Celonis Chief Scientist Devavrat Shah argues that enterprises must shift from probabilistic language models to deterministic graphical architectures for complex numerical operations.
Enterprises are increasingly encountering the limitations of Large Language Models when applied to structured, high-stakes numerical forecasting and operational planning. Devavrat Shah, the Chief Scientist for Enterprise AI at Celonis and a chaired professor at MIT, posits that the industry has conflated distinct AI architectures, leading to the deployment of probabilistic text-based models for tasks requiring deterministic, relationship-aware computation.
Shah joined Celonis following the strategic acquisition of Ikigai Labs, bringing his specialized research into Large Graphical Models (LGMs) to the vendor’s operational platform. He argues that while LLMs excel at tokenized, sequential text processing, they lack the native capability to model the complex, multi-dimensional relationships inherent in enterprise data cubes. LGMs are designed to learn the structural dependencies between disparate datasets, effectively fusing information across time-series and tabular structures without requiring predefined schemas.
The fundamental issue, according to Shah, is that organizational data often exists in silos where relationships remain implicit and undefined. Graphical models provide a mechanism to map these connections, enabling the platform to perform rigorous what-if scenario modeling. This architectural distinction allows Celonis to position its technology as a neutral, numbers-native layer that operates beneath the conversational interface of standard LLMs.
The shift toward specialized architectures reflects a broader market trend where vendors move away from monolithic AI approaches. Companies like ServiceNow and Blue Yonder are similarly implementing layered models, utilizing domain-specific engines for computation while reserving LLMs for orchestration and natural language interaction. Celonis aims to differentiate its platform by applying this logic horizontally across the enterprise rather than limiting it to specific functional silos.
Shah emphasizes that the unit of deployment should be the functional role rather than the individual task. By focusing on a unit of decision-making, organizations can implement incremental improvements rather than attempting wholesale, high-risk transformations. This role-based approach allows for the modular integration of AI, where specific forecasting or planning functions are supported by the LGM architecture while maintaining human oversight.
Demand planning serves as a primary example of this methodology, where historical data, analyst adjustments, and strategic goals must be reconciled. Shah notes that each stage of this process functions as a distinct role, allowing for targeted automation and optimization. The integration of these roles into a cohesive context model enables a more stable and reliable output than a general-purpose language model could provide.
The industry is arriving at this conclusion through iterative trial and error, discovering that probabilistic models struggle with the heavy, structured work of enterprise optimization. Shah characterizes the resulting system as an enterprise world model, where the LGM handles the underlying computation, and language models manage the interface between human users and autonomous agents. This layered strategy ensures that the business logic remains grounded in the actual operational context of the organization.
The significance of this transition lies in the move toward deterministic reliability in enterprise software. By separating the reasoning layer from the computational layer, organizations can mitigate the hallucinations and inaccuracies often associated with large-scale probabilistic models. This separation of concerns is essential for high-stakes environments where numerical precision is a prerequisite for operational integrity.
Stakeholders should view this as a maturation of the enterprise AI market, moving beyond the initial hype cycle toward specialized, fit-for-purpose architectures. The integration of LGMs into the broader Celonis operational ontology suggests that the future of enterprise automation will be defined by hybrid systems. These systems leverage the strengths of different architectures to solve problems that were previously intractable with a single-model approach.
Future deployments will likely hinge on the ability of enterprises to define these functional roles with precision. As organizations move toward this architecture, the focus will shift from the novelty of the model to the reliability of the underlying decision-making framework. The success of such systems depends on the ability to reconcile complex, evolving data structures while maintaining the flexibility required for strategic business adjustments.
