"So we draw worlds and fit them like tracings against the world about us, and crumple them when we find they do not fit and draw new ones." - John Steinbeck, The Log from the Sea of Cortez (1951)
Steinbeck was writing about science, but he was describing something universal: the moment a model of reality meets reality itself and fails. Enterprises have been drawing their operational world for decades, in the language of process documentation, policy manuals, and institutional knowledge passed between people who already generally understood the context because of our capacity for human judgment. That tracing worked fine when the intelligence reading it was human. A new class of natural language-processing system is now in the room, and it does not carry the experiential human grounding that gives language meaning, nor does it reliably process implications. Large Language Model-based systems are probabilistically capable across many tasks, but if we want them to participate in support of human judgement, the tracing needs to be redrawn.
This matters more than almost anything else in the current AI conversation. Automation capability has become abundant. The barriers to deploying an LLM agent that can retrieve information, draft a document, or summarize a meeting are essentially gone. The barriers to deploying an agent that can recommend a course of action, coordinate work across systems and people, execute a transaction, or govern a process at scale remain significant. And they are not technical barriers in the model capability sense. They are representational architecture barriers in the enterprise sense. What organizations have is a representation problem and therefore, deploying AI is a context problem.
This is the first post in a series that works through that problem agent by agent, layer by layer. It builds on two primary ideas. First, the idea that context is not a single thing but is a stack of various abstracted representations of enterprise artifacts that need to be governed as a complete system. The second is a recognition of the judgment gap between human intelligence and machine execution that must be compensated for. It is concerned with the space between what an agent can technically be programmed to do and what an organization can safely authorize. This post synthesizes both into a framework for understanding why enterprise AI efforts hit a scaling wall, and what it takes to break through it into more meaningful and valuable use cases.
The Scalability Chasm
The first two questions to ask are not about agents or models. They are about what kinds of work need to remain distinctly human and what parts of work an agent can safely perform.
There are generally seven classes of enterprise work. Two are scaling now: Retrieve and Explain, and Create and Transform. One is emerging at scale: Analyze and Diagnose. Four remain hard to scale for most enterprises: Recommend and Decide, Coordinate and Orchestrate, Execute and Transact, and Govern and Assure. This is where impressive pilots fail at scale.
The dividing line between these two groups is not model capability. It is a different kind of line entirely.
Above the line, the agent is primarily descriptive and acts as a delegate. It explains what happened, produces a draft, or identifies where reality diverged from expectation. A human is still in the loop before anything changes. Errors are visible and correctable once verified. Accountability remains with the human. The agent's output is an input to a human decision, not a decision itself.
Beneath the line, the work requires outputs that direct judgment, coordinate work, or change enterprise state. An agent is no longer providing outputs for consideration by a human. It is committing the enterprise to a course of action, routing work to people who will act on it, or participating in assuring compliance against a standard. The human review loop either narrows dramatically or disappears. And this is where the scaling wall appears.
I call this the Scalability Chasm. It is the point where descriptive context in the form of prompting or "RAGging" is no longer enough and enterprise intent must become explicit, centrally governed, and machine-usable.
The Judgment Gap
For AI investments to realize a valuable return, companies must cross the chasm. They cannot rely on primarily asking machines to retrieve information. They need agents that can act, not only report. Current efforts fail because operating across only fragmentary cognitive scaffolding opens the door to silent and invisible agent drift. I am using the term cognitive scaffolding instead of context because context is always in service of cognition and the highest value potential of agentic AI lies in producing outputs that enable authoritative organizational action. This is why context is important, but it is not entirely enough.
Human judgment draws on layered cognition: language, semantics, accumulated and lived experience, situational awareness, and something closer to practical wisdom; the ability to know what the right action is in a specific and fluid context, given real constraints, competing priorities, and organizational accountability. Much of what constitutes human judgment is also not fully capturable in data fed to LLMs. Machines currently simulate language form and semantics from linguistic symbols very well, but do not reliably replicate the rest of human cognitive architecture. These capacities are embedded in tacit knowledge and in the irreducible variability of real organizational context, neither of which fully exists in a training corpus.
The judgment gap is the space between what an agent can technically do and what the enterprise can safely authorize it to do. Most organizations assume that gap is a model problem. It is not. It is a cognitive architecture problem. The model output may constitute a technically valid decision. The enterprise has not yet built the infrastructure to specify what outputs the model is authorized to produce, under what conditions, with what evidence required, and when it must stop and escalate.
This is not fixed by a better prompt. It is not fixed by better retrieval. It is fixed by compensatory cognitive architecture: a systematic approach to supplying the layers of context that agents require as input but organizations have not provided in a governed and structured way.
Managing Context as a System
Context is not a single thing; it is four layers that converge in a critical way:
Data answers "what happened?" Facts, events, transactions, and current state. The foundational layer that every agent requires at every level of agency. SAP S/4HANA, SAP Business Data Cloud and SAP Dremio provide this as an open, unified foundation.
Semantics answers "what does this mean?" Meaning, identity, and relationships between entities. Representing that an invoice, a purchase order, a supplier, and a goods receipt are related, and in what way, is what allows an agent to generate outputs that span objects rather than remain within them. SAP Knowledge Graph and SAP Reltio provide semantic relationships and entity truth across the enterprise data and application landscape.
Process answers "how does work move?" Expected flow, observed behaviour, exception patterns, and the business rules embedded in application logic. An agent that can only access what happened, without a structured representation of what should have happened and why it did not, cannot provide outputs that diagnose, coordinate, or govern. SAP Signavio Process Intelligence provides this layer, grounding agents in both designed and observed process behaviour through event logs, process models, and business rules.
Prescriptive Intent answers "what is the enterprise actually trying to achieve, and what is allowed?" Objectives, authority, rules, evidence requirements, and stop conditions. This is the layer that makes execution governable centrally. Without it, an agent may produce outputs that are factually correct and logically coherent but organizationally wrong: wrong decision-maker, wrong threshold, wrong escalation path, wrong timing. The artifact that operationalizes this layer is the Process Atom: a discrete, machine-usable unit of enterprise behavioural intent that encodes what an agent is authorized to do, under what conditions, with what evidence required, and when it must stop and escalate. Process Atoms are not documents, prompts, or business rules in the traditional sense. They are governed, versioned artifacts that carry organizational authority at the resolution a machine can act on. SAP Company Memory is the repository that holds these atoms and manages their lifecycle: authored by humans with organizational authority, approved through a governance process, versioned so that changes are traceable, and called at runtime by any agent that requires them.
The critical insight is that these four layers must be managed as a system, not assembled ad hoc. An enterprise that has excellent data and no governed process context will fail when deploying diagnostic agents. An enterprise with excellent process intelligence and no prescriptive intent will fail at agentic decision-making. Managing context as a system means treating each layer as a governed asset with its own lifecycle, authoring process, and governance regime, and ensuring that the layers are integrated rather than siloed.
The matrix below scores each layer's dependency across all seven agent types on a 0 to 5 scale, where 0 means not required and 5 means the layer must be explicit, governed, and machine-usable with no gaps tolerated. The jump from the left side to the right is where the scaling problem lives.
Compensatory Cognitive Architecture
The term I use to bridge this chasm is compensatory cognitive architecture. It is the set of systems, artifacts, and governance structures that compensate for what machines do not inherently encode: organizational accountability, practical wisdom, and the institutional and situational memory of how decisions were made and why.
The key word is compensatory. We are not trying to give machines genuine judgment. We are building the architecture that constrains and guides machine outputs to remain within the boundaries of what the enterprise has explicitly authorized. The agent does not require a representation of why the three-way match policy exists. It needs the policy to be encoded in a form it can act on, with clear authority, clear conditions, and clear stop rules.
Process Atoms are that encoding: each carries enterprise intent at the resolution a machine can act on, with governance that the organization can manage and an auditor can verify. SAP Company Memory governs their lifecycle as described above.
Enterprise Architecture as the Governing Frame
None of this works without enterprise architecture as the surrounding governance frame. EA plays two distinct roles in the age of AI agents.
The first role is runtime governance. As agent deployments proliferate, the enterprise needs a registry of what agents exist, what systems they touch, what authorities they hold, and what their boundaries are. Without this, agent proliferation becomes invisible and ungovernable. SAP LeanIX AI Agent Hub provides this as a vendor-agnostic governance layer, discovering and registering agents across SAP, Microsoft, Google, Amazon, and other platforms. EA teams own this layer. It is not optional.
The second role is scaffolding context. The enterprise architecture map, covering systems of record, integration topology, application ownership, and business capability boundaries, is a source of cognitive scaffolding. An agent supplied with the context of which system owns a business object, which team owns a process, and which application enforces a business rule is better scaffolded than one operating without that structural context. LeanIX Application Portfolio Management is the source of this context, and it feeds directly into the cognitive stack that agents draw from.
Human Agency and Machine Execution in the Same Enterprise
The ultimate ambition of this framework is not to replace human judgment. Every company is a human accomplishment, after all. It is to create the conditions under which human agency, machine execution and automation can coexist within shared, auditable boundaries.
The agents on the other side of the scalability chasm are not replacing human workers. They are executing the coordination, transactional, and governance tasks that currently consume enormous human attention on low-to-medium judgment, high-volume activity. The humans who remain in the loop are there for the decisions that genuinely require accountability, practical wisdom, and contextual sensitivity that no architecture can fully encode.
Making that division of labor work requires the enterprise to reimagine work by being explicit for the first time about exactly where those boundaries are. What can an agent execute without approval? What requires a human sign-off? What evidence must exist before an action can be taken? What conditions must trigger an escalation? These are not technical questions. They are organizational governance questions that the enterprise must answer, encode, and maintain.
That is the work that the SAP Business AI Platform exists to support. Not to make the answers automatic, but to make the answers explicit, governed, and machine-usable.
What This Series Does
Each subsequent post in this series examines the classes of enterprise work, explores the specific context demands across all four layers, and shows where the scaling wall appears and what breaks it.
A blocked invoice scenario in the P2P process will be the context to show the concrete example throughout. It is a common, well-understood enterprise process that cleanly illustrates how context demands escalate from a simple retrieval agent all the way to an agent that governs compliance across the complete procure-to-pay cycle.
The goal is not to be abstract. The goal is to make the scaling problem concrete enough that an enterprise leader can look at their current AI program and know exactly which layer is missing and what it would take to fill it.
In Steinbeck's time, the scientists could crumple their tracings and start again. Enterprises deploying agents at scale cannot afford that luxury. Tracing the new world of the enterprise correctly before agents execute is what this series is for.
Post 2 covers the agents that are already scaling: Retrieve and Explain, and Create and Transform. And it introduces the question that most organizations have not yet asked: if these agents work because humans are still in the loop, what happens when you try to remove that loop?
Ready to move from framework to action?
Download our practical guide — 4-step guide to AI agent excellence — to see how leading enterprises are closing the judgment gap and scaling AI agents beyond pilots.