Share with your CIO
More than 80 percent of enterprise AI projects never reach production, according to RAND Corporation, and Arango’s COO Ravi Marwaha argues the culprit isn’t model quality but the fractured data architecture agents inherit the moment they leave a controlled pilot. Speaking on the AI in Business Podcast alongside IBM’s US industry CTO Sumedh Chaudhary, Marwaha describes a compounding pattern where decades of “consolidation” actually multiplied data copies, leaving agents with no coherent operational picture to reason from. The fix isn’t another data lake. It’s connected, decision-ready context.
What this means for your business
The organizations most exposed here are those that have already run successful pilots and are now planning production rollouts. If your current AI infrastructure relies on copied datasets, layered BI tools, or siloed document repositories, the agent that performed well in a curated test environment will degrade in production, not because the model changed, but because the environment did. The gap between “it worked in the demo” and “it works at scale” is almost always an architecture gap, not an intelligence gap.
The Marwaha and Chaudhary framing, delivered through content sponsored by Arango, a vendor selling exactly this contextual infrastructure, tilts toward making the problem sound uniquely unsolvable by existing data platforms. That’s worth noting, but it doesn’t make the diagnosis wrong. The “pilot purgatory” pattern the Carnegie Endowment flagged in January 2026 is real and well-documented outside vendor circles. The architectural argument, that agents fail when they can’t trace what changed across connected systems, is consistent with how production failures actually present. Where the framing oversimplifies is in implying that a unified context layer is a clean, bounded purchase rather than a multi-year integration challenge on top of the data problems you already have.
Regulated, document-heavy workflows, think loan processing, insurance underwriting, clinical documentation, are the sharpest diagnostic for whether your enterprise is actually agent-ready. Chaudhary’s point that error rates in these workflows expose architectural failures before model limitations is a useful inversion of how most AI programs are evaluated. Most organizations benchmark model accuracy in clean conditions. The better test is whether semantic continuity, the ability to preserve meaning across page breaks, system handoffs, and document relationships, holds under real operational conditions. If it doesn’t, adding more agents makes the problem worse, not better.
The multi-agent orchestration framing is where this argument earns its sharpest edge. A single agent operating over fragmented data is a liability. Multiple agents operating over the same fragmented data is a coordination disaster. The enterprises that will move fastest aren’t the ones deploying the most agents but the ones that build a single operational picture first and let agents share it. I’d revise this assessment if a major production deployment demonstrated that federated, agent-specific data pipelines could sustain accuracy thresholds in regulated workflows, but nothing in the current evidence points that direction.
Concept deep-dive: Temporal awareness
Temporal awareness is an agent’s ability to track not just what the current state of a system is, but what changed, when it changed, and why that change matters to the decision at hand. Think of it as the difference between reading a snapshot and reading a ledger. Without it, an agent approving a contract or flagging a compliance risk is working from a photograph of a river, not the river itself. In high-stakes workflows, that distinction is the difference between a defensible decision and a liability.
Based on reporting from Unified Context as the Missing Foundation for Enterprise AI, originally published 2026-07-07 11:47:00.
