Beyond "agent washing": how to build AI systems that actually deliver ROI

TL;DR

• AI agents are not the same as chatbots
• AI agents can act autonomously to achieve an objective without human intervention
• Automating meaningless work won't lead to ROI
• Enterprise AI is moving beyond RAG to a "knowledge layer"

In the world of enterprise technology, buzzwords move fast. Right now, we are squarely in the era of "agent-washing," where almost every basic chatbot is rebranded as an autonomous AI agent.

But what does it actually take to move beyond the hype and deploy AI that transforms business operations? In a recent episode of the SuperDataScience podcast, host Jon Krohn sat down with John Roese, Global CTO and Chief AI Officer at Dell Technologies, to discuss how Dell achieved massive ROI on its AI investments, the emerging architecture of "knowledge layers," and where enterprise AI is heading.

Here are the seven key takeaways from their deep-dive conversation:

1. Defining "true" agentic AI vs. "agent washing"

The market is currently flooded with basic chatbots rebranded as "agents," a phenomenon Roese terms "agent-washing." True agentic AI represents a complete paradigm shift from basic text generation.

• Autonomous execution: true agentic AI is defined as a system that can take a high-level objective from a human, independently reason through the steps required, navigate digital ecosystems, and execute the work without a human in the loop.
• The digitization of skills: while the first wave of generative AI was about unlocking and talking to data, the agentic wave is about the "digitization of skills"—shifting AI from a passive assistant to an active digital worker.
• Objective-driven action: Instead of just responding to a linear prompt, a true agent continuously evaluates its progress toward a designated goal, adjusting its strategy as it encounters obstacles.

2. Why 95% of AI projects fail (and how Dell got a 30x ROI)

Many organizations struggle to see tangible financial returns from their AI implementations. Dell avoided this trap by being incredibly deliberate about its deployment strategy.

• Stop automating "meaningless" work: Roese notes that if you use AI to automate a task that was already low-value or meaningless when humans did it, you won't impact revenue or cost structures. Generative AI acts as a great detector for unnecessary business processes.
• Scale a few, don't scatter many: instead of letting thousands of fragmented AI proofs-of-concept bloom, Dell strictly limited its deployment to under 30 highly targeted use cases across sales, supply chain, services, and engineering.
• Massive financial returns: by scaling these select use cases across the entire global enterprise rather than keeping them localized, Dell achieved massive efficiency gains, with ROIs ranging from 10-to-1 up to 30-to-1.

3. The shift to "knowledge layers"

As AI models become a commodity, the real competitive advantage for an enterprise shifts from the underlying model to how it accesses and structures data.

• Moving past RAG: basic retrieval-augmented generation (RAG) is no longer enough for complex enterprise tasks. Businesses must transition to building a dedicated "knowledge layer."
• Enterprise knowledge graphs: a knowledge layer organizes corporate data into complex, semantic knowledge graphs that allow AI systems to understand relationships, context, and organizational logic rather than just scanning text.
• The new corporate backbone: this layer acts as a permanent corporate memory, allowing disparate AI agents to query a single, trusted source of truth across the entire enterprise.

4. Bridging the gap with model context protocol (MCP)

To perform actual work, autonomous agents need a standardized way to interact with tools, databases, and the web. Anthropic's open-source Model Context Protocol (MCP) is emerging as a critical framework, though it poses challenges.

• A standard for tool use: MCP acts as an open standard protocol that lets AI models seamlessly connect to data sources and secure environments without custom, fragmented API integrations.
• Perceive and act: by leveraging MCP, an agent moves beyond being a passive language model and gains the structural pipes necessary to perceive environment states and take action on external software.
• Enterprise security gaps: while powerful, Roese warns that MCP is not natively enterprise-grade out of the box. Dell manages this by centralizing MCP servers in a highly controlled environment to wrap the protocol in necessary security architecture.

5. The rise of "AI factories"

AI is changing how IT data centers are fundamentally designed, shifting infrastructure from standard computing architectures into specialized environments.

• Purpose-built architecture: Traditional data centers are designed for general-purpose application hosting. An "AI Factory" is a highly specialized data center topology built entirely to handle massive data throughput and continuous vector processing.
• Accelerating customer deployment: Dell’s clients are increasingly utilizing these AI factories to rapidly ingest enterprise data, fine-tune open-weight models, and host their proprietary knowledge layers.
• Predictable AI workloads: Treating AI infrastructure as a factory allows enterprises to transition AI from a highly variable, experimental R&D expense into a predictable, stable operational utility.

6. The mandate for sovereign AI by 2027

AI is no longer just a corporate priority; it’s a matter of national interest and national infrastructure.

• National AI strategies: Roese predicts that by 2026 or 2027, virtually every country will possess and mandate a dedicated "Sovereign AI" strategy to retain control over their citizens' data and cultural contexts.
• Protecting critical backend systems: sovereign infrastructure will extend beyond government administrative use. It will be legally required for hosting long-term autonomous agents operating in critical utilities, healthcare, and finance.
• Physical data localization: to meet these emerging legal requirements, enterprise knowledge graphs and backend systems will need to physically and legally reside inside environments structurally tied to that country's sovereignty.

7. What 2026 and beyond looks like for tech leaders

For CIOs, CTOs, and data scientists, the operational playbook is changing rapidly as AI moves out of the sandbox.

• Harmonized regulatory frameworks: leaders must proactively work alongside internal compliance teams and government guidelines to ensure their autonomous agents are certified and auditable.
• Transition to agent-to-agent coordination: the future workflow won't just be humans prompting an AI. It will involve specialized, cross-functional agent-to-agent protocols where AI workers coordinate with one another to solve multi-step problems.
• Rigorous governance against "slop": as AI-generated data begins to flood enterprise data streams, establishing absolute data governance and tracking provenance will be crucial to preventing system degradation.

To hear the full conversation on the democratization of skills, AI security, and Dell's operational roadmap, check out SuperDataScience Episode 953.