SAP Outlines Four Strategic Decisions for CIOs Before Deploying AI Agents
October 11, 2026 · 3 min read
Also published in Türkçe
At the SAP Connect 2026 event, SAP Chief Executive Officer Christian Klein challenged the prevailing approach of building isolated artificial intelligence agents on top of large language models. The company argued that general-purpose models lack the specific knowledge required to identify which vendor fulfills an order, what controls govern a payment, or how a specific enterprise handles exceptions. In response, SAP unveiled a new architecture designed to bridge this gap and placed four critical strategic decisions on the agenda for chief information officers before they deploy autonomous agents into production environments.
The proposed architecture integrates several existing components, including Joule Work, SAP Business Data Cloud, SAP Knowledge Graph, and SAP Signavio. This combination aims to function as a unified layer that provides business context, coordinates work across different applications, and captures operational learning. Within this framework, Joule Work serves as the engagement layer, while assistants organize tasks by function and agents execute specific duties in finance, procurement, supply chain, human resources, and customer experience. Klein stated that the SAP Knowledge Graph currently maps more than seven million fields, half a million tables, over 50,000 application programming interfaces, and approximately 400 data products. However, the advisory firm Forrester noted that these figures, provided by the vendor itself, indicate scope but do not prove semantic accuracy within a client environment or demonstrate actual business value.
The first strategic decision requires organizations to define exactly which process intelligence the software provider should supply versus what remains under corporate control. Forrester advises that chief information officers must retain authority over company-specific definitions, policies, and decision criteria before authorizing agents to execute transactions. The second decision involves identifying and repairing context debt, which refers to the accumulation of missing, conflicting, or inaccessible business context that prevents work from being completed correctly. Sebastian Steinhaeuser, Chief Operating Officer of SAP, warned clients against paying excessive token fees for issues that could have been resolved at the database level. He also directed product teams not to code the final missing percentage of automation into an agent and label it innovation. Real-world examples highlighted these challenges, with the Döhler Group reporting that incomplete and duplicate master data limited its sales order processes without human intervention, while a procurement leader at Novartis recommended starting automation in areas where incorrect decisions remain reversible.
The third decision concerns ownership of the knowledge accumulated by agents during operation. SAP described a planned corporate memory capability designed to ingest process documents, policies, and execution information to provide condensed context to agents. Corrections, substitutions, and conversations generated during agent-mediated work could create a learning cycle where human interventions influence future executions. Forrester identified this accumulation as an architectural risk, noting that corrections, evaluations, and decision logs would be costly to recreate if clients cannot export them in a reusable format. The fourth decision addresses who measures generated value and who absorbs the risk of failure. SAP is moving premium agentic capabilities to a consumption-based pricing model with charges linked to agent actions, keeping artificial intelligence units as the commercial currency. The company is also developing value calculation tools and agent mining features via SAP Signavio to connect activity to process outcomes. Despite these tools, clients continue to absorb the costs of retries, human rework, and failed executions. Forrester recommends separating the value meter from the billing meter and demanding raw data on execution, exceptions, intervention, and consumption before deploying to production.
SAP is assembling these components to position itself as the orchestration layer for corporate work. The unresolved question, according to the analysis presented at the event, is how much control customers are willing to transfer to the company. It remains uncertain whether process knowledge, learning history, and resulting economic signals will remain portable and extensible in ecosystems featuring multiple agent providers.