Enterprise AI brings real opportunity, but the vital question is what happens when it breaks. Investment is evolving quickly, and so is the attack surface. Agents are being deployed in live environments, but often without a clear plan for recovery if they are compromised, corrupted, or make the wrong call at the wrong time.

To address this reality, Cohesity recently released its enterprise AI resilience strategy. The company’s position is that resilience must come before AI scale, not after it. Organizations that try to skip past establishing resilience can end up spending more time recovering from disruptions than benefiting from AI. Sanjay Poonen, CEO and president of Cohesity, said, “Enterprises need the confidence to manage AI-driven risk and recover quickly when disruptions occur. Cohesity provides the resilience foundation that protects AI infrastructure, governs data access, mitigates agent-driven risk, and unlocks the transformative power of trusted enterprise data.” While this sounds ideal — and Cohesity does have a long track record of delivering for enterprises — this framing also puts a lot of weight on execution.

The Prevention-First Security Model Needs to Evolve

What I continue to see across enterprises are security budgets that are still heavily weighted toward keeping attackers out. Cohesity estimates organizations spend four to five times more on prevention than on recovery. And that ratio made sense when the threat landscape was slower and more predictable.

Unfortunately, the cost of launching an attack is falling, while the number of vulnerabilities has increased. Attackers are using their own agents to automate reconnaissance and accelerate breach timelines. Organizations that assume their defenses are good enough may be working from an outdated model.

AI Agents Extend the Attack Surface into Unfamiliar Territory

One of the challenges organizations face is that the tools they have built their recovery programs around were not designed for the workloads they are now running. Agents are emerging as a new application model. They continuously access data, execute decisions, and interact with other systems.

Agents can be compromised through prompt injection, corrupted inputs, or logic errors. Any of those can cascade across systems in seconds. Restoring data is not the same as restoring an agent to a working state. The agent’s permissions, tool configurations, memory, role definitions, and associated databases all need to come back together correctly. Cohesity is working to extend its platform to cover the full stack, including agent memory, vector databases, model configurations and policies, training and fine-tuning datasets, and enterprise data stores.

According to its strategy, the company plans to preserve immutable snapshots of AI environments and enable synchronized, point-in-time recovery across those components, which depends on accurate dependency mapping and orchestration across heterogeneous systems — all with the goal of reducing downtime without requiring full system rebuilds. Whether the recovery tooling holds up in complex, multi-agent production environments remains an open question, and Cohesity will need to demonstrate its answer in production.

Cohesity’s Ecosystem Approach Addresses a Gap

Cohesity isn’t trying to deliver a fully vertically integrated stack, which reflects the reality of how enterprise AI environments are evolving — and I believe that’s a smart move. Instead, the strategy relies on an ecosystem model where different vendors contribute distinct capabilities across orchestration, observability, security, and recovery. ServiceNow provides workflow and orchestration to manage agent-driven processes, while Datadog delivers observability into system and application behavior. Cohesity’s role centers on data protection, governance, and recovery, aiming to ensure that underlying data and AI-related state can be restored to a trusted point in time.

This type of integration is increasingly enabled by emerging standards such as the Model Context Protocol (MCP), which aims to standardize how AI agents connect to enterprise tools, access data, and execute actions across systems. While still evolving, frameworks like MCP could help link orchestration, observability, and data protection into a more cohesive operational model, reducing friction in multi-vendor environments.

Cohesity is leaning into a more automated response model for AI-driven threats. Anomalous behavior or policy violations can trigger API-driven, point-in-time restoration, potentially reducing the time to contain incidents. No single vendor covers the full threat surface of an AI-enabled enterprise, and connecting best-of-breed tools is more realistic than replacing them.

The harder question is what happens when the ecosystem itself becomes the point of failure. Multi-vendor environments introduce coordination risk. How conflicts between systems are identified, escalated, and resolved will matter just as much as any individual capability. That said, there is no certainty that single-vendor environments will be better, given that no single vendor can detect all threats.

Data Visibility Remains a Prerequisite

Resilience starts with knowing what data you have and where it lives. Cohesity’s DSPM, powered by Cyera, is built to continuously discover, classify, and monitor sensitive data across cloud, SaaS, and AI environments. Cohesity reports high classification accuracy at scale, though results will vary depending on data types, environment complexity, and implementation maturity, particularly in large, distributed environments where data sprawl and inconsistent tagging are ongoing issues. Classification accuracy directly affects how well organizations can prioritize risk and decide what needs protection. Tying that visibility into Cohesity’s broader cyber resilience framework is the right approach, since it connects insight to action.

The challenge is adoption at scale. Organizations need to be able to trust the outputs enough to act on them, and that doesn’t happen automatically. Classification is only part of the equation. Success still also depends on governance, ownership, and consistent processes, and those tend to take longer to put in place than the technology itself.

Sovereign Cloud Requirements Are Adding Complexity

Regulatory and data sovereignty requirements are adding another layer of complexity to enterprise resilience, especially across Europe, the Asia-Pacific region, and Canada. Compliance is no longer something handled after deployment; it has to be designed into the system from the start. Cohesity’s sovereign-by-design approach is intended to absorb this complexity by effectively addressing requirements such as data residency, local control, and regulatory alignment. By allowing implementation details to vary by region and partner, customers should be able to operate across regions without constantly reworking their environments as regulations evolve.

Partnerships with providers such as AntemetA, Singtel, and Micrologic extend that model into markets where local presence and trust make a difference. At the same time, Cohesity is trying to serve both large enterprises and midsize organizations. That expands the opportunity, but it also introduces tension. Simplicity at the mid-market level doesn’t always align with the flexibility and control that enterprise customers expect, and balancing both without fragmenting the platform will again require disciplined execution.

Alongside this, Cohesity is pushing a broader shift from treating data as something to protect to something that can be used. Capabilities such as the Cohesity Gaia AI application and federated semantic search enable organizations to access and analyze backup data without copying it or breaking governance controls. As Zubin Irani, VP of partnerships at Glean, put it, “AI agents are only as useful as the information they can securely access. Enterprises have decades of critical knowledge stored across systems, but much of it remains locked away.” Cohesity wants to help change that.

What Cohesity Still Needs to Prove Before the Strategy Becomes a Foundation

It’s in the nature of multi-agent systems that they will fail in ways that are difficult to predict, and this means that Cohesity’s ecosystem approach introduces real coordination risk across vendors. On top of that, classification at scale only matters if organizations act on it, and shifting backup data from a recovery layer to an active AI input is as much a behavioral shift as it is a technical one.

What Cohesity needs now is evidence. I’m looking forward to seeing production case studies that show measurable impacts, clearer detail on how failures are handled in real environments, and concrete examples of backup data being used in AI workflows. Real-world results will determine whether the company’s strategy becomes a true foundation for operations or remains simply a well-structured framework.