Generative artificial intelligence has burst into corporate environments, promising to automate processes, improve decision-making and facilitate collaboration between teams. However, a recent study published on arXiv (2607.11053v1) reveals a critical limitation: large language models (LLMs) fail in pragmatic cooperation when working with asymmetric information. This finding has profound implications for companies integrating these models into multi-agent systems, customer service chatbots or internal analysis tools, especially in scenarios where participants do not share the same context.
The research formalizes the concept of 'collaborative epistemic asymmetry', linking objective task success with Grice's cooperative principles. In practice, when an LLM acts as an interlocutor in a human or automated team, it often fails to recognize that its counterpart lacks key information, leading to incomplete responses, misunderstandings or erroneous decisions. For example, in a customer service system where the AI agent does not know that the user has already provided certain data, the model may request repeated information or ignore contextual cues, generating frustration and loss of efficiency.
From a technical perspective, LLMs exhibit some pragmatic ability when trained or prompted, but still stumble on Grice's maxims — quantity, quality, relevance and manner — especially when information is partial. This is not a minor failure; it directly affects the reliability of AI-based solutions that many companies are adopting. At Q2BSTUDIO, as a software development and technology company, we understand that implementing AI in collaborative environments requires careful design to mitigate these communication gaps. It is not enough to deploy an LLM; it must be integrated with systems that manage shared context, conversation memory and user expectations.
For organizations seeking to leverage artificial intelligence without falling into these errors, the key lies in custom software development. Generic software rarely captures the particularities of a workflow with asymmetric information. That is why at Q2BSTUDIO we offer custom software services that allow designing LLM pipelines with context verification mechanisms, specialized agents and human escalation routes when pragmatic cooperation fails. Moreover, integration with cloud platforms like AWS or Azure facilitates the orchestration of these agents, ensuring scalability and security.
Cybersecurity also plays a crucial role. When an LLM operates with asymmetric information, it may inadvertently expose sensitive data or generate responses that violate privacy policies. Therefore, at Q2BSTUDIO we combine our AI solutions with cybersecurity and pentesting practices, ensuring that models not only cooperate well but also protect critical information. Likewise, using Business Intelligence tools like Power BI allows monitoring the performance of these systems, detecting patterns of communication failure and adjusting models in real time.
Autonomous AI agents, increasingly popular in process automation, are especially vulnerable to epistemic asymmetry. Imagine a system of agents managing inventory, orders and logistics: if one agent fails to correctly communicate a stock shortage because it assumes another already knows, the supply chain breaks. At Q2BSTUDIO we develop automation solutions that incorporate pragmatic validation layers, ensuring that each agent shares the necessary context without redundancies or omissions. Additionally, the cloud provides the ideal infrastructure to implement these distributed systems, with cloud services on AWS and Azure that guarantee high availability and low latency.
Another area where this limitation manifests is in data-driven decision making. Teams using LLMs to generate reports or BI summaries may obtain misleading results if the model does not consider that the recipient lacks certain background. For example, an executive summary generated by AI might omit key assumptions because it considers them obvious, when in reality they are not for the reader. The Business Intelligence and Power BI solutions we implement at Q2BSTUDIO are designed to include contextual validation layers, ensuring that the presented information is complete and relevant for each audience.
The research also points out that certain failures correlate with unrecognized violations of Grice's maxims. This suggests that current LLMs lack a sophisticated theory of mind to infer what their interlocutor knows or does not know. For companies, this means that blindly trusting an LLM as an orchestrator of collaborative processes is risky. At Q2BSTUDIO we advocate a hybrid approach: combining the power of generative models with business rules, verification systems and human oversight at critical points. Our process automation services are designed to integrate these pragmatic safety layers, minimizing the risks of misunderstandings.
Cloud implementation also allows centralizing information and ensuring that all agents — human or artificial — access the same updated context. For example, by using a shared knowledge base on AWS or Azure, an LLM can query the complete history of interactions before generating a response, reducing asymmetry. At Q2BSTUDIO we help companies design these architectures, from model selection to integration with storage services and message queues.
In conclusion, the failure of LLMs in pragmatic cooperation with asymmetric information is not an insurmountable barrier, but it requires a careful technical and strategic approach. Companies that want to adopt these technologies should do so with the support of experts in custom software development, artificial intelligence, cybersecurity and cloud computing. Q2BSTUDIO offers precisely that: an ecosystem of services that address each of these dimensions, from creating personalized applications to implementing robust and secure AI agents. The key is to recognize the limitation and design systems that compensate for it, not ignore it.
