From answers to actions
An agent that recommends a payment needs less authority than one that prepares or executes it. Ted Way, SAP Business AI product engineering’s vice president and chief product officer, describes a progression from advice to supervised execution: the agent learns which systems are involved, receives permission to act and follows constraints set by the company.
A single payment may depend on the customer, supplier, product, contract, invoice and purchase order. Controls include approvals, access rights, compliance requirements and an audit record of every action.
The process can move through stages: the agent reports an approval, recommends a payment, prepares it for a person to execute, and eventually executes and reconciles it under strict rules while keeping a human in the loop. Way calls this supervised execution.
The case for automation is that it can reduce work between an answer and a completed task. SAP says PwC Business AI cut a complex invoicing operation from 35 minutes to five while improving accuracy.
The context problem is data, not just access
The information an agent needs is often spread across applications. A supplier might appear as IBM in one system, International Business Machines in another and IBM Inc. in a third. Even the meaning of “critical supplier” can vary by company.
Way argues that agents need an ontology, a knowledge graph and reconciled data; a large language model cannot be expected to work out the right interpretation on its own. A context layer can bring separate records about an entity into one verified profile and show how it connects to other business objects.
That layer cannot be treated as a one-time setup. Companies add data and change processes, so the context can quickly become stale. Keeping it useful means reconciling records continuously, weighing source reliability and freshness, and tracking how processes and agent workflows actually run.
Companies also set thresholds for when an agent must stop and request review or approval. For regulated financial work, they can require a fixed sequence of steps. Where compliance demands a prescribed order, Way says, the agent should not improvise; where there is room to solve a problem, discretion can be useful.
Context has to cross company boundaries
SAP’s newer and expanded Joule Assistants are designed to coordinate specialized agents around a business task rather than a single software product. Each assistant coordinates multiple Joule agents, which perform particular processes across SAP and other vendors’ systems.
A logistics disruption shows why that distinction matters. If an item is unavailable where expected, an agent may need to check stock at other warehouses, whether it is reserved for another order, shelf-life rules, alternative transport and which decisions require approval. SAP says one customer that supplies natural ingredients to manufacturers turned a multistep logistics emergency into a process started with one click, returning about a third of the team’s working time.
SAP cites other results from applying company-specific context:
Joule uses aligned business entities and semantic links through SAP Reltio in SAP Business Data Cloud and SAP Knowledge Graph. SAP Reltio continuously reconciles records, weighing source trust, data freshness and combined information. SAP Signavio can track business processes and agent workflows. Way offers a telling example from procurement: a company may think it has one procure-to-pay process, while Signavio analysis finds 400.
My read is that the strongest part of SAP’s case is also its least glamorous: agent autonomy depends on the quality and upkeep of the company’s shared context. The announcement says less about how organizations will judge whether that context is complete enough for an agent to act, or how they will manage disagreement between systems when it is not.
The task must carry its rules with it
The challenge grows when a workflow crosses providers. SAP uses the Agent2Agent (A2A) protocol so Joule can interact with agents outside SAP; Way points to two-way integration with Microsoft Teams as an example.
Communication alone is not enough. As a task moves between systems, its data meanings, entity relationships, company rules and controls must move with it. Way acknowledges that few companies operate within a single technology stack. In that environment, interoperability is not just a matter of connecting agents; it is whether the context that makes their actions reliable survives the handoff.
The harder test for enterprise agents will be whether companies can keep that context current across systems—and preserve the boundaries that make delegated action safe.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X