The market is becoming polarized and creating a gap

The budget standard segment

At one end of the spectrum are streamlined, cost-effective solutions designed for a quick, largely self-sufficient start. They answer calls, provide business hours, answer simple questions, schedule appointments, or take messages. They are practical and efficient for clearly standardized use cases and small organizations. Their limitations become apparent when issues become more complex, require deeper knowledge, involve more sophisticated integrations, or entail heightened data protection and operational requirements.

The Enterprise Segment

At the other end of the spectrum are comprehensive contact center platforms designed for large, integrated service environments with high call volumes, sophisticatedrouting and agent processes, and dedicated automation teams. For large corporations, these are often the right choice. For many medium-sized and institutional organizations, however, they are too comprehensive, too process-heavy, and require too much integration.

The gap between them

Between these two extremes lies a large portion of the SME sector: organizations with many recurring but substantively complex phone inquiries that want to professionally design a genuine service channel — without resorting to a purely off-the-shelf solution or a large-scale contact center project. For this group, the usual decision-making frameworks are only of limited use.

Why the price per minute is the least reliable criterion

The price per minute of conversation is the most easily comparable metric on the market, which is why it dominates the discussion. That is precisely where the problem lies: comparability is not the same as informative value. The price per minute obscures the factors that determine actual service quality: the design of the conversations, the quality and verifiability of the stored knowledge, integration with specialized systems, the data protection architecture, and day-to-day operations.

In addition, the per-minute price listed is rarely the total price in practice. Depending on the offering, costs for telephony, language models, data storage, support, compliance options, and development services may be added. A sound decision is therefore not made by comparing individual per-minute rates, but by considering the overall performance over the system’s lifecycle. The relevant question, then, is not “How cheap is a minute of talk time?” but rather “How do we design a phone service that fits our organization and can be operated responsibly?”

How an AI phone assistant is technically structured

To evaluate this, it helps to take a look at the architecture underlying most systems. At its core, it consists of two levels and a tension between them.

The Telephone and Voice Layer

It converts spoken language into text in real time (speech-to-text), generates speech from the responses (text-to-speech), and controls the flow of the conversation, including events and interruptions. Good voice quality and low latency — that is , the time between the end of a statement and the start of the response — are crucial for a natural listening experience.

The AI and Knowledge Level

It determines what the assistant does in terms of content: how it categorizes a request, what knowledge base it uses to respond, when it asks for clarification, and when it hands the request off to a human. This level is configured, tested, and refined over time.

The Conflict of Objectives Between Latency and Depth

These two levels are linked in a practical way that is often underestimated. A very large language model with a very broad dataset can handle more in terms of content, but tends to respond more slowly—which, unlike in chat, is immediately perceived as an uncomfortable pause on the phone. In many telephone use cases, therefore, a deliberately smaller model with a focused, well-maintained dataset is the better choice. Call quality is not achieved through maximum model size, but through the proper balance of model, knowledge, and response time.

A natural voice is a prerequisite

Synthetic voices have now advanced to the point where callers often can hardly tell they’re machine-generated. This is a significant step forward, but it’s a basic requirement—not a quality feature in and of itself. An assistant that sounds natural but miscategorizes inquiries raises expectations it cannot meet. The fact that technical naturalness alone does not build trust is also evident in the numbers: In the 2024 Customer Service Barometer, only 43 percent rate the responses from chatbots and digital assistants as trustworthy. True performance, therefore, is demonstrated through behavior: whether the assistant correctly understands the inquiry, provides reliable rather than merely plausible-sounding responses, asks meaningful follow-up questions when something is unclear, and recognizes when it’s necessary to hand the matter over to a human.

What You Need to Know for Chat and Phone Support

In many organizations, website FAQs, chatbots, and phone hotlines are managed separately — by different teams with varying levels of expertise. This creates extra work and, as callers can tell, inconsistent information across channels. The real added value of a platform-based solution therefore lies less in the fact that a system can handle phone calls, and more in the fact that phone and chat draw from the same knowledge base: knowledge is maintained once rather than multiple times, insights from one channel improve the other, and answers remain consistent. A shared knowledge base does not, however, mean identical conversation styles — phone calls follow their own rules for greetings, follow-up questions, confirmations, interruptions, and a style suited to the spoken word, and are specifically designed for the spoken medium.

Integration and Handover to People

Service calls rarely end with just providing information. It is often necessary to retrieve data, create a ticket, schedule an appointment, or continue a process in a specialized application. It is therefore crucial how flexibly a system can be integrated with existing systems such as CRM, ticketing, or specialized applications — including more specialized systems that fall outside the standard framework of common contact center integrations. Equally important is the structured handoff to staff members: A good telephone assistant recognizes their limitations and transfers the call rather than continuing a conversation that goes beyond their expertise.

Privacy by Design

Information requiring special protection is regularly discussed over the phone. For regulated service providers, public organizations, and applications in the healthcare sector, data protection is therefore not just a general selling point but a specific selection criterion — one that cannot be added after the fact. It must be clarified early on where calls and data are processed, what data flows arise, and whether the AI and knowledge infrastructure can be operated in an EU-hosted, private, or on-premises environment when requirements are more stringent. These questions belong at the beginning of a project, not in the final phase.

A phone bot is a project, not a switch

These criteria lead to a practical conclusion: A reliable phone assistant capable of handling complex inquiries cannot be created simply by having a well-worded prompt and an active phone number. Among other things, call objectives, typical pitfalls, escalation and transfer procedures, data access, and tone of voice must be designed — and all of this must be tested before regular operation begins and continuously refined thereafter. This calls for an approach that combines the platform with ongoing support and views day-to-day operations as a distinct quality service. Availability, reducing the workload on employees, consistent service across all channels, and controlled refinement are the key metrics by which value is measured.

How to make this happen: the Phonebot from Kauz.ai

This is exactly where the Kauz.ai Phonebot comes in. It is designed for discerning small and medium-sized businesses and institutional organizations — that is, for the space between affordable standard phone assistants and enterprise call center suites, where the criteria described above take center stage.

The starting point is deliberately not an isolated telephony product, but rather the shared AI assistant platform: Phone bots are created, configured, and further developed in aiStudio and share the same knowledge base as chatbots. This allows the phone bot to answer specific queries just like the website chat, while its conversational style is specifically tailored for the voice channel. For quality control, the platform incorporates tools proven in chatbot operations — knowledge management, transparency regarding the source of responses, tests and test reports prior to changes, insight into real conversations, and, for particularly sensitive topics, expert-reviewed answers drawn from topic catalogs.

At the telephony level, Kauz.ai relies on an EU-hosted voice and telephony infrastructure; the AI and knowledge layer can be integrated with EU hosting or, for more demanding requirements, with private cloud or on-premises options . Existing systems—such as those for data retrieval, ticketing, CRM, or specialized applications—can be integrated via flexible programming interfaces, as can the handoff to employees. The platform is complemented by support for design, setup, testing, and further development—depending on the project, either as managed operations or under shared responsibility with internal teams and partners.

Phonebots for Your Customer Experience

For small and medium-sized businesses, AI-powered phone assistance is no longer a question of “if,” but of “how.” This “how” determines far more than just a single metric. Whether a phone assistant adds value depends on the interplay of several factors: the ability to control call behavior; a shared knowledge base for chat and phone interactions that allows for independent call handling; the ability to integrate into existing systems; a data protection architecture designed from the outset; and operations designed to deliver consistent quality performance.

Anyone who uses these points as a benchmark shifts their perspective: away from a tool that you activate, toward a service channel that you design. For organizations whose issues are complex and whose reputation hinges on every single call, this is the real opportunity. They can tailor the phone — often their most direct and most frequently used line of communication with customers and citizens — so that accessibility, reliability, and consistency are not left to chance. This is precisely the difference between a number answered by an AI voice and a telephone service that fits the organization.

Smart AI Solutions for Small and Medium-Sized Businesses

1. Technology-agnostic AI platform

A technology-agnostic AI platform is characterized by several features. Support for and flexible use of various LLMs form the foundation for a future-proof AI strategy. This is because not only is the performance of these models evolving rapidly—their optimal areas of application are also constantly changing. A model that is the best choice for a specific task today may be replaced tomorrow by a more powerful or cost-effective alternative.

Open APIs and standardized interfaces are equally crucial. They enable seamless integration into existing enterprise systems and lay the foundation for an interoperable AI landscape. New technologies and standards—such as the Model Context Protocol (MCP) for the standardized connection of AI agents to enterprise systems—can thus be easily integrated. This keeps the platform open to future innovations, reduces proprietary dependencies, and protects existing investments.

2. Data Sovereignty Through Flexible Operating Models

Companies vary significantly in terms of their data protection, compliance, and IT infrastructure requirements. While a traditional cloud environment may be sufficient for some use cases, the location and operator of the cloud infrastructure also play a decisive role in choosing the operating model. Cloud offerings from major international providers such as Microsoft Azure, AWS, or Google Cloud may be suitable for many scenarios, while regulated industries in particular are increasingly turning to European cloud providers such as Open Telekom Cloud, Hetzner, or OVHcloud to meet stricter requirements for data sovereignty and compliance. For particularly sensitive use cases—such as in the financial or healthcare sectors—a private cloud or on-premises solution may be required, in which confidential information is processed within the organization’s own control boundaries. A sovereign AI platform should therefore support various hosting options and give companies the freedom to choose the operating model that best suits their needs—and to flexibly adapt or switch it as needed.

3. AI must not be a black box

Who uses AI in the company? What is the knowledge base behind the results? What are the actual costs of using AI? Where can it be improved? And can AI results be reviewed and, if necessary, adjusted manually?

As AI usage increases, it becomes difficult to maintain an overview of content and organizational structure. Different AI tools, decentralized licenses, and unclear responsibilities make it difficult to use AI securely and cost-effectively. At the same time, the risk of “shadow AI” increases when employees resort to unauthorized applications, with all the associated data protection implications. A in-house CompanyGPT as a centralized, approved work environment is therefore one of the most effective measures for curbing unauthorized AI use and preventing data leaks.

A central AI platform also provides transparency and control: It consolidates various AI applications into a unified working environment, manages access through role-based and permission-based models, and simplifies license management. At Kauz.ai, this transparency extends beyond day-to-day operations: AI chatbots can also be specifically controlled in terms of content and continuously optimized. Dialogues are documented and can be reviewed and improved as part of human-in-the-loop and feedback processes. This ensures that not only the usage but also the quality of the AI applications remains traceable at all times.

We combine governance, cost control, and quality management into a single interface, making AI transparent and manageable for IT, business units, and management alike. This transforms AI into a professionally operated enterprise application.

4. Maximum Cost Control

The costs of professional AI deployment are difficult to calculate: token prices fluctuate constantly, and usage varies depending on the task, frequency of use, and language model. Systematic cost monitoring thus becomes a crucial management tool for assessing which use cases are worthwhile and which employees actually need which licenses.

To this end, Kauz.ai offers a cost dashboard that transparently displays the consumption of AI points—a unit that makes the costs of different models comparable—by time period and use case. In addition, KauzSelection optimizes costs by automatically assigning each request to the appropriate LLM: Expensive top-tier models are used only where they are truly needed.

The licensing model is also designed to be flexible: Instead of rigid flat rates per user—which often force unnecessary upgrades—the pricing packages include a shared AI point allowance. If additional points are needed, they can be purchased on an ad hoc basis. This makes AI costs predictable, and companies pay only for what they actually use.

5. Think About Compliance from the Very Beginning

With the EU AI Act, the GDPR, and regulations such as DORA, the regulatory framework for AI use is becoming increasingly concrete. In particular, the traceability of AI decisions is evolving from an option to a requirement: The AI Act mandates comprehensive logging for certain AI systems to ensure traceability and human oversight.

A future-proof AI platform should not address compliance only after the fact, but should incorporate it from the ground up. That’s why our architecture is designed from the outset to meet the regulatory requirements of the EU AI Act and the GDPR—from the hosting infrastructure and data processing to the contractual framework. A finely granular role- and permission-based system ensures that only authorized individuals can review and approve AI decisions: a key component of the required human oversight.

6. Independent of external specialists

Data sovereignty has both a technical and a human dimension. Those who rely on external developers or consultants for every adjustment to their AI applications are simply trading dependence on the technology provider for dependence on the service provider. No-code platforms solve this problem by empowering business departments to independently create, maintain, and scale AI chatbots, AI agents, and AI workflows—without any IT support.

aiWorkplace and aiStudio, designed for developing AI chatbots, are consistently tailored for business users: content management, control of response behavior, and dialogue analysis are handled through a user-friendly content management system. Pre-built templates and over 500 actions in the aiWorkflow Builder also enable a quick start with multi-step AI process chains. This ensures that knowledge about the company’s own AI applications remains within the organization—and with it, control.

Data sovereignty is a matter of choice

Data sovereignty is achieved through specific architectural and selection decisions: a technology-agnostic platform with flexible model selection, open standards such as MCP, selectable operating models—including on-premises options—full transparency regarding usage, quality, and costs, integrated compliance, and the empowerment of internal business units. Companies that consistently apply these criteria when selecting a platform future-proof their AI investments—regardless of how the market and regulations evolve. The best time to set the course is now.

AI Impulses for AI-First Companies

The Right Model for Every Task

The Kauz Selection is an intelligent routing system within aiWorkplace that automatically selects the optimal model for each request. Instead of requiring users to manually choose between GPT, Claude, Gemini, or other models, Kauz-Selection analyzes the requirements of a task and decides within milliseconds which model offers the best combination of quality, speed, and cost.

Users no longer have to worry about selecting a model themselves. They simply enter their query as usual, and the appropriate model is selected automatically in the background.

For example, a translation or a short email can be processed by a cost-effective model, while complex contract analyses or demanding research tasks are automatically routed to more powerful reasoning models.