Ninety-five percent of AI pilots failed to deliver measurable ROI according to a 2025 paper from the MIT Media Lab. Companies exploring AI are running proofs of concept, hosting demos, and briefing their boards. But when the question turns to when this can reach production scale, the answer is usually: Not yet.

Why does the path to realizing AI value at scale prove harder than expected?

The common thread is not the lack of AI capabilities; it's the missing context that the AI needs to work. What gets left behind in AI pilots is the full process logic, business rules, security constraints, and relationships that make the data meaningful and that allow users to securely interact with it.  In fact, data quality or availability issues was cited by 73% of respondents as the primary barrier to AI ROI in a 2026 survey of 2,600 business leaders by SAP and Oxford Economics.

Closing the context gap is not a matter of extracting more data and deriving more ontologies in an external data environment. It requires a fundamentally different approach: Bringing the AI to the Enterprise Resource Planning (ERP) application suite so it can work in a context-native environment.

Why business context Is critical for enterprise AI

AI models are trained on publicly available internet content and while remarkably capable, are not trained on your customer contracts, purchase orders, approval hierarchies, or other operational logic that governs your business. That knowledge is embedded inside your business applications and needs to be understood by the LLMs.

When we talk about business applications, it is with ERP at the center. Why? Because ERP is the brain of every company. Every event that matters to a business ultimately resolves into a transaction executed in and governed by the ERP suite. A customer places an order and revenue accrues. A supplier part hits the dock and inventory is recorded. An employee joins the company and payroll is activated.  These are not simply representations of events. They are the events themselves as the business has formally recognized and committed to them.

An AI agent that can act on a business situation, like proposing a supplier substitution or recommending a production plan adjustment, needs to do more than read data. It needs to understand the rules governing the data, the process context it sits within, and the system it will need to act on to complete the loop from insight to outcome.

This requires more than sending data to Large Language Models (LLM). Your data was created from thousands of business rules that embed nearly infinite interdependencies: A customer is linked to a contract for a product that is tied to a bill of material containing a purchased material delivered by a supplier that carries a risk score. A macro-economic event at the supplier’s location ripples back through your business and onto your customers. AI needs to understand those relationships to warn you of impacts and recommend actions.

AI that works directly with your ERP suite does not need to be taught how your order management and procure-to-pay process works. It already knows.

Why context-native AI outperforms simulated AI environments

If we agree that AI needs broad and deep context to work effectively, how do we provide that to the AI?  Should we bring the AI to the context, or should we build a simulated environment and let the AI work with that?

Imagine the AI as a fish.

In the first scenario, the fish swims in the ocean. The ocean is the live business environment: your data, your transactions, your process logic, your security and compliance rules. The fish doesn't need anyone to explain the ocean to it. It lives in it. Every decision it makes is informed by the full, living context around it, in real time.  This is context-native AI.

In the second scenario, the fish swims in an aquarium. To be fair, it is a sophisticated aquarium. There are plants in there. A filter running. Gravel on the bottom. Someone has invested real effort in making it look like the ocean.

That filter running on a schedule? That is your data pipeline.

Those artificial plants? That is your custom-built ontology painstakingly recreated by your experts.

And the glass walls your fish keeps bumping into? Those are your boundaries where everything outside those walls is invisible to the AI.

This is an external AI platform approach.

Some AI vendors will tell you the answer is to build an aquarium. Aquariums make sense when you want to isolate and study a few fish. They fail when you need an entire ecosystem operating at business speed.  It works for a pilot, but when you need to deploy across your business, you’ll need a bigger aquarium, or more aquariums. That means mirroring more data, building more ontologies, and creating more simulations of your real environment. This comes with effort, duplication, and cost.

And no matter how big you make the aquarium, or how many aquariums you have, it is still not the ocean.

The moment a real business event occurs—a supplier fails to deliver a critical part, a key customer announces bankruptcy, or a key executive leaves the company—the ocean changes instantly. The aquarium does not.

How context-native AI turns business context into action

With context-native AI, AI agents swim inside the full context of the ERP application suite, not in a simulated environment outside of it. This is an architectural distinction that changes how AI works at operational scale.

The context-native approach means AI has direct access to business rules, relationships, and logic that make business data meaningful. Deterministic business rules continuously check probabilistic AI outputs, catching the errors before they propagate. Security and compliance rules are enforced natively, extended from applications and users to agents, not rebuilt in a parallel framework. When an AI agent reaches a decision point, it asks a human for guidance and then executes directly with the system of record, completing the loop from insight to action.

Key to making this work is a native knowledge graph that represents the rich relationships across the data based on your business rules, giving AI agents process context across the full application landscape. There is no need to extend or rebuild the graph as you add new organizations, modules, and business models. Context scales with the business.

Extensibility is also important. A context-native architecture does not require every agent to be pre-built. Instead, customer-specific apps, agents, and extensions to standard agents can be built with the same business context and governance model as the core application suite. That means organizations can innovate without disconnecting AI from the operational context that makes its recommendations trustworthy and actionable, and without creating additional technical debt.

Context-native AI is AI that navigates, thinks, and executes from the inside. An agent that can read a purchase order is useful. An agent that can translate a late shipment from a purchase order through to a bill of material, customer order, and product alternatives to propose to the customer, all within the same authorized context, is transformative.

Why business context is so hard to recreate outside ERP

External AI platforms can deliver value and are appropriate for use cases like pilots or to solve isolated challenges, but they become increasingly difficult to govern and maintain at enterprise scale across operational business processes. Understanding why reveals the depth of the architectural differences.

External AI platform approaches can access business data through open APIs, but APIs expose raw data, not the knowledge graph that encodes how that data relates to business rules, process flows, and organizational structures.

When an AI outside the application suite calls an API to retrieve a supplier record, it receives just the data, not the full graph of relationships that makes that data meaningful, like the contracts, risk classifications, approval thresholds, and compliance rules that govern any transaction involving that supplier.

To compensate for the missing context, you must do one of the most expensive things in enterprise technology: Reverse-engineer the semantic layer with help from internal subject matter experts and build a custom ontology that mirrors what the application suite already encodes natively.  Then you must maintain that ontology as the application context changes which is a permanent operating cost.

Security and compliance must also be considered. When AI operates inside the platform, it inherits these controls automatically. When AI operates outside the platform, those controls must be rebuilt independently, creating a parallel governance framework that must be maintained in sync with the source system of record. Every time a role changes in the business application, someone must manually reconcile that change in the AI layer. Every audit now covers two perimeters instead of one.

McKinsey on why ERP is the Foundation for AI at Enterprise Scale

In January 2026, McKinsey & Company published an article specifically addressing the relationship between ERP and enterprise AI scale, and their conclusions align closely with the architectural argument above.

McKinsey's report, "Bridging the great AI agent and ERP divide to unlock value at scale," makes three arguments that validate a Context-Native approach:

1. That ERP defines how value flows through the business
2. That an ERP foundation unlocks value at scale
3. That embedding AI into the workflows where the work gets done is crucial for adoption

The implication is clear:  Organizations should treat ERP as a key enabler and “Companies that close the divide between AI ambition and ERP readiness will move fastest from experimentation to real, defensible P&L impact.”

Building an AI strategy that scales beyond

If your organization is currently evaluating AI data platforms, or has already made investments in building an AI pilot on top of your business applications, the most important question to ask is this: What problem are you actually solving, and is building simulated environments the answer?

An external AI platform can deliver value in a pilot or answer a specific question within a specific business domain. The concern is not the pilot, it is the path to scale. Beyond piloting, every step towards a bigger external AI architecture increases the cost of context maintenance, multiplies the governance and risk surface, and moves you further from the live business context that makes AI genuinely useful.

Our recommendation: Prioritize modernization of the ERP suite in parallel with your AI exploration. A modernized ERP application suite means creating the conditions that AI needs to operate at enterprise scale: Standardized processes, clean master data, a built-in semantic layer, and embedded governance. With a modernized ERP suite, the ocean is already there. The decision is simply whether your AI will swim in it.

Bottom Line: Context is the foundation of enterprise AI

Enterprise AI is not a technology problem. It is a context problem. The organizations that will extract value at scale from AI are not the ones that build the most sophisticated data pipelines and simulated environments. They are the ones that give AI the most complete, accurate, and current view of their business, and the ability to act on it directly in a secure and governed manner.

SAP's approach brings AI into the application suite across finance, supply chain, procurement, HR, and customer experience. Bringing the AI to the context is how SAP’s approach to Business AI is different.

To do this, SAP's Knowledge Graph uniquely maps application rules and data relationships across 452,000 tables and 7.3 million data fields.  This semantic layer is what transforms data into context, and context is what separates AI that can reason about your business from AI that can only guess at it.

And bringing the AI to the application suite allows SAP to extend the standard governance framework to the AI so you don’t need to duplicate it externally.

Context-native AI swims in the ocean. External AI lives in an aquarium. No matter how big you build the aquarium, or how much you invest in making it look like the real thing, it will never be the ocean.

Hit the ground running and scale seamlessly with AI cloud ERP from SAP, built for companies of any size.