A customer types ‘where is my order?’ into your e-commerce support bot. The bot asks for an order number. The customer provides it. The bot replies ‘your order has been shipped.’ The customer writes ‘I have been seeing the same message for three days.’ The bot repeats the same response. The conversation ends there, and the customer picks up the phone. This scenario captures why many SMEs in Turkey fail to extract the value they expect from chatbot investments. The problem is not a lack of technology — it is the failure to treat intent, sentiment and context as a unified system.
In chatbot architecture, intent recognition is the process of classifying what a user wants to accomplish with a given message. ‘I want to see my invoice,’ ‘show me my last payment’ and ‘what are my account transactions?’ are phrased differently but carry the same intent. In a natural language processing (NLP) framework, these sentences are mapped to a shared intent label — for example, ‘invoice_query.’ But intent recognition alone does not complete a dialogue. When a user asks the same question for the third time, it signals a gap between expectation and delivered service. Designing a bot that is blind to this gap wastes the technical investment underneath it.
This is where sentiment analysis enters the picture. This layer classifies the emotional charge of a user’s message — positive, negative or neutral — giving the bot not only ‘what is being said’ but ‘how it is being said.’ ‘My order hasn’t arrived yet’ and ‘my order hasn’t arrived in a week, this is unacceptable’ carry the same intent; the second, however, carries a clearly negative and elevated emotional load. In a well-designed dialogue flow, when this intensity crosses a defined threshold, the bot should escalate the conversation to a human agent or offer resolution options rather than cycling through standard responses. Without this escalation mechanism, a chatbot does not improve customer satisfaction — it actively damages brand perception.
Context tracking is the memory of the conversation. When a user asks ‘what is the phone number of the Istanbul branch?’ and then writes ‘opening hours?’, the second message contains no explicit reference to Istanbul. A system without context tracking interprets this as a fresh query and produces a meaningless response. A context-aware system keeps the entity from the previous turn — ‘Istanbul branch’ — active and resolves the follow-up question within that frame. In enterprise customer service scenarios, particularly multi-step transactions such as return requests, quote inquiries or appointment scheduling, coherent dialogue is structurally impossible without context tracking.
The point where these three layers — intent, sentiment, context — must operate together is precisely where technical capability meets dialogue design. A pattern observed frequently in SME projects across Turkey is this: the development team builds the NLP infrastructure and trains the intent classifier, while the team responsible for dialogue flows writes conversation scripts without a shared understanding of how these three layers interact. The result is a bot that works technically but feels broken to the user. Closing this gap requires dialogue designers, product managers and NLP engineers to operate within the same conceptual framework from the outset.
In practice, one of the most persistent challenges stems from Turkish morphology. Turkish is an agglutinative language, meaning a single root word can appear in dozens of surface forms. ‘Fatura’ (invoice), and its inflected variants all refer to the same entity but look entirely different at the character level. Training a model to recognize these as the same concept demands significantly more data and domain-specific effort than adapting off-the-shelf English NLP libraries. Commercial chatbot platforms vary considerably in their actual maturity of Turkish language support. Before committing to a vendor, an SME manager should ask specifically how many training examples underpin the Turkish model and which spelling and dialectal variations it covers — not simply whether ‘Turkish is supported.’
When evaluating a chatbot investment, an SME manager should apply three concrete tests. First, does the system consistently classify the same intent regardless of how differently users phrase it? Second, does it have a defined mechanism to escalate conversations to a human agent when negative sentiment intensity exceeds a set threshold? Third, in multi-turn dialogues, does context carry forward reliably from one message to the next? A system that cannot answer ‘yes’ to all three does not improve the customer experience — it relocates operational cost to a different line item without reducing it. A chatbot architecture that treats intent, sentiment and context as an integrated whole, on the other hand, both raises customer satisfaction and genuinely reduces human agent load. That reduction is the primary variable in any honest ROI calculation for this category of investment.
This article was originally written in Turkish by Gökhan MERCANOĞLU on June 11, 2018 and has been automatically translated into English and other languages using machine translation.