
Opinion
The real cost of AI security is what goes unchecked
"A model’s accuracy matters only if its decisions arrive in time to prevent damage," writes Bold Security CEO Nati Hazut, "at a cost the organization can sustain across thousands of computers."
Security tools using AI are getting better at recognizing sensitive information. But the way that intelligence is deployed determines how much protection an organization can actually sustain. If checking more activity makes a system too expensive or slows employees down, teams start checking less. Accuracy is only part of the equation; the cost and speed of applying it matter just as much.
Data loss prevention (DLP), which is meant to stop sensitive information from leaving an organization, makes that tradeoff especially clear. DLP earned its bad reputation for good reason. Tools ran on fixed rules, which meant whole teams existed to write and maintain policies that went stale within days or weeks of shipping. The rules were crude: they often couldn’t tell a customer list from a template that mentioned customers, so they fired constantly. And a system that fires constantly gets turned down to logging. It didn’t prevent much. It produced work.
AI changed what these systems could understand. A model can recognize confidential information in a document without anyone having to manually create rules and labels for it. It can weigh context rather than match patterns, which cuts the noise that made blocking unworkable. But putting a better model into the same architecture creates a different problem: the system now has to pay for that intelligence every time it uses it.
Many “next-gen” or “AI-native” security tools bolt AI onto the architecture that was already there. When the model runs in the cloud and each inspection requires a remote verdict, the file travels to it and the answer travels back. That is expensive, because each inspection is an inference someone is paying for, and the bill scales with how much of the company you inspect. It also adds a round trip before the system can allow or stop a transfer.
For an employee trying to send a file, that delay means waiting for the security check before work can continue. Examining stored files in advance reduces the work left until that moment. But it also means processing information before anyone tries to move it, and keeping those classifications current. Repeatedly sending that work to a cloud model raises the bill. A design that reduces the delay at the point of transfer can still be too expensive to run continuously.
Those costs and delays lead to compromises, and they rarely look dramatic. To control spend, one team scans less frequently while another excludes large files or a busy group of employees. To keep work moving, a policy that once blocked a risky action is changed to record it instead. Each decision eases an operational burden, and each one reduces the protection the system actually delivers. A system cannot make accurate decisions about activity it has been configured to ignore.
My work in connected-device security at SAM taught me to think about the resources available where protection has to run. Cloud security at Polyrize taught me something else: knowing where data is stored, or who can access it, does not mean you can intervene when it moves. AI brings both lessons together. Understanding the information has to be efficient enough to keep running and fast enough to stop a risky transfer before it is complete. Those requirements have to shape the architecture from the start.
That is why the location and timing of the analysis matter. Small AI models built for specific tasks can analyze sensitive content on the employee’s computer, removing the need to send the file to a cloud model and wait for its verdict. Classifying a file when it reaches the computer gives the system information it can reuse across related checks. When someone tries to share the file, the system combines that analysis with the context of the action to decide whether to allow it. Less work is left until the employee is waiting for an answer.
Local processing replaces a per-inspection cloud inference charge with work on the computer’s own processor and memory. That distinction matters: it removes that recurring service cost, but the employee’s applications still need those resources. A tool described as lightweight while idle has to stay that way when examining files all day. Otherwise, moving the model has simply moved the bottleneck. Efficient local models, early classification and reuse of previous analysis have to work together to keep protection running without slowing the employee down.
This is what an accuracy-focused proof of value can miss. A predefined dataset demonstrates whether a model recognizes sensitive information; it does not, by itself, show what happens when the system runs continuously across 50,000 computers. Response time, resource use and the cost of maintaining coverage belong in that evaluation, with the files and workloads employees actually use.
The danger is buying a system for its accuracy and then configuring away the coverage that makes that accuracy useful. The model can keep performing exactly as advertised while fewer files and fewer actions are checked. Cost and latency have then stopped being implementation details: they are deciding how much of the organization is protected.
Nati Hazut is the co-founder and CEO of Bold Security.














