Unified Context as the Missing Foundation for Enterprise AI

This article is sponsored by Arango and was written, edited, and published in alignment with our Emerj sponsored content guidelines. Learn more about our thought leadership and content creation services on our Emerj Media Services page.

​Scalable AI requires coherent, connected context, since no model can compensate for an enterprise architecture that presents multiple versions of truth.

RAND Corporation found that more than 80 percent of AI projects never reach production — twice the failure rate of conventional IT projects — and identified inadequate data infrastructure and leadership misalignment as leading causes.

Evidence of that challenge spans industries and geographies. The Carnegie Endowment for International Peace noted in a January 2026 practitioner paper that many AI initiatives become trapped in “pilot purgatory” because production environments require robust data flows, governance frameworks, and institutional readiness that pilots rarely test. Solutions that succeed in one environment often fail to transfer cleanly to another, forcing organizations to rebuild critical infrastructure repeatedly.

As agents move from assisting employees to acting on behalf of the enterprise, governance becomes as important as model capability. NIST’s AI Risk Management Framework stresses that trustworthy AI requires transparency, accountability, monitoring, and traceability throughout deployment. Without those foundations, organizations cannot consistently explain, audit, or trust AI-generated decisions.

The pattern shows that capability is often capped by architecture rather than intelligence.

In a recent series on the AI in Business Podcast, Ravi Marwaha, Chief Operating Officer and Chief Technology Product Officer at Arango, and Sumedh Chaudhary, CTO US Industry Market at IBM, dug into why fragmented data, missing context, and brittle workflows push enterprises into an AI failure zone — and what it takes to build agentic systems that act accurately and explainably inside real, high‑stakes operations.

This article examines four insights that clarify why enterprise AI repeatedly stalls at scale —  and what leaders must rebuild to make agentic AI accurate, explainable, and economically defensible.

• Unified context for accurate agent decisions: Agents need real‑time context — what changed, how systems relate, and why it matters — to stop guessing and produce decisions that match expert judgment.
• Fragmented Data the Root of the AI Failure Zone: Unified, decision‑ready context built from connected systems, rather than consolidated copies, eliminates the architectural drift that drives AI failure.
• Regulated workflows as the proving ground for trustworthy AI: High‑stakes processes force the temporal awareness, traceability, and evidence standards required to scale AI safely across the enterprise.
• Multi‑agent orchestration for end‑to‑end automation: Agents only work when they can operate over the same contextual layer, enabling coordinated actions that replace isolated pilots with real workflows.

Listen to the full episodes:

Episode 1:  The Architecture Shift Behind Reliable Enterprise AI – with Ravi Marwaha of Arango​

​Episode 2:  How Unified Context Turns AI Into Real Enterprise Performance – with Ravi Marwaha of Arango

​

Guest: Ravi Marwaha, Chief Operating Officer & Chief Technology Product Officer at Arango​

Expertise: Enterprise AI, Product Strategy, Data Infrastructure, Digital Transformation

Brief Recognition: Ravi Marwaha is a technology executive with more than 30 years of experience leading product, engineering, and digital transformation initiatives across enterprise software and financial services. He currently serves as Chief Product and Technology Officer at Arango, where he leads product strategy, engineering, customer success, and implementation, with a focus on AI data infrastructure and enterprise AI platforms. Prior to Arango, Ravi was Chief Product Officer for the Commercial Bank Digital Platform at J.P. Morgan. He has also held senior leadership roles at Uptake, GE Digital, SAP, and Informatica, overseeing global product, engineering, cloud, and customer success organizations. Ravi holds a Bachelor of Engineering in Mechanical Engineering from Karnatak University.

​Episode 3:  Why AI in Document-Heavy Workflows Fails Without the Right Foundation – with Sumedh Chaudhary of IBM

​

Guest: Sumedh Chaudhary, CTO US industry market at IBM

Expertise: Enterprise AI, Enterprise Architecture, AI Go-to-Market Strategy, Hybrid Cloud

Brief Recognition: Sumedh Chaudhary is a technology executive and enterprise architect with more than 20 years of experience spanning AI, cloud, enterprise architecture, and consulting. He currently serves as CTO for GSI & Industry Tech at IBM, where he leads technology strategy and AI platform adoption through Global Systems Integrator partnerships. Prior to this role, he served as CTO for IBM’s GSI Ecosystem Partnerships and as Associate Partner for Strategic Deal Solutions. Earlier in his career, Sumedh held consulting and leadership roles at Thomson Reuters and Deloitte, delivering enterprise technology, data, and analytics initiatives across the public and private sectors. His work has been recognized with multiple IBM honors, including the IBM Architect of the Year Award for Watson. Sumedh holds both an MBA in Business Analysis & Marketing and an M.S. in Data Science from Georgia State University.

Ravi Marwaha opens the series by arguing that production failures stem from the environment around the model rather than the model itself. Pilots operate inside curated conditions where data is pre‑selected, workflows are simplified, and inconsistencies are shielded from the agent. Production introduces the full complexity of enterprise systems built for human navigation — CRMs, ERPs, ticketing tools, data lakes, and document repositories — each holding a different slice of truth.

“Your data is never in one place, and even organizations with mature lakes, warehouses, and MDM systems still operate with structured, unstructured, and multi‑model data spread across environments built for humans rather than agents. When agents are fed irrelevant information or stitched‑together fragments, they don’t become more intelligent — they lose the grounding required to make reliable decisions. Models hallucinate because they don’t have context, not because they’re weak.”

— Ravi Marwaha, Chief Operating Officer & Chief Technology Product Officer at Arango

Sumedh Chaudhary sees the same dynamic in regulated workflows. In document‑heavy environments, missing context shows up immediately as measurable error because the agent cannot maintain continuity across pages or systems. Leaders often misdiagnose this as a data‑quality issue, but both guests emphasize that the underlying failure is architectural: agents cannot make defensible decisions when the operational information they rely on is incomplete or contradictory.

Executives looking to operationalize this shift can begin by clarifying what an agent must actually know at the moment of decision. In both conversations, three moves emerged consistently:

• Define the decision context — identify the specific signals an agent must rely on, not the entire historical record.
• Map where that information lives — understand which systems hold which fragments of truth.
• Establish temporal awareness — ensure the agent can track what changed and why it matters.

For leaders, the implication is clear: AI readiness depends on whether an agent can access the operational information required to understand what changed, where information lives, and how systems relate. Without that foundation, even strong models behave unpredictably once they leave the controlled conditions of a pilot.

Across both conversations, fragmentation emerges as a primary barrier to reliable agent behavior at scale. Marwaha describes enterprises as environments shaped by decades of accumulated architectural decisions — data copied for convenience, systems layered atop systems, and reporting tools built atop those layers. None of this was designed for autonomous decision‑making, and all of it becomes visible the moment an agent tries to act across systems.

“People said they were consolidating, but in practice they were fragmenting even more — copying data from one place to three places, then ten places, and eventually consolidating those copies into an eleventh location. Over time, BI tools created yet another layer of data, each version slightly different from the last. This pattern has been compounding for a very long time, and agents inherit all of those inconsistencies the moment they try to act across systems.”

— Ravi Marwaha, Chief Operating Officer & Chief Technology Product Officer at Arango

Sumedh Chaudhary sees fragmentation manifest differently in document‑heavy workflows. A model can interpret a page, but if page‑to‑page meaning is lost during extraction or processing, error rates rise quickly — especially in regulated industries where accuracy thresholds are explicit. The agent is not failing to understand the content; it is failing to understand the relationship between fragments.

For leaders, the takeaway is that fragmentation is not solved by centralizing everything. Centralization often creates new copies and new inconsistencies. The real challenge is ensuring decision consistency — tracing where meaning breaks between systems, versions, or documents, and treating those breakpoints as architectural issues rather than data‑quality problems.

Regulated, document‑heavy workflows reveal enterprise readiness faster than any other environment. According to Chaudhary, these workflows fail not because the models are inadequate, but because enterprises cannot preserve the continuity agents need to reason across pages, documents, and systems. Error rates spike the moment an agent loses the semantic thread that connects the workflow.

“A document‑heavy workflow presents a third level of challenge because you’re not dealing with just unstructured text — you’re dealing with images, tables, and page breaks that disrupt the semantic thread. When systems flip from one page to the next, they often lose the contextual layer that connects the information, and the agent cannot reconstruct what the workflow requires. In regulated industries, error rates are measured explicitly, and tools are abandoned quickly when those thresholds aren’t met.”

— Sumedh Chaudhary, CTO US Industry Market, IBM

These workflows force temporal awareness, traceability, and evidence into the architecture. They require agents to understand what changed, when it changed, and how that change affects the decision at hand. When continuity breaks, compliance thresholds fail long before model capability becomes the limiting factor.

Chaudhary highlights several early indicators leaders should monitor before attempting scale:

• Error‑rate improvement week by week — stagnation signals architectural issues, not model issues.
• Temporal reasoning — agents must demonstrate awareness of what changed and when.
• Semantic continuity — page‑to‑page and system‑to‑system meaning must remain intact.

For executives, the implication is that regulated workflows are not niche — they are the clearest diagnostic for determining whether an enterprise can support agentic AI. If an agent cannot maintain continuity in these environments, it will not behave reliably anywhere else.

The final insight across the conversations is that agentic AI becomes operational only when multiple agents can act over the same connected operational picture. Marwaha notes that enterprises often attempt to scale by adding agents, but without shared operational information, each agent becomes another silo. The challenge is not the number of agents; it is the coherence of the environment they operate within.

Sumedh describes how multi‑agent systems function in document‑heavy workflows. No single agent can maintain context across extraction, page transitions, document relationships, and cross‑system meaning. Instead, OCR agents, vector agents, splitter agents, and matching agents each carry part of the reasoning load.

“In a well‑designed system, you have an OCR agent to extract the digital footprint, a vector agent to store embeddings, a splitter agent to preserve page‑to‑page continuity, and a matching agent to link documents across the workflow. Each agent carries a different part of the reasoning load, and none of them can succeed alone because the workflow depends on how they interact. A good orchestration layer is like a concert master — it coordinates the musicians, but it cannot produce the performance without them.”

— Sumedh Chaudhary, CTO US Industry Market, IBM

Orchestration binds these agents into a coherent chain of reasoning. Supervisor agents can oversee decisions, and separation of duties can reduce error in high‑stakes workflows. But orchestration only works when all agents operate over the same operational information — not separate pipelines or inconsistent copies.

For executives, the takeaway is that multi‑agent systems are an emerging architectural approach for coordinating complex enterprise workflows. The difficulty is not deploying agents; it is ensuring they share the same operational picture of the business.