Most companies do not fail to measure Generative AI ROI. They measure the wrong thing and mistake the result for the real number. Since 2023, the dominant reflex has been to count output efficiency: documents generated, emails drafted, minutes saved per task. These metrics are real, but they are insufficient. Did decision quality improve? Did error costs fall? Did the employee’s attention capacity shift toward higher-value work? Without answers to those questions, any ROI figure is a polished estimate wearing the clothes of an analysis.Consider a mid-sized insurance brokerage in Izmir — 312 employees, agencies across seven cities. The firm deployed a RAG-based policy query system in early 2024. After six months, they had a clean finding: representative access time to policy information dropped from 4.5 minutes to 48 seconds. A solid operational number. But when the systems manager presented it to the CFO, three questions came back immediately: Did customer satisfaction improve? Did error rates fall? What are representatives doing with the four minutes they recovered? Two of the three had no answer. No baseline had been established before go-live; the destination of the recovered time had never been tracked. Excellent operational data failed to convert into a management decision — not because the technology underperformed, but because the measurement architecture was never built.Baseline discipline remains the weakest link in AI ROI calculations, and the reason is organizational rather than technical. Someone must ask, before any system goes live: ‘How do we currently run this process, and how do we measure it?’ In most Turkish SMEs, no one asks that question because measurement culture has not taken hold. An Ankara-based construction materials distributor — 445 employees, domestic B2B sales — added an LLM-assisted proposal generation engine to its tendering process in mid-2025. Before the project started, there was no record of how long proposals took; the informal estimate was ‘around two hours.’ After go-live, the question ‘how much did we improve?’ had no defensible answer. The outcome was likely positive. It was unmeasurable. The lesson is not comforting but it is clear: if measurement infrastructure is not in place before the AI budget is approved, the ROI analysis is over before it begins.Agent-based systems have made value attribution considerably harder. The agentic AI deployments gaining enterprise traction in 2025 and 2026 typically enter the middle of a human workflow — automating part of it, accelerating another part, and returning a third part to human judgment. In that hybrid model, the question ‘what exactly did this agent contribute?’ requires an attribution model designed into the system from the start. A Bursa automotive parts exporter — 378 employees, direct sales into Germany and Poland — deployed an agent system for trade documentation in late 2025. From day one, the team logged documents processed, human correction rate, and rejection events. Four months in, the correction rate had fallen by seventeen percentage points from the initial measurement. That figure captures both system learning and genuine operational value. The critical difference: attribution architecture was designed before deployment. ‘Human step, agent step, error or correction’ was coded into the workflow log. Deploying an agent and then asking how much it earned is like opening a factory and trying to install the electricity meter six months later.Portfolio framing has moved from optional to necessary. Enterprise Generative AI no longer lives in isolated pilots — it operates as a system family: multiple agents, multiple LLM connections, multiple affected workflows. In that complexity, single-project ROI calculations become misleading. If one agent consumes another agent’s output, whose contribution counts? Among the financial services and large retail teams in Turkey where this problem is most acute, the pattern that holds up looks like this: manage the AI portfolio in two distinct buckets. Bucket one — efficiency agents — carry documented time and error savings, weekly tracking, and a clear ROI threshold. Bucket two — decision support systems — carry a value horizon of twelve to eighteen months, tolerate short-term ambiguity, and are evaluated on ‘change in decision quality,’ not quarterly cost savings. Applying identical metrics to both buckets simultaneously undervalues the first and prematurely kills the second. The EU AI Act’s high-risk system classification reinforces exactly this division: decision support systems carry different governance obligations and therefore require a different measurement framework. Turkish companies selling into EU markets are already navigating this compliance layer alongside the ROI question.What should you do on Monday morning? First: for every AI project already running or planned in the next quarter, attach a mandatory baseline requirement — before the system goes live, the current process must have a two-page answer to ‘how does this work today, and how do we measure it?’ Second: for any agent system, build the value attribution mechanism into the technical design with your engineering team before deployment — ‘human step,’ ‘agent step,’ and ‘error/correction’ distinctions must be in the event log, not retrofitted afterward. Third: divide your AI projects into two buckets, assign each a separate measurement rhythm and success threshold, and restructure your CFO presentation around ‘which bucket is this in, and does it meet that bucket’s criteria?’ rather than ‘how much profit did this project return?’ Three years of enterprise Generative AI deployment have clarified one uncomfortable truth: the ROI problem is a measurement problem, not a technology problem. Deploying the system is hard. Knowing what to measure, when, and against what comparison point is harder. In Turkey, the number of organizations doing this correctly is still very small — and that gap is the actual competitive opening.
This article was originally published in Turkish by Gökhan MERCANOĞLU on March 9, 2026. The English edition has been reviewed and edited by the author.