# TAHSİLDAR: A New Era in Collection Operations

- Published: 2026-10-26
- Source: https://www.cbot.ai/tahsildar-a-new-era-in-collection-operations/

Success in voice AI is not limited to delivering a natural, fluent conversation. The real value comes from handling the complexity of real operations, supporting business processes end to end and producing measurable results. With TAHSİLDAR, the voice digital employee we have built at CBOT, we approach collection processes from an operational perspective, not just a technological one. By bringing together years of sector experience, the voice technologies we have developed and the knowledge we have gained from real operations, we aim to deliver a secure, efficient and sustainable collection experience.
10:24 a.m. Seven different customers from the same collection list are being called. One customer picks up the phone; another's spouse answers. Someone asks to be called back later. One customer starts speaking in English, another says they can't pay. Some ask for a restructuring, while others only want to speak to the agent they talked to before.

Operationally, all of these calls fall into the same category: a collection call. Yet each one calls for a different communication approach, decision logic and process.

Collections are often seen as one of the best candidates for automation, on the assumption that they follow a fixed set of steps. Our experience in the financial sector at CBOT tells a different story: **Collections are one of the most demanding operational tests for AI.**

That is because a single call tests four critical capabilities at once: handling the unexpected, turning the conversation into a concrete business outcome, protecting customer trust, and doing all of this at a sustainable cost.

## From Demo to Real Operations: The Test for AI

CBOT CEO Mete Aktaş often makes this point:

> "The market is full of demos right now, not scars."

This captures the key difference in how generative AI applications move into enterprise use. A system that performs well under controlled conditions needs far more comprehensive technology and process design to handle the complexity of real operations just as well.

In a collection demo, the customer answers the phone, confirms the debt and gives a payment date. In real operations, calls rarely follow such a predictable path.

When building [TAHSİLDAR](https://www.cbot.ai/tahsildar/) at CBOT, we put this reality at the center of the design. We treated the call process not as a linear dialogue flow, but as a complete structure covering different scenarios and decisions.

The result is a comprehensive call architecture made up of **61 decision points, 111 transitions and 20 possible outcomes**.

The architecture goes beyond standard collection scenarios. A customer who says they have lost a loved one, a family member reporting that the debtor has passed away, a person in hospital, or an employee who wants to plan their payment around payday each require different business rules and communication approaches.

For example, the answer "I'll pay when my salary comes in", common in Türkiye, does not mean a firm promise to pay from the system's point of view. First the payday needs to be confirmed, and then the payment plan is built around that date.

That is why we designed TAHSİLDAR's decision logic not only around what it should say, but also around which actions it must not take under which conditions.

No debt information is shared before identity verification is complete. Legal disclosures are never skipped. Ambiguous statements are not treated as explicit consent. A customer who says they cannot pay is not offered a new proposal beyond the limits set by the institution. The preference of a customer who does not want to be contacted is recorded.

These rules are not safeguards added to the system afterwards; they are among its core design principles.

Under their [AI governance](https://www.cbot.ai/ai-governance-safety/) framework, institutions can also set the digital employee's level of authority for each action. The digital employee can complete the action on its own, wait for an authorized approval, or hand the process over to the relevant team.

**This way, decision and control mechanisms are managed within the business rules set by the institution.**

https://www.youtube.com/watch?v=YqxQyDDnEPw

## From Conversation Quality to Operational Efficiency

Voice AI systems are usually judged on natural conversation, answer accuracy and low latency. At enterprise scale, however, these criteria alone are not enough for success.

The real value appears in the concrete results that reach business processes once the call is over.

**The success of a voice digital employee should be measured not only by how it handles the call, but also by what it gets done after the call.**

With this in mind, we built TAHSİLDAR to support the collection process end to end, from preparation before the call to the actions that follow it.

**Managing call lists:** The details of the customers to be called come directly from the collection system. Call frequency, timing and prioritization follow the rules set by the institution.

**Recording call outcomes:** Promises to pay, payment plans and other information gathered during the call are passed to the collection system as structured data. Information does not stay as a call note; it becomes usable in operational processes.

**Scheduling callbacks:** The callback time the customer asks for is checked against the institution's contact rules. The customer is not told that a callback has been scheduled until the action has been completed successfully in the system.

**Handover to a human agent with context:** Restructuring requests, objections and cases that need a live agent are transferred to the relevant team. The agent's screen shows the call transcript, the detected customer intent, the information collected, the reason for the handover and a summary of the call.

After the handover, AI keeps [supporting the agent](https://www.cbot.ai/agent-experience/). It summarizes conversations, suggests replies and reminds the agent of next steps. Final decisions and control stay with the agent.

**Seamless communication across channels:** When needed, the conversation can continue in another language or on a text channel such as WhatsApp. Information the customer has already shared and the context of the conversation are kept across channels.

This approach turns [omnichannel customer communication](https://www.cbot.ai/tahsildar/omnichannel-debt-collection/) from a set of separate channels into one unified experience.

TAHSİLDAR works with existing PBX and IVR infrastructure over SIP and MRCP, and in the background with CRM, core banking and collection systems. The digital employee's data access is limited by role, and calls are recorded in a traceable, auditable way.

## A New Dimension of Security in Voice AI

In collection calls, trust is one of the most critical parts of the conversation. For the call to go well, customers need to be sure of the institution calling them and know that the information they share is safe.

As generative AI technologies advance, voices can be synthesized realistically. This opens up new possibilities for institutions, but it also brings new responsibilities for data security and digital identity management.

The rise of deepfake technologies in particular makes it harder to treat a voice on its own as a reliable means of identity verification.

That is why we treat enterprise voice technologies not only as a tool to improve customer experience, but as a digital asset that must be managed securely.

When developing our voice AI architecture at CBOT, we focus on three key questions: Where is voice data stored, on which infrastructure do the models run, and who controls the institution's voice technologies?

**On-premise deployment options**

Our platform, including the language model, speech recognition, speech synthesis, enterprise knowledge management, workflows and analytics components, can run [entirely on-premise](https://www.cbot.ai/deployment-options/).

This means sensitive customer data can be processed within the security boundaries set by the institution. In cloud and hybrid architectures, infrastructure choices are shaped by the institution's data security requirements and operational needs.

**Voice technologies we build ourselves**

At CBOT, we develop our speech recognition and speech synthesis [engines](https://www.cbot.ai/tahsildar/ai-voice-agent-for-debt-collection/) with our own R&D capabilities.

This lets us work directly on the pronunciation, stress and regional accents of Turkish. It also lets us build architectures that do not require voice data to leave the institution, and optimize latency for the institution's own infrastructure.

**Managing the corporate voice identity**

We apply the same security approach to custom voice development and voice cloning for institutions.

Voice generation is based on explicit consent. Voice profiles are stored in environments approved by the institution. Usage rights are defined by role, and changes are logged.

This approach helps voice AI applications meet enterprise governance requirements, especially in banking, finance and other sectors where personal data security is critical.

**Security is not only about preventing risks, but also about making them traceable and manageable.**

## GPU Optimization and Sustainable Cost Management

Running voice AI systems on-premise gives more control over data, but it also requires efficient management of computing resources.

Especially in high-volume operations, GPU capacity becomes one of the major components of total cost of ownership.

Collection operations are a striking example. Call volumes are not steady throughout the month. Demand rises around due dates and the end of the month, so infrastructure capacity has to be managed flexibly.

Fixing capacity at peak demand alone can lead to resources being used inefficiently the rest of the time.

That is why **we see GPU optimization not only as a matter of technical performance, but as a direct part of operational efficiency and cost management.**

Since our founding, we have operated as an R&D-focused technology company. Efficient inference and the optimization of computing resources are among the key focus areas of the [R&D capabilities](https://www.cbot.ai/cbot-research/) we have built through university collaborations, TÜBİTAK-supported work and European Union projects.

In this context, we combine several technical approaches.

**Choosing the right model for the job:** Instead of using the largest model for every task, we choose models that match the capacity the task requires. Models such as [CBOT LLM](https://www.cbot.ai/cbot-llm/), which can run on both GPU and CPU, add flexibility to infrastructure planning.

**Separating workloads:** We separate tasks that need GPUs from tasks that need less computing capacity.

**Independent scaling:** We build architectures that can scale speech recognition, speech synthesis and language model components separately. When demand rises, only the services that need it get more capacity.

**Warm capacity management:** We use methods that reduce the delays that can occur while new capacity comes online during busy periods.

**Performance monitoring:** We assess system performance by tracking [key infrastructure indicators](https://www.cbot.ai/ai-infrastructure-inference/) such as time to first response, queue load and GPU utilization.

There is no single infrastructure model that fits every institution. We plan capacity based on call volume, the number of concurrent calls, latency targets and data security requirements.

## Measuring Success by Results, Not by Call Counts

When assessing the success of AI in collection operations, looking only at the number of calls made or the promises to pay collected is not enough.

What matters is how far the calls turn into actual payments.

For example, a high promise-to-pay rate does not point to strong performance on its own if a large share of those promises are not kept.

That is why we look at several indicators together when [measuring performance](https://www.cbot.ai/tahsildar/debt-collection-ai-kpis/).

We compare the promise-to-pay rate with the kept-promise rate. We track the right-party contact rate to assess the quality of customer data. We follow actual payments at 1, 3, 7 and 30 days after the call.

To understand the real contribution of AI, we value controlled comparisons with customer groups that have similar risk profiles.

So rather than offering a single predetermined success rate for every institution, we believe performance should be measured against the institution's own data and operational goals.

**Sustainable success comes less from making more calls than from turning more calls into accurate, measurable business results.**

## Six Critical Questions Leaders Should Ask

For leaders evaluating voice AI in collection operations, operational readiness matters as much as the choice of technology.

That is why we recommend answering six key questions before deploying a digital employee:

1. **Scenario management:** Have you defined the conditions under which calls branch in different directions and the business rules that apply in each case?
2. **Process integration:** When a call ends, which information and actions are passed to existing systems?
3. **Handover to human agents:** When a handover is needed, what information can the agent access, and is the context of the call preserved?
4. **Data security:** Where are voice data, AI models and corporate voice profiles stored, and how are they managed?
5. **Infrastructure efficiency:** How is the capacity needed during peak periods planned, and how are costs optimized?
6. **Performance measurement:** Is success judged only by the number of calls, or by actual payments and kept promises?

These questions offer a useful evaluation framework not only for collection processes, but also for enterprise AI projects across other sectors.

## Conclusion: Real Value Shows Up in the Field

At CBOT, we have been developing enterprise AI technologies since 2017. Today we work with more than 40 financial institutions, 15 of them banks. Our market share of more than 80 percent among companies providing AI infrastructure to the banking sector in Türkiye, and the 26,000 active digital employees we reached by the end of 2025, show the scale of our experience in this field.

TAHSİLDAR is the result of years of operational experience in the financial sector, the technologies we have developed and what we have learned from real use cases.

This experience has shown us something important: **Success in enterprise AI does not come from using an advanced model alone, but from reliably integrating technology into the institution's real business processes.**

This approach is also shaping the future of voice digital employees. The systems that stand out are those that give accurate answers, manage business processes end to end, work alongside human agents when needed, protect data security and do all of this at a sustainable cost.

With TAHSİLDAR, our goal is to bring these capabilities together with the real needs of collection operations.

Because the real value of AI shows up not in what it can do in a demo, but in the results it delivers in the field.