victory, Several Defeats",{"path":239,"title":240},"\u002Fturkey-national-ai-platform-evren","National AI Infrastructure Is More Than a GPU Count",{"path":242,"title":243},"\u002Feu-ai-act-after-risk-classification","EU AI Act: What Should You Do Once You Know the Risk Level?",{"path":245,"title":246},"\u002Fstart-with-the-business-problem-not-the-ai-model","Start With the Business Problem, Not the AI Model",{"path":248,"title":249},"\u002Fwhen-is-one-ai-model-enough","When Is One AI Model Enough?",{"path":251,"title":252},"\u002Fthe-same-ai-does-not-create-the-same-competitive-advantage","The Same AI Does Not Create the Same Competitive Advantage",{"path":254,"title":255},"\u002Fwhen-ai-changes-work-training-employees-is-not-enough","When AI Changes Work, Training Employees Is Not Enough",{"path":257,"title":258},"\u002Fopen-publishing-ai-monthly-seo-reporting-experiment","Open Publishing and an AI-Assisted Monthly SEO Reporting Experiment",{"path":260,"title":261},"\u002Fwho-captures-ai-productivity-gains","Who Captures the Productivity Gains from AI?",{"path":263,"title":264},"\u002Fwriting-with-ai-means-thinking-with-your-archive","Writing with AI Means Thinking About Your Archive, Too",{"path":266,"title":267},"\u002Fwhat-does-the-eu-ai-act-actually-regulate","What Is the EU AI Act, and What Does It Regulate?",{"path":269,"title":270},"\u002Fai-will-spread-across-turkiyes-businesses","AI in Türkiye: What Adoption Figures Leave Out",{"path":272,"title":273},"\u002Fdo-we-need-ai","Do We Need AI, or Are We Solving the Wrong Problem?",{"path":275,"title":276},"\u002Fwhat-makes-an-ai-system-specific-to-your-business","What Makes an AI System Specific to Your Business?",{"path":278,"title":279},"\u002Fwhen-does-enterprise-ai-need-rag","When Does Enterprise AI Actually Need RAG?",{"path":178,"title":179},{"path":282,"title":283},"\u002Fai-smaller-teams-business-capacity","AI Makes Smaller Teams Possible. What Does the Business Gain — and Lose?",{"path":285,"title":286},"\u002Fdo-you-know-how-dependent-your-company-is-on-ai","Do You Know How Dependent Your Company Is on AI?",{"path":102,"title":288},"AI in Healthcare: The Difference Between Chatbots, Triage, and Medical Devices",{"path":290,"title":291},"\u002Fai-value-is-not-always-more-work","AI's Value Is Not Always More Work",{"path":293,"title":294},"\u002Fwhen-process-automation-actually-needs-ai","When Does Process Automation Actually Need AI?",{"path":296,"title":297},"\u002Fthe-threshold-collapsed-to-zero","The Threshold Collapsed: What ProductLog Taught Me About Building in Public",{"path":269,"title":270},[],[301,303,305],{"path":296,"title":297,"date":302},"2026-06-28",{"path":278,"title":279,"date":304},"2026-08-29",{"path":306,"title":307,"date":308},"\u002Fbeyond-the-bot-lessons-from-building-a-chat-system-for-global-patients","What We Learned Building a Healthcare Chatbot for International Patients","2025-07-13",1790550152382]