AI voice bots: opportunities and risks, and a revolutionary new voicebot
VOICE-BASED AI BOTS: OPPORTUNITIES AND RISKS, AND AKQUINET’S REVOLUTIONARY NEW VOICEBOT
Voice-based support systems are currently experiencing a significant surge in development. More and more interactions that previously had to be handled by staff are now being taken over by digital voice assistants. Whether in call centres, technical support or information desks, voicebots offer the potential to automate communication whilst improving the quality of customer service processes. Further areas of application will soon follow, for example in humanoid robotics: assistants equipped with voice capabilities can accompany tourists in cities or support the care of people in need of assistance, combining physical strength with linguistic skills.
AI VOICEBOTS TODAY: RARELY LIVING UP TO THEIR PROMISES
Despite these attractive prospects, the reality of many voicebot projects still falls short of expectations. Companies link their implementation to clearly defined objectives: increasing customer satisfaction, reducing the workload on staff, improving process efficiency, or even reducing staff turnover in high-pressure areas of telephone customer service. However, when standard voicebots are confronted with real, organisation-specific requirements, fundamental technical limitations quickly become apparent – something the AKQUINET team has observed time and again during consultations. Why does this happen?
One of the main reasons is that many systems rely solely on large language models optimised for general conversation. These work well for small talk or when answering general questions. However, as soon as access to specific process data is required or an action needs to be triggered on the system side, functionality breaks down or becomes severely limited. For example: if a customer wants not only to find out when her order will arrive, but also to change the delivery date, many voicebots are unable to handle such a scenario. Some of them rely on rigid, static lists of information or instructions, but the depth and up-to-date nature of this data are insufficient.
HALLUCINATIONS: WHEN AI PREFERS TO MAKE SOMETHING UP RATHER THAN NOT ANSWER
This leads to a second, equally significant risk: hallucinations. Large language models tend to generate responses that sound convincing but are inconsistent with reality. A 2023 study shows that nearly 20 per cent of all responses generated by LLMs contain hallucinations (see HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models, October 2023). Of course, language models are evolving very rapidly. Even if it is difficult today to unequivocally determine the current level of hallucinations in widely used LLMs, one thing is certain: hallucinations do occur, and almost every user of language models encounters them. The consequences can be serious – erroneously ‘invented’ system statuses, purported procedural decisions or false information regarding contract terms can mislead customers and pose a real risk to businesses.
AI IS SUPPOSED TO HAVE ACCESS TO DATA – BUT HOW?
Added to this is the issue of the technical integration of systems. A voicebot’s direct access to operational systems raises a number of problems. On the one hand, it increases the risk of unintended AI interference with production data sets. On the other hand, it often leads to performance bottlenecks. ERP and ticketing systems are not designed to handle the additional load generated by AI queries. This results in pauses in the conversation lasting five seconds or longer, which is absolutely unacceptable in the context of a telephone call. Operational experience also shows that rigid voicebot architectures quickly reach their limits. As soon as requirements change – for example, when processes are modified or new topics are added – traditional voicebots prove difficult or costly to expand.
HOW THE AKQUINET AI VOICEBOT WORKS
To meet these challenges, the AKQUINET team’s new technical concept is based on three interlinked principles:
- Ensuring reliable responses
- Secure storage of feedback
- Precise decoupling of system components
What have we done?
The central element of the AKQUINET AI voicebot is the introduction of an intermediate layer – the so-called ‘staging area’. Corporate systems no longer transmit data directly to the AI, but to this intermediate layer, which is constantly synchronised in real time. The voicebot accesses only this layer. As a result, the AI model does not interfere with operational systems in an uncontrolled manner, whilst still having access to up-to-date data. This approach is complemented by short-term memory based on a retrieval mechanism. Unlike traditional long-term training, the AI receives the necessary information only within a specific context and then discards it. This reduces the workload whilst minimising the risk of hallucinations, as the AI always operates on precise, context-specific data.
AI NEVER INTERACTS DIRECTLY WITH THE SOURCE SYSTEM
Another key element is the way in which actions are triggered. The AI does not make changes to systems on its own. Instead, it has pre-defined tools at its disposal – for example, functions or automated process modules. The voicebot decides which tool is required, launches it, but never communicates directly with the source systems. This ensures that security requirements are met and that all operations are fully auditable.
The key advantages of this approach are:
- a reduction in errors thanks to a clear separation of access to data and action logic,
- full auditability of all processes executed,
- stability even with complex queries.
We have also further developed the communication itself. The AKQUINET AI voicebot utilises a real-time audio model. In this case, speech is not first converted into text, but is processed directly. Instead of a speech-to-text and then text-to-speech sequence, we therefore have direct speech-to-speech. This results in more natural dialogues, faster response times and greater resilience to accents, dialects and different pronunciation variations. A fallback mechanism ensures reliability – should the audio model fail, the system automatically switches to a predefined text-based language model and continues the dialogue, with a slight delay but without interrupting the connection.
ARCHITECTURE IS CRUCIAL FOR LLM INTEGRATION
Integration with existing IT landscapes is equally important. The AKQUINET AI voicebot operates entirely on the Azure platform. If deployed within an existing Microsoft tenant, company data never leaves the organisation’s infrastructure. At the same time, companies benefit from the stability and extensive documentation of standardised technologies.
In summary, it is becoming increasingly clear that the success of a voicebot depends less on the size of the language model used and more on the solution’s architecture: how data is made available, how uncontrolled AI behaviour is prevented, how responsiveness is ensured, and how the system maintains flexibility in a dynamic corporate environment. The approaches described demonstrate that efficient voicebots can only be created when technological capabilities are closely aligned with operational requirements. They represent an important step towards secure, scalable and genuinely useful voice assistance systems in day-to-day business. Thanks to the AKQUINET AI voicebot, these steps have already been planned with your business in mind.