At SAP Connect 2026, SAP CEO Christian Klein asked, “Why would anyone build agents standalone on an LLM platform?” General-purpose models do not know which supplier will deliver, which controls govern a payment, or how a company handles an exception. “LLMs still have no idea how your business runs,” he said.

SAP presented an architecture intended to close that gap through Joule Work, AI assistants and agents, SAP Business Data Cloud, SAP Knowledge Graph, and SAP Signavio. The company is competing to become the layer that supplies business context, coordinates work, and captures operational learning across applications. CIOs should evaluate that ambition through four decisions.

Decide The Type Of Work Intelligence That SAP Should Supply

Enterprise applications traditionally combined screens, workflows, business logic, and records. SAP’s architecture distributes these functions. Joule Work becomes the engagement layer, assistants organize work by role, and agents execute across finance, procurement, supply chain, HR, and customer experience. Applications continue to record transactions and enforce controls. If the architecture works as intended, users will complete more work without opening them.

SAP is betting that executable process knowledge will remain scarce as interfaces and agents become cheaper to create. Klein said the SAP Knowledge Graph maps over 7 million fields, half a million tables, more than 50,000 APIs, and roughly 400 data products. These vendor-reported figures indicate scope. They do not prove semantic accuracy inside a customer environment or demonstrate business value. Their significance is that SAP is attempting to convert decades of application structure into machine-readable business context for agents. This reflects a broader shift in which process intelligence can help ground and govern AI agents, although our research finds a gap between vendor positioning and productized capabilities, accessible data, and meaningful customer adoption.

Use SAP’s context where its transactional, regulatory, or industry knowledge is differentiated, but we recommend that tech leaders keep company-specific definitions, policies, and decision criteria under enterprise control. Apply Forrester’s AEGIS framework to define enterprise guardrails across governance, identity, data, application security, threat operations, and Zero Trust before agents can execute business transactions.

SAP says its gateway allows other AI experiences to call its agents and Joule Work to invoke registered third-party agents. That establishes connectivity as a stated capability, not portability. Connectivity lets customers integrate an agent into another environment. Portability lets them replace it without losing capability, history, or control.

Decide The Context Debt That Deserves Repair

Every complex enterprise carries an accumulation of missing, conflicting, or inaccessible business context that prevents work from completing correctly. We call this context debt.

A 20-year-old SAP ECC estate exposes context debt through custom code, duplicate records, and local process variants. An S/4HANA estate faces the same problem when a process depends on CRM data, specialist services, or controls documented outside the workflow.

COO Sebastian Steinhaeuser warned customers against paying “incredible token bills for something that they could have fixed at the database level.” He also told product teams, “Don’t just encode the last 5% of missing automation in your backlog into an agent and call that an innovation.”

Use agents to expose the debt that matters. Classify failed executions as data defects, process defects, integration gaps, authorization problems, model errors, or cases requiring human judgment. Döhler reported that incomplete and duplicate master data constrained its developing no-touch sales order process. A Novartis procurement leader advised companies to start where a wrong decision is recoverable. The first example reflects operating experience; the second came from an early pilot. Neither establishes repeatable marketwide outcomes.

Repair a defect when it repeatedly blocks valuable work or creates material control risk. Do not turn every agent failure into a migration program. Forrester recommends that technology leaders govern enterprise resource planning modernization as a continuing innovation system, addressing integration and testing debt while keeping analytics, orchestration, and AI replaceable.

Decide Who Owns What Agents Learn

The richest process knowledge often appears when standard execution breaks down. A planner overrides a recommendation, a buyer selects a higher-priced supplier, or a controller resolves an unusual posting. Transaction systems record the result but often lose the reasoning.

Agent-mediated work can capture parts of that reasoning through corrections, overrides, and conversations. SAP described a planned company-memory capability that can ingest process documents, policies, and execution information, then provide distilled context to agents. SAP executives discussed future approaches for using observed interactions to refine rules and exception handling. Together, these capabilities could create a learning loop in which human corrections influence future execution. Agents curate knowledge from enterprise artifacts and events while people contribute intent and judgment. The governance question is about who controls the resulting knowledge and whether it remains usable outside the originating platform.

This is an architectural risk, not a proven source of lock-in. Accumulated corrections, evaluations, and decision traces would be expensive to recreate if customers cannot export them in reusable form.

Require access to prompts, corrections, overrides, evaluations, and resulting memory artifacts. Define whether employee interactions can improve customer-specific agents or broader models. Test exportability before production deployment. An exit right is hollow if a replacement agent must relearn years of exceptions.

Decide Who Measures Value And Bears Failure Risk

SAP is moving premium agentic capabilities toward consumption pricing, including charges linked to agent actions, while retaining AI units as its commercial currency. It is also developing business evaluations, value calculators, and SAP Signavio agent mining to connect activity with process results.

SAP could operate the agent, provide the benchmark and calculation tools, and meter usage. Customers still absorb retries, human rework, and failed execution. The deeper governance question is who defines successful work. Organizations should approve the baseline and reproduce the calculation independently.

Separate the value meter from the billing meter. This separation is increasingly important as consumption pricing and workflow orchestration shift leverage toward software vendors and make usage, dependency, and business outcomes harder to evaluate independently. Define successful completion before deployment. Require raw execution, exception, intervention, and consumption data. Scale when completed outcomes improve faster than consumption and exception handling declines.

SAP is assembling the components required to become the orchestration layer for enterprise work. The unanswered question is how much control customers are willing to transfer to it and whether the resulting process knowledge, learning history, and economic signals remain portable and extendable for a further where the agents will go across vendor ecosystems.

Turn Your SAP AI Strategy Into A Defensible Decision

Forrester clients can schedule an inquiry or guidance session to evaluate a proposed SAP agent use case, pressure-test its modernization and consumption economics, and define the governance, learning ownership, and exit rights needed before deployment.

Bring one candidate process, its current performance baseline, and the applications and data sources it crosses. The session will help establish deployment conditions, evidence requirements, and a clear scale-or-stop decision.

Share