An operations director at a retail chain said something last year that stopped me mid-conversation: ‘Since we deployed our AI system, the team spends three hours less each morning on reports. But our rate of wrong inventory decisions hasn’t changed.’ That single sentence captures the central fault line running through AI deployments in Turkey in 2026. Projects are being measured — but the wrong things are being measured. Time saved and transaction volume get tracked. Decision quality gets left in the grey zone where managers say it ‘seemed to improve intuitively.’ The thesis I want to defend here is this: measuring an AI project’s value through process efficiency alone is the equivalent of celebrating reaching the wrong destination faster. The real value sits in whether the decisions produced by that process are actually getting better.Consider a mid-sized retail chain in Ankara with 312 employees. The company integrated an agentic system into its stock planning workflow in early 2025. The system combines supplier delivery delays, shelf turnover velocity, and seasonal demand signals to generate weekly order recommendations. After six months, the company reported: ‘Planning now completes in 14 hours; it used to take 31.’ Is that a success? Yes — but an incomplete one. Over the same period, the number of products running out of stock early dropped from eight items to six. There is improvement. But given the investment, the question ‘where does this sit relative to expectations?’ remains unanswered. Because at the start of the project, nobody asked how decision quality would be measured. Efficiency was tracked. Value was not.Here is a distinction worth making explicit: is the AI system co-producing the order decision alongside a human, or is it replacing the human decision altogether? That question determines which metrics make any sense. In a model where the system generates recommendations and a human approves them, you cannot declare the system successful or unsuccessful without isolating the quality of the human’s contribution. If the human approves every recommendation, the system is effectively making the decision — and the quality of that decision is not being monitored. If the human modifies half the recommendations, the system is genuinely a decision support tool — and whether those modifications improve or worsen outcomes is not being tracked. Both scenarios can look identical on a dashboard showing ’31 hours reduced to 14.’ Neither tells you whether the company is stocking better. Without this separation, no AI governance framework — however well constructed — will produce a meaningful verdict on project value.In Turkey, the structural problem is that measurement gaps and data infrastructure gaps are tangled together. A useful parallel comes from a building materials wholesaler in Izmir. The company deployed a RAG-based system to classify incoming customer requests and route them to the appropriate sales representative. The system went live in late 2025. Six months later, the system’s correct classification rate was measured at %46. That sounds like a reasonable first-pass result. But correct classification is not the same thing as decision quality at the point of sale. Of the correctly classified requests — did the customer actually purchase? Did the representative offer a complementary product? Did the complaint category show up again three weeks later? None of that was tracked. The technical metric looked acceptable. The business outcome remained ambiguous. This is the pattern repeating itself across Turkish AI deployments: technical benchmarks are defined during procurement, business outcome benchmarks are defined — if at all — after the project is already running.Turning value measurement into an operational practice comes down to three concrete steps, taken in the right order. The first: before the project begins, identify which decisions the system will directly influence and record the current error rate or cost associated with those decisions. Without a baseline, there is no comparison — only narrative. The second: map technical metrics onto business metrics explicitly and in advance. If the recommendation acceptance rate is %61, what is its documented relationship to order fulfilment accuracy or margin? That linkage should be written down before deployment, not constructed post-hoc to justify the investment. The third — and the one most frequently skipped — is to log every instance where a human overrides or departs from the system’s recommendation. This ‘override’ data is the most honest signal available. It reveals both where the system’s reliability boundary actually sits and where employee capability gaps are concentrated. In the majority of Turkish AI projects I have seen, the first of these three steps is not completed before the system is installed. Value is then reconstructed retrospectively, which rarely produces a coherent picture.The EU AI Act’s phased enforcement from 2025 onwards converts this measurement gap from a management problem into a compliance problem. For systems operating in high-risk categories, you are now required to document how human oversight is applied and where the decision boundary sits between human and machine. Turkish companies exporting to EU markets or reporting to European parent entities are feeling this requirement in practice. A pharmaceutical distributor using a decision support system for product return authorisations must maintain auditable records to the standards expected by an EU buyer. A company that cannot show, in writing, whether ‘the system recommended’ or ‘the system decided’ is in a technically non-compliant position regardless of how well the system performs technically. This removes value measurement from the category of abstract management discussion and places it firmly inside risk governance. The question ‘how do we measure this?’ now has a legal dimension that was not present two years ago.Measuring an AI project’s value through process efficiency alone is celebrating reaching the wrong destination faster — that is the core of what I am arguing here. The stock planner at the Ankara retail chain is still asking the same question every Monday morning: right product, right shelf, right time? The system returned three hours to the morning routine. What happened to those three hours, and whether the decisions improved, remains uncharted. In 2026, AI projects need to earn their place not by securing budget approval but by demonstrating measurable business output. The path to that demonstration runs through one question asked before deployment, not six months after it: which decisions should be better, and how will we know when they are? Answering that question in advance is considerably less expensive than trying to reconstruct the answer once the contract has been signed and the system is already live.
This article was originally published in Turkish by Gökhan MERCANOĞLU on September 9, 2026. The English edition has been reviewed and edited by the author.