A company that writes its data security policy after deploying AI agents loses oversight the moment those agents go live — and walking that back turns out to be far harder than expected. That sentence would have sounded like contrarian caution in 2024. By mid-2026, it is an operational summary from a mid-size insurance company in Istanbul. With 312 employees managing policy administration and claims processing, the firm ran three separate AI agent pilots over twelve months. All three delivered technical results. All three surfaced the same gap: no one had formally written down which decisions belonged to the agent and which required a human signature. Errors surfaced — in the wrong place and at the wrong time.The underlying error is structural, not technical. Most organizations frame AI agents as automation tools and leave the human-control question to the architecture team. That is the wrong delegation. Where human oversight is required is a liability decision, not a configuration choice. The EU AI Act — in force since 2025 — reaches Turkish companies that sell goods or services into EU markets or transfer data to EU-based processors. Its mandatory human oversight requirements for high-risk AI systems place Turkish finance, insurance, and healthcare sectors at the front of the compliance line. Yet the majority of organizations in these sectors still have nothing more than a draft policy: no risk classification, no accountability map, no documented escalation path.So where exactly must human control stay, and where can it safely step back? The starting point is exposure calculation, not vendor capability. A claims-processing agent can read incoming documents and assign categories — low exposure, no human sign-off required. If that same agent issues a payment decision or initiates direct customer communication, a human signature is mandatory. Placing an ‘approval step’ in the workflow is not sufficient; that step must be logged, reasoned, and queryable after the fact. The Istanbul insurer completed a 6-week internal audit to produce precisely this distinction. The result: out of 47 end-to-end workflows, 29 moved to full automation while 18 were classified as human-signed — and that ratio was embedded as an SLA clause rather than left as informal practice.Data privacy introduces a separate dimension that organizations routinely underestimate. The primary risk in feeding corporate data to generative AI systems is rarely the headline breach scenario. It is data persistence. When a customer record passes through an external LLM API, how long does it remain on a third-party server? Under what conditions can it be used as training data? Does that transfer qualify as an overseas data transfer under KVKK? These were speculative questions in 2023. By now they are standing audit items from Turkey’s Personal Data Protection Authority. A substantial number of Turkish SMEs still evaluate cloud-based LLM services entirely outside their KVKK compliance framework — treating the AI vendor relationship as a SaaS procurement rather than a data processing agreement. A single data breach notification or regulatory inquiry tends to correct that confusion at significant cost.This is where the SLM — small language model — conversation earns practical weight. A task-specific model running on local infrastructure eliminates data egress entirely. The tradeoff is real: general-purpose reasoning is limited, and fine-tuning demands sustained technical investment. A medical device distributor operating in Ankara, with 284 employees managing patient-record workflows and post-sales service, ran a direct comparison: a cloud LLM integration (low setup cost, high data-transfer exposure) versus a locally deployed SLM (11-month development cycle, full KVKK alignment). The company chose the second path. The ROI calculation turned positive at the 24-month mark. The harder management challenge, as the IT director put it, was not the technical work — it was explaining to senior leadership for eleven consecutive months why this was taking longer than the vendor’s demo suggested it would.Measuring productivity gains accurately is the question organizations skip most reliably. During pilots, processing speed tends to be the only metric tracked. ‘Average claims processing time dropped from 4.2 days to 1.1 days’ is a real result — but it is an incomplete one. The questions that actually matter run alongside it: what happened to the error rate per decision? Did customer complaint volume increase? How frequently did human reviewers correct agent output in the first 90 days? These three indicators together tell you whether the gain is real or redistributed. Returning to the Istanbul insurer: a 52% improvement in processing speed was recorded. Over the same period, customer complaints rose by 63% in the first quarter — because the agent was issuing rejection decisions without sufficient explanatory reasoning. Reading those two numbers in isolation produces opposite conclusions about whether the project succeeded.All of this consolidates into a single framing: granting authority to an AI agent is a corporate commitment, not a configuration parameter. If you have not documented which decisions belong to the agent and which require a human today, you will be forced to answer that question under incident pressure later — with KVKK exposure, EU AI Act liability, and customer trust all waiting in the same queue. The Istanbul insurer’s answer to three failed pilots was a published AI accountability matrix: every workflow now carries an explicit agent authority level, a human approval threshold, and a deviation log. Progress is slower than before the framework existed. But the organization knows where it is going — and more importantly, it knows who is responsible when something goes wrong.
This article was originally published in Turkish by Gökhan MERCANOĞLU on August 19, 2026. The English edition has been reviewed and edited by the author.