This is part 1 of a 3-part series about AI and its role in cyber resilience. Please watch for the next post, which will discuss what happens when the AI service itself is what needs recovering.
Picture the moment a business needs to recover from a cyberattack.
The security team is still working out what happened. Infrastructure teams want to get systems running. Business leaders need to know when operations can resume. Then someone asks the question that changes the conversation: Which protected data copies can we trust?
That’s where I want the discussion about AI and cyber resilience to start. Not with a chatbot. Not with a promise of autonomous recovery. With the decision people have to make when the business is waiting and the evidence is incomplete.
My view is straightforward: AI should help us reach that decision with better evidence, less manual work, and clear accountability. It shouldn’t try to make uncertainty disappear behind a confident answer.
Key Takeaways
- Effective cyber recovery should depend less on automated recommendations and more on evidence-based decisions that organizations can explain, validate, and defend.
- Faster attacks increase the value of predefining your organization’s recovery authority, validation criteria, and governance long before an incident occurs.
- A recent backup should not automatically be a trusted recovery point; organizations need security, integrity, and operational evidence before restoring production services.
- Recovery automation delivers the greatest value when it accelerates approved tasks while preserving human oversight for high-risk decisions and exceptions.
- Recovery readiness should be measured by decision quality, accountability, and business service restoration outcomes rather than by AI adoption alone.
AI Changes the Pace, Not Just the Vocabulary
Calling something an “AI attack” can obscure what actually happened. An attacker may still be stealing credentials, sending phishing messages, or abusing legitimate tools. AI just helps accelerate those methods.
CrowdStrike’s 2026 Global Threat Report describes adversaries using AI to optimize existing techniques and reports an 89% increase in attacks by AI-enabled adversaries compared with 2024. That is not the same as saying 89% of all attacks involved AI.
IBM’s Cost of a Data Breach Report 2026 found that 25% of the breaches studied involved attackers using AI tools, most often for deepfake impersonation.
These findings matter. But I don’t think the answer is to put “AI-powered” in front of every security capability. Asking this AI-related question can be more useful: Where can AI help us remove delay without removing control?
For recovery, I would put that effort into gathering evidence, narrowing recovery options, and explaining why one option deserves more confidence than another. Give the person making the decision something they can inspect and challenge. Don’t just give them a recommendation.
Twenty-Nine Minutes Is Not a Recovery Deadline
CrowdStrike reported that average eCrime breakout time fell to 29 minutes in 2025. The fastest observed breakout was 27 seconds.
Breakout time measures the interval between initial access and lateral movement to another system, not the time to complete an attack or recover from one. This distinction matters. We shouldn’t turn a striking statistic into a claim that every business must recover in seconds.
What I take from it is this: We need to do more to prepare before an incident.
I don’t want the first serious discussion about recovery authority, application dependencies, or validation criteria to happen while systems are unavailable. Those are decisions we can work through beforehand.
Which actions can run automatically? Which require approval? What evidence is sufficient to move forward? What causes the team to stop? These questions are less exciting than a demonstration of an AI agent. But they’re the questions I would want answered before giving that agent authority.
A Recent Protected Data Copy Is a Candidate, Not a Conclusion
“What data should we restore?” sounds like a simple question.
I’d rather ask: “Which recovery point do we have enough evidence to trust?”
My recommendation is to make that evidence explicit. Review anomaly history, malware and indicator scanning, integrity checks, and the dependencies of the service being recovered. Then validate the selected recovery point in isolation before reconnecting it to production.
And be careful with the word clean. I would not treat a scan with no findings as proof that every risk has been eliminated. The recovery decision should include both security validation and confirmation that the application works as the business needs it to work.
Consider an order-processing service. Starting the virtual machine is one milestone. Confirming that users can authenticate, the application can reach its database, and an order can complete are all separate, subsequent milestones.
This is the standard I want recovery plans to work toward: a functioning business service, with evidence supporting its return.
Keep the War Room. Change What It Does.
I’m not arguing for taking people out of recovery. I’m arguing for being more deliberate about where we need them. My recommendation is to automate repeatable, approved work and reserve human attention for exceptions, conflicting evidence, business priorities, and decisions that carry meaningful risk.
The war room should govern recovery. It shouldn’t have to manually execute every step.
At the same time, pre-authorized doesn’t mean unrestricted. A runbook should define the recovery scope, permissions, approval gates, and stop conditions. If the evidence falls outside those boundaries, the workflow should stop and bring a person in.
That’s the kind of automation I’m comfortable advocating: useful, inspectable, and bounded.
Measure the Outcome, Not the AI Label
There is evidence that AI and automation can also make a meaningful difference on the side of security defenders. IBM’s report found that organizations using these capabilities extensively had average breach costs $1.9 million lower and breach lifecycles 65 days shorter.
However, those are study findings, not a guaranteed result for an individual organization or product. For me, the practical test for recovery readiness is asking these questions:
- Can your team explain why it selected a recovery point?
- Can it show what was tested?
- Does everyone know who can authorize the return to production?
At Commvault, the perspective I want to bring to this conversation is recovery first. AI is valuable when it helps people answer those questions sooner and with stronger evidence, but the business doesn’t need a more confident-sounding answer. It needs a recovery decision it can stand behind.
FAQs
Q: What role should AI play in cyber recovery decisions?
A: AI should help teams gather evidence, analyze recovery options, and identify risks more quickly. The final recovery decision should remain accountable to people who can evaluate business impact, security concerns, and operational trade-offs. The goal is better-informed decision-making, not autonomous decision-making.
Q: Why is selecting a trusted recovery point critical after a cyberattack?
A: Restoring the wrong data copy can reintroduce malware, corruption, or operational issues into production systems. Organizations should evaluate recovery points using multiple forms of evidence, including anomaly detection results, integrity checks, security validation, and application testing. Trust should be established through verification rather than assumed from backup recency.
Q: How do faster cyberattacks change recovery planning requirements?
A: Accelerated attacks can reduce the time available to investigate and respond once an incident begins. This makes advance planning more important, including documented recovery procedures, approval workflows, validation standards, and decision-making authority. Organizations that prepare these elements beforehand can help reduce confusion during high-pressure events.
Q: What is the difference between restoring systems and restoring business operations?
A: Recovering infrastructure is only one step in business recovery. Organizations must also validate application functionality, user access, data integrity, service dependencies, and business processes before declaring recovery complete. A system that is running but not delivering business outcomes is not fully recovered.
Q: What are some best practices for governance in automated cyber recovery workflows?
A: Recovery automation should operate within predefined boundaries that include approval requirements, permissions, recovery scope, and stop conditions. Governance frameworks should define when automated processes can proceed and when human review is required. This approach helps balance operational speed with accountability and risk management.
Q: How should organizations measure cyber resilience and recovery readiness?
A: Effective measurement focuses on outcomes rather than technology labels. Key indicators include the ability to justify recovery decisions, validate restored services, document testing results, and demonstrate clear ownership of production recovery approvals. These capabilities reflect practical resilience during real-world incidents.
Curtis Moyer is Field CTO for Security and AI at Commvault.