Enterprise AI Agents Clash Over Meaning Even When Data Matches

Identical Data, Divergent Decisions

Enterprise AI agents that draw from the same data sets can still arrive at conflicting outcomes because raw data rarely includes the business context—processes, policies and priorities that give the numbers meaning. As agents assume more decision‑making roles, those gaps become increasingly difficult to manage.

Three problems emerge

When multiple agents operate without shared business semantics, three issues appear: identical figures lead different functions to different conclusions; agents that work correctly in isolation can combine into the wrong overall outcome; and accuracy degrades as data accumulates.

Why the gap exists

Enterprises have spent years loading ever more records into central data platforms, separating them from the policies, processes and business context where they originated. Irfan Khan, president and chief product officer for SAP Data & Analytics, says this separation is at the root of the problem. “If you access data from a source environment and it doesn’t have the context, the likelihood is you’re going to make an ill‑informed decision,” he explains. “Users learn from history only when institutional knowledge helps them separate signal from noise. An agent starts with no such knowledge unless the company supplies it.”

Traditional data strategies extract records from source systems and centralize them, which strips away the semantics and process context they carried at the source. The shift toward separating storage and compute let companies scale each independently and encouraged many to build multiple data lakes alongside on‑premises estates, flat files and object stores, leading to fragmented data increasingly separated from its original business context.

Coordination failures in multi‑agent workflows

Missing context becomes a coordination failure once agents divide a business process. For example, a procurement agent may negotiate longer payment terms to improve margins, a treasury agent may delay payments to preserve liquidity, and a compliance agent may hold payments until every validation clears. Each performs its assigned job correctly, but together they stall supplier payments long enough that critical vendors start deprioritizing the company, and no internal system flags anything as broken.

Few business processes run on a single agent from end to end. An order entry agent might hand off to one that checks inventory, which hands off to another that files a procurement request or decrements stock. If those steps run out of sequence, one agent can decrement an account before another has confirmed the goods exist, and the task can be completed, but the outcome can still be wrong. “It’s almost like pushing out a task to 20 individual people to go and get it done for a client, and then you realize that nobody held a team meeting,” Khan says. “Nobody kept an inventory of where you were, or where you are within the process.”

Preserving meaning with a business data fabric

Giving agents access to distributed enterprise data solves only part of the problem. They also need a common understanding of how the business expects that data to be interpreted and acted upon. SAP’s business data fabric architecture is designed to preserve that meaning as data moves across systems rather than separating it from its original business context. It connects data and its semantics with the processes, policies, and other business knowledge that govern how the data is used, while allowing agents to reach information across distributed environments.

Khan describes the shared context as “company memory”: the KPIs, operating practices and accumulated business knowledge that a company wants its agents to apply consistently. “It’s almost like having a village elder who has all the historical understanding, and you would consult that village elder to make an informed decision,” he said. “There isn’t such a thing in the agentic world, so let’s consider that to be company memory.”

Embedding constraints and objectives

The goal is to give agents a common frame of reference rather than allowing each to optimize its own piece of a workflow in isolation. An inventory agent, for example, shouldn’t make the locally optimal decision if that decision conflicts with a broader business objective or policy. The shared semantic and policy layer gives each agent a way to determine whether it is still on the right execution path—and when it should fall back rather than act.

That requires encoding constraints as well as objectives. Agents need to understand not only what they are trying to accomplish, but what they must never do and which conditions must be satisfied first, such as confirming goods are available before authorizing a shipment. Within SAP’s business AI platform, the agent harness and agent hub provide the control layer for defining those permitted actions and constraints. The aim is to prevent an individually rational agent decision from producing an outcome that violates policy, governance requirements or the objectives of the larger process.

Leveraging existing enterprise knowledge

The good news is that enterprises don’t have to create all of this context from scratch. Much of it already exists in the systems they use to run the business. Business processes, master data, operational rules and governance policies represent years of accumulated knowledge about how a company actually operates.

By binding that knowledge to the data agents consume, companies can reduce the risk of contradictory decisions and keep AI‑driven workflows aligned with corporate intent.