The measurements that earn an AI agent autonomy are the ones your CFO has been asking for.
The quarterly security slide has looked the same for a decade. Vulnerabilities discovered: 14,000. Triaged: 9,200. SLA attainment: 94 percent.
Every number on that page describes what the security team did. None of them describes what the business got.
Alan Mitchell, a global CISO whose career runs from mainframes through client-server, the cloud, and now AI, made the distinction the spine of a recent Backline Unfiltered conversation. Volume metrics count activity. Value metrics count outcomes, and security programs have spent fifteen years reporting the first while their executives needed the second.
Counting alerts made sense when alerts were hard to get
Fifteen years ago, scanner coverage was a genuine achievement. Running a scan across every repository, host, and container image took budget and a year of work, so counting what came back proved the program worked.
The counting habit outlived the achievement. Detection is close to free now: open-source scanners and public vulnerability data do the mechanics, and every commercial platform runs broadly the same engine underneath a nicer UI.
A number that used to prove capability now measures how much code you own and how many scanners you bought.
The habit also generates its own busywork. Plenty of teams burn real hours arguing a Critical down to a High so the SLA clock resets, which produces a better-looking dashboard without moving risk.
The intake side moved as well. Agents now write code faster than any review gate reads it, so the discovery number climbs every quarter on its own, without anybody doing better security work.
A metric that rises whenever your team ships more measures throughput rather than exposure.
Three weeks to a day
Mitchell described a security process that used to consume three weeks of human effort and now finishes inside a day of machine reasoning. The number has the shape of a value metric: capacity returned to the organization, denominated in a currency your CFO already recognizes.
Volume asks how many issues you found. Value asks what the company got back.
For an AppSec lead, the practical version is a short list:
- Exposure genuinely reduced this quarter, rather than backlog re-ranked
- Acceptance rate on security-generated fixes
- Age curve of your dependency debt
- Time from disclosure to merged fix on the assets that carry revenue
The acceptance rate deserves a second look, because it’s the only line on the list that measures whether your output will be used. A team with excellent coverage and a 30 percent acceptance rate has a distribution problem wearing a detection problem’s clothes.
As good as a human, measured honestly
The blocker on value metrics is autonomy, and the blocker on autonomy is a standard nobody applies to people.
The clock adds its own pressure. Time from public disclosure to active exploitation has compressed from years into hours, which leaves human-paced remediation permanently behind the attacker no matter how elegantly you rank the queue.
Mitchell’s litmus test runs on mistakes per number of operations. Can the agent do the work as well as a competent human? Humans introduce regressions and merge risky changes late on a Friday, yet nobody asks a senior engineer for a zero-defect record before granting merge rights.
Hold an agent to perfection and you’ll hold it there permanently, which leaves your backlog exactly where it sits today.
So Mitchell inverts the old mantra: verify and trust, in place of trust but verify. Trust becomes the output of measurement rather than its precondition.
In practice, verification lands before a human ever reviews the fix. The existing unit tests and regression suites run first, and the agent records why it picked that fix over the alternatives.
Humans stay in the loop early, then move to spot checks and telemetry once the error rate holds steady, and hands-off comes last, if the data earns it.
Every one of those gates produces a number. The same numbers that earn an agent more autonomy are the numbers that belong on your board slide.
Budget follows the scorecard
A program measured on detections keeps buying detection. Renewal season rewards the tool that produces the biggest number, and the biggest number has never been the same thing as the lowest risk.
A program measured on exposure reduction buys remediation capacity instead, because nothing else moves the metric.
Mitchell’s “garbage in, garbage out” warning applies to the scorecard itself. Value metrics depend on a clean, auditable record of what actually got fixed, by whom or by what, and whether it held under test.
If your remediation history lives across three ticketing systems and a spreadsheet, the new numbers will be as fictional as the old ones.
Rewrite the slide
Take two volume metrics off your current board deck and swap them out. Exposure reduction replaces vulnerabilities discovered. Acceptance rate replaces tickets opened.
Your first quarter under the new scorecard may look worse, because honest numbers usually do. The second gives you numbers a CFO can fund.
Take a leaf from our client’s story, which shows what the reporting looks like when remediation runs autonomously. Report what the business got back, and the budget conversation (almost) takes care of itself.
