The Agentic AI Storage Shock
How enterprise agents turn data lakes, workflow logs, and generated artifacts into the next infrastructure gating layer
We spent the early part of the week at Dell Tech World, and a specific takeaway from conversations with Dell executives, customers, and practitioners was that the enterprise data layer is moving back to the center of the AI discussion. Agentic AI makes that inevitable once you grasp how the workflow is changing. Claude Code/Cowork and Codex are useful early indicators because they show what happens when AI moves beyond a cloud chat interface and gains read/write access to a working file system. Once the model can operate inside the project environment, the workflow leaves behind more than a final output. It creates a durable record of how the work was produced, what context informed it, and what changed before the result was accepted.
That is the same shift enterprises are preparing for across broader workflows. Agents will need more than model access or an application interface. They need enterprise context that can be retrieved, permissioned, verified, acted on, written back against, and retained. That requirement puts pressure on the entire data stack, from storage and retrieval to governance, observability, and the software layer that controls access to the data.
A useful way to understand the shift is the movement from cold data to warm data. Cold data is stored and retained and archived, or technically accessible, but not necessarily prepared for live use inside a workflow. It may sit in documents, old ticket histories, file shares, application databases, compliance archives, email threads, or data lakes. A human user can often work around that fragmentation by searching, interpreting, asking someone, or applying judgment. An agent cannot reliably do that unless the context is represented in a way the system can retrieve, permission, verify, and act on.
Agentic AI pushes more of that data into a warm operating layer. Warm data is not necessarily hot transactional data, but it is close enough to the work to be useful. It is indexed, governed, permission-aware, observable, and current enough to support an action. A support agent needs prior tickets, product documentation, entitlement rules, customer history, and an audit path. A procurement agent needs supplier terms, shipment data, approval thresholds, exception history, and writeback into the system of record. A coding agent needs the repository, dependencies, test logs, build artifacts, and security scans. In each case, the agent is not simply searching information. It is using enterprise context to complete work.
That observation changes how we should think about storage and the software layer above it. In an agentic enterprise, data has to be stored economically, protected, and recovered, while also becoming available for retrieval, reuse, governance, and audit. More workflow history becomes operating context. More access paths require permissioning. More agent actions create records that need to be logged for compliance, quality control, and future use. A growing share of the storage layer therefore participates directly in execution.
This is why we think the next phase of enterprise AI is best understood as a data architecture transition. The agent interface will get attention because it is what users see. The durable operating change sits underneath: trusted retrieval, identity, permissions, approvals, tool calls, writeback, observability, and retention. The outcome here is: agents need enterprise context to act, but every action creates new enterprise context. If that record is retained, governed, and made retrievable, the output of one workflow becomes an input into the next.
The uncomfortable version of this thesis is that if agentic AI works, the enterprise storage stack is underbuilt. Agents consume context, generate workflow records, create artifacts, and turn prior work into future operating memory. That means the data layer has to support both sides of the agentic loop: fast access to warm context during execution and durable retention of the new data created by the workflow. We explored the hardware side of this in our earlier report, Storage Wars, where we argued that flash is moving from a persistence layer into an active extension of the inference memory hierarchy. The agentic enterprise extends that logic from the AI rack into the enterprise data estate.
That loop gives stakeholders a more useful way to track the theme. Agentic AI spend is unlikely to appear as one clean budget line. It will show up through services work, workflow modules, governance products, observability and security tools, data platform modernization, and infrastructure refresh. The best positioned companies are likely to be those closest to the work, closest to the data, and closest to the retained record of what the agent actually did.
For paid subscribers, the full report includes:
Our full framework for the Agentic AI Data Flywheel and why enterprise data shifts from passive information to active operating memory.
A breakdown of the enterprise agent execution stack, including workflow ownership, systems of record, identity, permissions, retrieval, observability, writeback, and retention.
A deployment evidence map showing where agentic AI is moving first across ITSM, support, procurement, payroll, coding, underwriting, legal, and regulated workflows.
A framework for the three spend pools: software control plane, data readiness, and memory/storage infrastructure.
A detailed view of the software control points and which categories of vendors are best positioned.
Beneficiary map of what vendors across the stack are poised to benefit from this dynamic
A storage and memory stack analysis covering HBM, DRAM, SOCAMM, enterprise SSDs, QLC NAND, object storage, nearline HDD, archive, controllers, and networking.
A directional sensitivity model for how much data different agent workflows may create, from text research and decks to images, video, code, legal workflows, and engineering simulations.
Our bull and bear case for the thesis, including the key indicators we will track to see whether agents are moving from assisted productivity into controlled execution.