Most large enterprises don't have a vulnerability problem. They have a fix problem. Scanners, risk models and exposure platforms have gotten very good at telling you what's broken. Getting it repaired is another matter, and that's where the industry keeps coming up short.
Look at what the market has built over the past decade: scanning, exposure management, threat intelligence, exploitability scoring, asset criticality models. Each layer makes the list of findings more precise. None of them moves a patch into production.
Where the backlog actually comes from
Take a critical vulnerability at a big enterprise. The security team rates it, writes it up, opens a ticket, assigns a priority and names an owner. On paper, that's a working process. In practice, it's usually where things jam.
The infrastructure team is running aging systems with tangled dependencies. The development group has a release calendar that won't bend. Production operations doesn't want more downtime. The business owns a revenue-generating service and wants it left alone. Some systems have maintenance windows measured in minutes. Others need weeks of regression testing before a patch can go in. Legacy applications rely on components nobody can upgrade without breaking something else.
Then there's the ownership question, which is often contested. Security teams are typically responsible for finding and reporting weaknesses but have no authority to force a fix. The teams that control the systems have their own priorities and limited capacity. The outcome is familiar: tickets that escalate, exceptions that pile up, deadlines that slip and a backlog that never shrinks. A better severity score doesn't resolve any of that.
Prioritization helps, but it stops short of the outcome
Nobody can fix everything, and the industry has accepted that. Risk-based vulnerability management exists because there's a real difference between an actively exploited flaw on an internet-facing critical system and a theoretical one sitting on an isolated, low-value box. Ranking those correctly is worth doing.
But a perfectly ranked backlog is still a backlog. In most organizations, the hard decision isn't which vulnerability matters most. Security and engineering can usually agree on that in a meeting. The hard part is agreeing on who fixes it, when the change happens, who can overrule a business unit that refuses downtime and who signs off on accepting the residual risk when a fix can't land in time. Those are governance questions, and governance is where the tooling market spends the least energy. Vendors keep pushing deeper into the "what." The bottleneck sits in the "who" and the "how."
AI makes discovery cheaper and the queue longer
AI is changing the economics of finding bugs. AI-assisted tools can analyze code and surface weaknesses at a pace that used to require large teams of human researchers. Google Threat Intelligence Group reports that monthly vulnerability disclosures more than doubled between January and August 2026, and that observed exploitation rose significantly over the same period. AI is part of that acceleration, though changes in disclosure practices also affect the numbers.
Here's the thing: discovery capacity is scaling fast, and remediation capacity isn't. If findings multiply while an organization's ability to patch stays flat, the result is predictable. Bigger backlogs, more pressure on engineering teams, more exceptions, more fights over priority. Exposure can grow even as visibility improves. The Cloud Security Alliance has already pointed to this gap between how quickly vulnerabilities are found and how well organizations can absorb and fix them.
That's the emerging problem. AI is industrializing discovery, while remediation still runs through organizational structures built for a slower era.
AI can help fix things, but it can't own the outage
AI isn't irrelevant to remediation. It can build proofs of concept, draft and test fixes, suggest configuration changes, map dependencies and automate parts of deployment. Organizations should adopt those capabilities where they make sense, and some of the gains will be large.
Still, the limits matter. AI can't decide on its own that a critical production system goes offline during peak trading hours. It can't absorb the accountability for disrupting customers, or referee a dispute between a business unit, an infrastructure team and the security function. It can't wave away legacy dependencies or contractual commitments. Even when an AI system produces a working fix, someone has to validate it, approve the deployment, confirm nothing else broke and verify that the vulnerability is actually gone. Some of those steps can be automated. Others need judgment, authority and coordination across teams. The more capable the tooling gets, the more exposed those human chokepoints become.
What large organizations actually need: an operating model
Better scanners won't close this gap. What's needed is a remediation operating model with clear accountability across security, development, infrastructure, operations and the business. That model has to answer some blunt questions. Who owns fixes for shared platforms and complex applications? Who allocates the people and budget to do the work? Who can authorize an emergency change? Who breaks the tie when remediation collides with a business priority? How are maintenance windows negotiated, and who enforces them? When a fix can't be completed in time, who accepts the residual risk, and who checks that the exception is still justified next quarter? And the question that matters most: who makes sure agreed actions actually happen? Those answers can't live in informal relationships between engineers or depend on one persistent security manager pushing hard enough. They need to be written into management processes, backed by executive authority and tracked through measures that someone reviews. Security provides risk intelligence and independent challenge. Technology teams own delivery inside their environments. Business leaders weigh in when downtime or investment trade-offs can't be avoided. The CISO's job is to build and oversee that machinery, not to chase tickets.
Measure fixes, not activity
Many programs still measure the wrong things. They track how many vulnerabilities were found, how severity is distributed, whether deadlines were met and how many findings remain open. Those numbers have a place, but they reward an administrative mindset in which closing a ticket stands in for reducing risk. The better question is whether material exposures are gone within an acceptable time, and whether the organization can prove a fix worked. Useful signals include recurring remediation failures, high-risk exposures that keep aging, the bottlenecks that slow delivery, the quality of exception handling and how long it takes to ship a verified fix. The gap between "marked resolved" and "actually remediated" is where a lot of risk hides. AI can produce more accurate and timely evidence of that progress. It can't take responsibility for the results. Auditability is part of the picture as well. Organizations increasingly need to show that vulnerabilities were addressed within reasonable timeframes, with documented decisions and traceable exceptions. As discovery speeds up and parts of remediation get automated, a clear chain of accountability matters more, not less.
Point AI at execution, not just discovery
This is where I think the industry should focus. Using AI to find more bugs and rank them more elegantly has diminishing returns when the queue is already too long. Vendors and practitioners should aim at the whole remediation lifecycle: mapping dependencies, identifying safe deployment windows, preparing changes, automating tests, coordinating workflows, validating results and giving management reliable evidence of progress. In heavily automated environments, AI agents may eventually run large parts of that process within tightly defined controls. But automation for its own sake isn't the goal. The goal is less friction between knowing about a problem and fixing it. That only works if the technology is designed around a governance model that already functions. Governance won't emerge on its own from a new platform.
The technology changes; the management problem doesn't
For years, I've argued that cybersecurity is a leadership and management problem that technology helps solve, not the other way around. Vulnerability management shows that clearly. The problems that kept organizations from patching critical systems fifteen years ago are still here: fragmented responsibilities, competing priorities, operational constraints, weak governance and thin follow-through. AI won't make those problems go away. By shortening the time between discovery and exploitation, it's exposing them more harshly than before. The organizations that come out ahead won't necessarily be the ones with the sharpest discovery tools. They'll be the ones that built the accountability and operational discipline to turn technical findings into fixes that stick. Finding what's wrong has never been the hardest part. Getting it fixed is, and that's a management job.