Picture a finance manager at a mid-sized Turkish trading company who processes hundreds of supplier invoices every month. The documents arrive in a dozen different formats: scanned images, PDFs, plain-text emails, even handwritten delivery notes. A conventional RPA robot cannot handle this variety. It reads data only when the data sits in the exact position it was trained to expect. The moment a supplier changes their invoice layout, the robot stops and a human steps in. This is the specific gap that ‘intelligent automation’ promises to close — and understanding whether that promise holds up in practice is exactly what Turkish operations managers need to examine heading into 2019.
RPA in its original form is a rules engine dressed up as a robot. It follows explicit, deterministic instructions: go to this screen, read this field, write this value into that system. For highly structured, repetitive processes this approach works reliably and delivers measurable cost reduction. A number of large Turkish companies, particularly in banking and insurance, have deployed RPA for tasks such as e-Invoice reconciliation, loan application data entry, and regulatory reporting. The consistent finding across these deployments is straightforward: the more standardized the process, the better the robot performs. Any deviation from the expected structure — a missing field, an unfamiliar document layout, an exception case — causes the robot to either throw an error or halt the process entirely. That ceiling is not a flaw in a specific product; it is a structural property of rule-based automation.
Machine learning components added on top of RPA address this ceiling in a specific and bounded way. Optical character recognition enhanced with machine learning models — sometimes called intelligent document processing — can classify documents of varying formats, extract relevant fields, and deliver structured data to the underlying RPA robot. The robot then writes that structured data into the ERP or back-office system as it normally would. Consider a logistics company managing inbound delivery notes from forty different suppliers: no two suppliers format their documents identically. A trained classification model can learn to identify the relevant fields across these variations and pass clean, structured output downstream. The full cycle runs without human intervention — but only for documents the model has been trained to recognize with sufficient confidence. Documents that fall outside the training distribution still require human review.
The decision-support dimension of intelligent automation is more contested and deserves a harder look. Several vendors claim that machine learning models embedded in RPA workflows can make autonomous decisions: approving a credit application based on a risk score, flagging a suspicious transaction, or routing a customer complaint to the correct department without human input. Evaluating these claims requires separating two distinct layers. The technical layer: a model’s decision quality depends entirely on the volume and cleanliness of the training data used to build it. Most Turkish SMEs do not have the historical data volume or the data governance infrastructure needed to train reliable decision models. The operational and legal layer: when an automated decision turns out to be wrong, accountability is unclear. Deploying autonomous decision automation into production without resolving this accountability question is a governance risk, not a technology achievement. A model that performs impressively in a controlled demo environment can behave very differently when it encounters the full diversity of real production data.
Turkey’s economic conditions at the start of 2019 make this conversation both urgent and complicated. Currency pressure and elevated inflation are pushing companies to cut operational costs, and automation is attracting attention as one of the more credible levers. But intelligent automation investment is not limited to software licensing. Data infrastructure work, model training, integration development, and ongoing model maintenance can collectively cost several times the license fee. A realistic project timeline for an intelligent document processing deployment in a manufacturing company’s procurement department — handling invoices from dozens of suppliers in varied formats — is twelve to eighteen months before the model reaches production-grade accuracy. Project plans that assume six months routinely end in disappointment, not because the technology is flawed, but because the data preparation and model validation work was underestimated from the start.
There is also a workforce dimension that Turkish managers often underestimate. Deploying intelligent automation does not eliminate the need for human oversight — it changes the nature of that oversight. Someone needs to monitor model performance, identify when accuracy drops, retrain the model with new document types, and handle the exception cases the model cannot classify. This role requires a different skill set than the staff who previously did the manual work. In organizations where the IT team is already stretched thin, adding model operations responsibilities without dedicated capacity is a setup for gradual performance degradation that goes unnoticed until a costly error surfaces.
For a Turkish SME manager evaluating this technology, the practical starting point is a clear-eyed process audit. Fully structured, invariant processes are still best served by conventional RPA — adding machine learning complexity where it is not needed creates maintenance burden without proportionate benefit. Semi-structured document workflows, such as supplier invoices or customer order forms, are the most credible candidates for intelligent document processing. Before committing to a project, three questions deserve honest answers: Is the available historical document data sufficient in volume and quality to train a model to the accuracy level the business requires? What is the operational cost of a misclassification or a missed field extraction — is it a minor inconvenience or a compliance or financial risk? Who will own model performance monitoring in production, and does that person have the time and capability to do it? If these questions cannot be answered clearly, deferring the project is the more disciplined choice. Intelligent automation genuinely expands what automation can reach — but that expansion requires infrastructure maturity that cannot be shortcut.
This article was originally written in Turkish by Gökhan MERCANOĞLU on January 14, 2019 and has been automatically translated into English and other languages using machine translation.