Automated conversation systems experienced a technical acceleration in 2024 across several simultaneous fronts. On the language model side, updates focused on the ability to handle real-time voice exchanges directly from a browser. On the regulatory front, the European regulation on artificial intelligence introduced transparency obligations that alter the very design of conversational assistants intended for the European market.
These two dynamics, technical and legal, redefine what companies can expect from their chatbots and voice assistants. They also raise open questions about the current limits of these technologies.
Real-time voice conversation: what language models are changing concretely
The most tangible advancement of 2024 concerns voice conversation. Large language models (LLMs) have integrated real-time oral exchange capabilities, accessible from a simple web browser. The latency perceived by the user has decreased to the point of making the interaction close to a traditional phone conversation.
This evolution is not limited to voice synthesis. Systems now handle conversational context over multiple turns, recognize certain nuances of tone, and adapt their responses based on the history of the exchange. For companies deploying customer assistants, this represents a change in usage: the voice assistant becomes an autonomous channel for processing requests, not just a simple complement to text.
A detailed overview of these technical advancements and their recent applications is available, as can be read on CN Blog in an analysis dedicated to the topic.
The open question concerns the reliability of these voice exchanges in high-stakes contexts: medical appointment scheduling, legal assistance, technical support for complex equipment. Public data on error rates in real conditions remains scarce, and most feedback comes from the publishers themselves.
Article 50 of the European AI regulation: transparency obligations for chatbots
The European regulation on artificial intelligence introduced, via its Article 50, a transparency obligation that directly concerns automated conversation systems. People must be informed when they interact with an AI system, unless this is obvious in the context of use.
This obligation only applies from August 2, 2026. In 2024, it therefore already existed in the regulatory text, but was not yet binding. Companies that anticipated this deadline modified their chatbot interfaces to display a warning at the beginning of the conversation.
Beyond the simple warning
Article 50 is not limited to signaling the presence of a chatbot. It also provides for transparency requirements for certain content generated or manipulated by artificial intelligence, particularly deepfake audio and video. A conversational assistant capable of producing or transmitting multimedia content potentially falls under this provision.
The regulation also distinguishes the responsibilities of technology providers and those of the organizations deploying them. A company using a chatbot developed by a third party remains responsible for the information provided to its users. This sharing of responsibility creates a zone of vigilance for outsourced customer services.
Multimodal interactions: text, voice, image in the same exchange
The other significant development axis concerns multimodality. The conversational systems of 2024 no longer rely solely on text. They integrate into the same conversation thread:
- Natural language voice exchanges, with context and tone recognition
- Analysis of images sent by the user (photo of a defective product, screenshot of an error)
- Generation of visual or video content in response to a request (explanatory diagram, demonstration)
This convergence opens up concrete applications in technical support, online commerce, and training. A customer can photograph a broken device, send the image to the chatbot, and receive a diagnosis accompanied by a troubleshooting video.
Multimodality transforms the chatbot into a unified interface rather than just a messaging channel. Current limitations concern the quality of image analysis under variable conditions (lighting, angle, resolution) and the computational cost that these processes entail for companies.
Data and conversation analysis: what companies are really leveraging
Automated conversation systems generate a considerable volume of conversational data. In 2024, associated analysis tools progressed on two fronts: real-time intent detection and extraction of weak signals from thousands of exchanges.
Companies specializing in customer relations, such as Armatis, deploy artificial intelligence at several stages of the process: recruitment of advisors, training, conversation analysis, and real-time assistance. AI does not replace the advisor but changes their workstation, providing response suggestions and contextual alerts during the exchange.
Known limitations of conversation analysis
Automated conversation analysis faces several documented obstacles:
- Language models may misinterpret irony, sarcasm, or culturally marked formulations
- The aggregation of conversational data raises privacy questions, particularly regarding GDPR
- Satisfaction metrics derived from automatic analysis do not always correlate with self-reported surveys from customers
Missing data primarily concerns the comparison between performance perceived by the user and that measured by analysis tools. Independent publications on this topic remain few.
Automated conversation systems in 2024 have crossed a technical threshold on voice and multimodality while entering a European regulatory framework that imposes new design constraints. The challenge for companies is no longer to adopt these technologies but to master the conditions of use: reliability of responses, legal compliance, responsible use of conversational data.
