Making the case for context-aware AI

Enterprise AI systems can perform tasks like generating recommendations, automating processes, and aiding decision-making. However, these systems often lack operational context—an understanding of how a business actually functions in practice. Without this context, even advanced AI models may produce recommendations that overlook critical constraints, such as workflow dependencies, resource limitations, or regulatory requirements. For example, an AI might suggest a financial forecast without considering a company’s cash flow restrictions, leading to costly errors. Devavrat Shah, chief scientist of enterprise AI at Celonis, emphasizes that AI agents must grasp the role of each task within broader business processes to avoid inefficiencies or mistakes. This gap between AI capabilities and real-world business operations is why many organizations struggle to scale AI beyond experimental phases.

Context in AI refers to the background information, constraints, and relationships that define how a business operates. Shah introduces the concept of decision intelligence, which combines AI with an understanding of business processes to improve decision-making. For instance, an AI agent tasked with approving purchase orders must know the company’s spending limits, supplier relationships, and approval hierarchies. Without this context, the agent might either delay decisions by overanalyzing or approve requests that violate policies. Shah argues that context enables AI to act as a role agent—a specialized assistant that handles specific tasks within a defined process. This approach ensures AI recommendations align with business goals and operational realities, reducing errors and wasted resources.

To provide AI with operational context, organizations must connect it to existing business systems, such as workflow tools and event logs. Celonis, for example, uses a context model—a digital layer that abstracts and organizes process data from systems like ERP or CRM software. This model serves three purposes: it extracts the right process details for AI agents, defines guardrails (rules and constraints) for those agents, and enables seamless integration with other AI systems. The goal is to create a real-time digital twin of the organization, a virtual representation of its processes that AI can interact with. For instance, if a supply chain disruption occurs, the context layer would help AI agents adjust procurement plans based on the company’s inventory levels and supplier contracts.

Deploying AI without operational context poses significant commercial risks. Shah highlights two primary dangers: excessive resource consumption and costly mistakes. Without context, AI agents may waste computational power trying to infer missing information, performing tasks inefficiently. More critically, they might overlook key constraints, such as legal requirements or financial thresholds, leading to decisions that harm the business. For example, an AI recommending a marketing campaign might ignore budget constraints or compliance rules, resulting in wasted ad spend or regulatory penalties. Shah notes that the impact of such errors is amplified in enterprise settings, where decisions often affect large-scale operations, finances, or workforce management.

While AI agents can automate repetitive tasks, human oversight remains essential for strategic decisions. Shah explains that AI excels at handling operational decisions—frequent, rule-based choices like approving invoices or scheduling shifts—but humans must guide strategic decisions, such as long-term investments or market expansions. The shift will redefine roles: employees will spend less time on manual tasks and more on analyzing AI outputs, validating recommendations, and making high-stakes choices. Organizations must invest in workforce education to ensure employees can interpret AI insights and understand their implications. Shah stresses that this transition requires thoughtful planning to avoid resistance or misuse of AI tools.

Organizations can start adopting context-aware AI by identifying a specific business decision to improve, such as optimizing inventory levels or streamlining approval workflows. Shah recommends breaking down the associated process into steps, identifying the most challenging or time-consuming parts, and testing whether AI can enhance them. Metrics like speed improvements or cost savings should be used to evaluate success before scaling the solution. Governance plays a key role in this process by defining clear rules for AI agents, such as what actions they can or cannot take. Additionally, companies must prepare their workforce through training programs to ensure employees can collaborate effectively with AI systems.