Dell’s Agentic AI Strategy Moves From Chatbots to Autonomous Work

Dell’s message at The Six Five Summit: AI Unleashed 2026 was that enterprise AI is entering a different phase. Instead of using chatbots mainly to search, summarize or create content, companies are beginning to deploy agents that carry out business work. Dell said it had reduced roughly 900 possible AI projects to 13 initiatives with clear business value, and described early productivity gains of about 20%–40%. 1

The claim is Dell’s, not an industry-wide measurement. But it offers a useful framework for understanding what changes when AI moves from experimentation to operational execution.

Agentic AI is about doing work, not just answering questions

Dell drew a line between first-wave generative AI and agentic systems. Chatbots and assistants can help employees find information in proprietary data, summarize documents or produce content. An agentic system is designed to digitize a process: it can reason through a task, use tools, coordinate steps and complete work with less direct human intervention. 1

That difference affects more than the model. Dell said production agentic AI requires changes to the technology stack, infrastructure, governance and organizational design. The strategic goal is to reduce the link between how much work a company can complete and how much human capacity it has available. 1

Dell’s production examples: CRM, code and pricing

Dell said it had worked on agents for about two years and moved them into production during the preceding year. The company identified three examples of operational use:

• CRM-data cleanup: agents can handle repetitive data-quality work across customer records.
• Software development: agents can assist with routine coding, annotations and parts of CI/CD coordination.
• Special pricing: agents can support a workflow where speed and consistent application of rules matter. 1

These examples are narrower than the idea of a fully autonomous enterprise, but they illustrate why process design matters. A useful deployment is not merely a model placed inside an existing application; it is an automated workflow with defined inputs, tools, permissions and outcomes.

Why Dell cut 900 potential projects to 13

Dell said it narrowed roughly 900 candidate AI projects to about 13 focused initiatives selected for explicit business value and managed through top-down prioritization. 1

The decision reflects a practical lesson for enterprise AI programs: a large inventory of experiments does not necessarily create an operating advantage. Dell’s argument was that organizations should concentrate resources on workflows where the expected business outcome can be measured, then build the controls and infrastructure needed to run those workflows reliably.

Dell attributed initial productivity improvements of roughly 20%–40% to targeted assistant deployments and said gains could be substantially larger when agents take over complete categories of work. It also claimed the focused program helped decouple revenue growth from the company’s cost structure. Those are company-reported results and should be treated as claims about Dell’s experience, not a guaranteed benchmark for every deployment. 1

“Agent washing” obscures the real difference

Dell executive John Roese warned against treating every AI feature as an autonomous agent. In Dell’s framing, agent washing happens when chatbots, limited assistants and systems capable of autonomous execution are described as if they were equivalent. 1

A useful distinction is:

• Chatbot: responds to a prompt, usually producing information or content.
• Targeted assistant: improves a bounded task while a person remains closely involved.
• Autonomous agent: can plan, call tools, coordinate actions and execute a workflow with substantially less direct supervision. 1

The distinction matters because autonomy raises the stakes. A mistaken summary may need correction; an agent with access to business systems may alter records, trigger downstream actions or expose sensitive information.

Production agents change the infrastructure equation

Marvell CEO Matt Murphy said production agentic inference creates different infrastructure pressures from traditional AI training workloads. Agents may need to preserve workflow state through key-value caches across repeated model calls, large contexts and long, multi-agent interactions. That can make memory capacity and bandwidth a constraint before raw compute becomes the primary bottleneck. 1

Murphy also said CPUs become more important because agentic systems require orchestration: branching, retrieval, tool calls, sandboxed execution and coordination between agents. In the discussion, the CPU-to-GPU ratio was described as shifting from roughly 8–16:1 in traditional AI toward about 1–2:1 for agentic workloads. 1

The implication for buyers is straightforward: an agent platform cannot be evaluated only by model quality or GPU capacity. Memory, networking, CPU orchestration, observability and the cost of repeated inference all affect whether a workflow is viable in production.

One architecture will not fit every agent

Dell argued against a single-model or single-location “monoculture.” Its proposed approach combines options such as on-premises open models, Dell-hosted frontier models, controlled virtual private clouds, external APIs and models running on devices. The right choice depends on performance, compliance, functionality and cost. 1

Economics also depend on the job being automated. Dell’s discussion contrasted a high-value executive decision-support agent, which could justify substantial monthly spending, with a high-volume CRM-cleanup workflow that needs very low per-record costs—described in the discussion as below roughly $0.50. 1

That suggests a better evaluation question than “Which model is best?” Enterprises should ask which model, deployment location and level of autonomy produce the required outcome at an acceptable cost and risk.

Identity is the control layer for autonomous agents

Dell said every autonomous agent that accesses its data, whether internal or external, should receive a Dell-issued digital identity. That identity would support fine-grained authorization and allow access to be revoked as a practical kill switch. 1

The security challenge becomes more difficult when agents do not simply inherit a user’s identity. A user-inherited agent acts under an identifiable employee’s credentials. A headless agent operates more independently, making attribution, permissioning and lifecycle management harder. Dell expected headless agents to account for about 70% of future deployments, according to the discussion. 2

Dell also pointed to agent-to-agent and agent-to-tool communication protocols, as well as post-quantum readiness, as parts of the security model for production autonomy. 1 The broader lesson is that an agent should be treated as a software identity with a defined owner, scope, audit trail and revocation path—not as an invisible extension of a chatbot.

Jobs will be decomposed into tasks

Dell said it analyzed 6,800 jobs by breaking them into categories including productivity, hygiene, coordination, expert and human-facing work. Its conclusion was that agents are more likely to remove portions of roles than automatically eliminate entire positions. 1

In software development, for example, agents may handle routine coding, annotations and parts of delivery coordination, while people spend more time on architecture, requirements, judgment and customer interaction. 1

That does not mean roles remain unchanged. Dell’s argument is that every job will be reshaped as automatable tasks are extracted and workers move toward responsibilities that depend more heavily on expertise, context, judgment and human relationships.

The production test for agentic AI

Dell’s position can be reduced to a demanding test: an agentic AI project is ready for production only when it has a measurable business outcome, a clearly bounded workflow, suitable economics and controls for identity, access and failure recovery.

The move from 900 possibilities to 13 initiatives captures the strategy. Agentic AI is not simply a more conversational interface. It is an operating model in which software performs work—and where infrastructure, security and job design must evolve alongside the models.