Agents need more than business process knowledge. They need to understand how your company actually works

Presented by SAP

Knowing standard finance, procurement and supply chain processes isn’t enough for AI agents to execute them reliably. They also need to understand a company’s configurations, policies, data relationships and exceptions.

That requires an understanding of both the business objects and relationships an agent acts on plus the rules and controls governing what it may do, says Ted Way, VP and chief product officer of SAP Business AI product engineering.

“An agent can understand global accounting practices like IFRS 15 or US GAAP,” Way says. “But to actually execute the work, it needs to understand that company’s revenue recognition configuration, the calculations, posting history, document flow and contracts.”

An agent also must account for company-specific process steps, business exceptions and any industry-related regulatory requirements; these factors determine whether a journal entry or payment is executed correctly. As companies increase agent autonomy within auditable workflows, the context those agents must understand grows with them.

What AI agents need before they execute multistep enterprise workflows

An agent can either produce answers or act and deliver outcomes. Doing the latter requires access to business context combined with governance.

“The base agent may only be able to tell the user, ‘I think you should do X,’” Way says. “We’re moving to an agent that can say, ‘I’m authorized to do X. I know which systems are involved. I can do it while preserving the controls around the process.’”

For a single payment, the business objects an agent must track can include the customer, supplier, product, contract, invoice and order, while the controls span approvals, access rights, compliance requirements and an audit record of every action the agent takes. The agent first reports the approval, then recommends a payment, then prepares the payment for a person to execute, and finally makes and reconciles the payment itself under strict policies with a person in the loop, a stage Way calls “governed execution.”

The potential payoff is not simply a better answer, but less work between the answer and the outcome. For customer PwC, SAP says Business AI reduced a complex billing task from 35 minutes to five minutes while improving accuracy.

AI agents need reconciled data across applications

Most of the context an agent needs sits scattered across applications. For instance, the same supplier can exist under different identifiers and attributes, so one record might say IBM, another International Business Machines and a third refers to IBM Inc.

“In my company, ‘critical supplier’ may mean something completely different than in your company,” Way says. “That’s why the ontology, the knowledge graph, and data harmonization matter so much. We definitely do not want our strategy to be, ‘We hope we get lucky and the LLM figures it out.’”

Giving an agent access to more data isn’t enough; it also needs to understand what the data means and how the records relate. The answer is a context layer above the raw data, which reconciles each entity’s scattered records into one trusted profile and tells the agent how that entity connects to others.

Keeping AI agents’ business context current

Enterprises add data and change processes constantly, so the context an agent relies on can quickly become stale. Keeping it current requires continuously reconciling records and determining which sources are most trusted and up-to-date. It also means monitoring how business processes and agentic workflows actually run.

When an agent reaches a threshold it shouldn’t cross alone, it sends a notification for review or approval. Customers define those thresholds and the rules the agent follows. For regulated finance processes, they also can require fixed workflows for steps that must be executed the same way every time.

“Some steps have to happen exactly the same way every time, for compliance or regulatory reasons, and we don’t need an agent to be creative there,” Way says. “Where there is room to solve a problem, it can be. Customers draw that line.”

Context must follow the work

Company-specific context becomes most useful when it can travel with a process across applications, systems and functions. That is the idea behind SAP’s new and expanded Joule Assistants, which coordinate specialized agents around a business objective rather than a single piece of software.

“Joule Assistants follow the business process, not the org chart of the software,” Way says. “Each assistant orchestrates multiple Joule Agents, and each agent executes a defined process across SAP and non-SAP systems.”

Consider a logistics exception. If the inventory needed for an order isn’t where it should be, resolving the problem requires more than identifying another source. An agent may need to know what inventory is available across warehouses, whether it has already been committed to another order, which expiration rules apply, what alternative transportation is available and which decisions require approval.

Way says one SAP customer that supplies natural ingredients to manufacturers used this approach to turn what had been a multistep logistics “fire drill” into a one-click process, recovering roughly a third of the team’s time.

The same principle applies in finance. SAP’s Revenue Recognition Assistant uses a company’s own configurations, calculations, posting history and document flow when categorizing transactions under IFRS 15 and US GAAP. SAP Pay, powered by Tereina, an SAP company, takes the model further into governed execution, making and reconciling an approved payment while preserving the controls around the process.

The impact can extend beyond automating individual tasks. SAP says Daimler Trucks North America increased its fleet contract bid win rate from about 10% to more than 40%, generating roughly $70 million in financial impact over 12 months.

Underneath these workflows, Joule draws on reconciled business entities and semantic relationships through SAP Reltio in SAP Business Data Cloud and SAP Knowledge Graph. SAP Reltio continuously reconciles records using what it calls dynamic survivorship to weigh source trust, recency and aggregation, while SAP Signavio can track business processes and agentic workflows as they run.

“Customers will say, ‘Oh, we have one procure-to-pay process,’” Way says. “We run Signavio and say, ‘Actually, you have 400 procure-to-pay processes.’”

Business context has to cross vendor boundaries

The problem becomes harder when a workflow moves beyond a single vendor’s applications and agents. SAP is using the Agent2Agent (A2A) protocol to allow Joule to interact with agents outside SAP; Way says its bidirectional Microsoft Teams integration is a terrific example.

But the larger issue is not simply whether two agents can communicate. The business context that makes an action trustworthy — the meaning of the data, relationships between entities, company policies and controls — also has to remain intact as work crosses systems.

Way acknowledges that few companies operate entirely within one technology stack. That makes interoperability more than an integration issue. If agents are going to execute work across SAP and non-SAP systems, the business meaning, relationships, policies and controls they rely on must travel with them.

The challenge for enterprise agents, then, isn’t simply learning how a business process is supposed to work. It’s staying grounded in how this particular business actually works, even as that business changes.

Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact sales@venturebeat.com.