Your RPA bot processes supplier invoices without a single error, sends inventory reports on schedule, and fills order forms automatically. That part works. Now look at the two-page complaint email your customer sent this morning: the bot sees it but has no idea what to do with it, because no rule captures what that text actually means. That is precisely where RPA’s wall begins. The Generative AI conversation that entered the business mainstream in early 2023 is fundamentally asking the same question: what changes when rule-based automation is combined with the capacity to read and produce language? But a word of caution — since ChatGPT’s December 2022 launch, every consulting deck seems to carry the same claim that ‘everything will be automated,’ and that claim is shaping up to be the least reliable promise of the year. My position is this: Generative AI does not replace RPA. It covers the unstructured text processes that RPA was never designed to see. These two technologies are not competitors — they are complements. But only when data quality and process maturity have already reached a workable threshold.Without a clear-eyed view of what RPA does well and where it stops, the Generative AI discussion stays in the air. Consider a mid-sized automotive spare parts manufacturer in Bursa with 312 employees, running customer order confirmations and supplier correspondence entirely through email. This company’s RPA bot has been moving structured data between the ERP and the accounting module for years without incident. Structured data — consistent formats, fixed fields, predictable values — works well in bots. The problem: between 40 and 50 customer emails arrive each day, some reporting shipment delays, some raising technical complaints, some simply asking ‘where is my order?’ The bot cannot read these. The accounting and sales team spends the first two hours of every morning sorting them manually. By RPA’s design, this territory was always empty — because a rule cannot define what a sentence means. That gap costs over fifty person-hours a month in a firm this size, and it compounds every week the inbox grows.What Generative AI adds to this picture is a layer that can read unstructured, free-format text — emails, scanned fax documents, customer feedback, technical specification PDFs — classify it with reasonable accuracy, and trigger a defined downstream workflow. An LLM-based tool can distinguish between a ‘shipment delay notification’ and a ‘technical complaint’ and route the message to the correct person or queue. In the Bursa spare parts case described above, adding this classification layer brought the morning email triage from roughly two hours down to about forty minutes — a drop observed by a senior operations consultant in a production environment, the firm’s name withheld but the measurement verified. One critical distinction: the system is not automatically replying to emails. It is only classifying and routing. The decision and the communication remain with a human. That boundary is not a limitation to apologise for; it is, for now, exactly the right design.Here is the tension that is receiving the least attention in Turkey this year: KVKK and Generative AI overlap in the same data pipeline, and for Turkish SMEs that creates a concrete legal exposure. When an LLM-based tool processes a customer email, the personal data inside that email — name, phone number, order detail — is sent to a cloud model API. Under Article 12 of the Personal Data Protection Law, this constitutes a transfer of personal data. If the API’s servers are located outside Turkey, cross-border transfer rules also apply. The Personal Data Protection Authority has not yet issued guidance specifically targeting Generative AI — but the existing definition of ‘data transfer’ is broad enough to cover it. In practice, many Turkish SMEs are currently testing commercial LLM APIs without recognising this exposure. A small internal pilot may carry manageable risk. But once a system moves into production and the data it processes qualifies as personal data, operating without a legal basis can become expensive very quickly, and not only in financial terms.So what does this mean for a Turkish SME manager on a Monday morning? Three concrete steps. First, run a process inventory. Which steps in your operation involve unstructured text — customer emails, complaint forms, technical documents, supplier correspondence? List them separately and measure how much manual time they consume today. That number is your baseline; without it, any pilot result is uninterpretable. Second, classify the data. Does personal data flow through those unstructured text processes? If yes, sit down with your KVKK counsel before testing any LLM tool — clarify the ‘disclosure obligation’ and the ‘data transfer mechanism’ questions before you touch production data. Third, start small and contained. Begin your Generative AI classification experiment with internal documents that carry no personal data — production reports, internal technical specs, non-customer correspondence. In that environment you can measure real accuracy and, critically, you will see hallucination risk first-hand: the model confidently producing an answer that simply is not there. Experiencing that behaviour in a low-stakes setting before it reaches a customer-facing workflow is not a detour. It is the shortest path to a decision you can defend.For five years, RPA was the only meaning the word ‘automation’ had in Turkish business conversations. Generative AI expands that territory — but it does not erase the map that already existed. For structured, rule-bound data processes, RPA remains more reliable, cheaper to operate, and more predictable than any LLM-based alternative. Generative AI’s real value sits in the unstructured text islands that automation has never reached. The Bursa spare parts team’s morning email marathon is not an abstract inefficiency; reclaiming those forty minutes a day is a measurable operational gain. But capturing that gain requires knowing exactly where your data goes, carrying your KVKK obligations before you move to production, and taking seriously the possibility that a model will produce a confident-sounding answer that is factually wrong. The question worth sitting with is this: how much of your organisation’s daily work lives inside unstructured text today — and is anyone actually measuring it?
This article was originally published in Turkish by Gökhan MERCANOĞLU on May 22, 2023. The English edition has been reviewed and edited by the author.