Is MTTR still the right success metric when AI can "resolve" symptoms faster than it understands root cause?

The shift

AI makes fast, plausible-looking mitigation abundant — restart the pod, roll back the deploy, reroute traffic, suppress the alert — because pattern-matching against thousands of past incidents and runbooks is now cheap and instant. What stays scarce is the causal model: knowing why the fix worked, whether it actually removed the fault or just hid it, and whether it will recur under different conditions. MTTR measures the abundant part and silently assumes it correlates with the scarce part.

The axioms

• MTTR is a proxy for "the problem is gone," not just "the symptom is gone." Rests on the assumption that resolving the alert and resolving the fault are the same act — true when a human has to understand the system to make the alert stop, false when a tool can make the alert stop without that understanding.
• A lower MTTR reflects better incident response capability. Rests on the scarcity of fast diagnosis — historically, speed only came from an engineer who already had a correct mental model, so speed was a proxy for skill/understanding.
• Postmortems and root-cause analysis happen after the incident is closed, on a slower clock than resolution. Rests on the scarcity of engineering time — RCA got deprioritized once MTTR was hit, because analysis was expensive and the incident was already "over" per the metric.
• On-call engineers are accountable for both stopping the bleeding and understanding the wound. Rests on accountability being scarce and human — one person owns the incident end to end, so speed and understanding were bundled in the same accountable party.
• Recurrence rate is a lagging indicator that eventually corrects for a bad MTTR obsession. Rests on the assumption that if root cause wasn't found, the same failure mode will resurface often enough, visibly enough, to get caught.

Invalid axioms

1. MTTR-as-proxy-for-resolution. AI flips the scarcity of fast mitigation — an agent can pattern-match "this looks like the OOM-kill incident from March, restart and scale" in seconds, closing the alert without any causal claim being verified. Teams that still price/staff around "time to green" as the north star are optimizing a number that AI can now hit while root cause quietly persists. The habit-trap: dashboards and SLAs built around MTTR keep getting reported as health signals when they've become a measure of AI's mitigation speed, not the org's understanding of its own systems.
2. RCA-as-separate-slower-phase. The reason RCA was deferred to a post-incident review days later was that synthesizing logs, traces, deploy history, and past incidents into a causal narrative was slow, expensive human work. AI makes that synthesis close to instant and cheap enough to run inline, during the incident, not after it. Keeping RCA as a scheduled Tuesday meeting long after the page cleared is now a scheduling artifact, not a technical necessity — the constraint that justified the delay is gone.

Unchanged axioms

1. Someone accountable for whether the fix was actually correct. AI can propose or execute a mitigation, but it cannot be liable for a recurrence, cannot be fired, cannot answer to a customer or regulator. Verification that the causal story is true — not just plausible — still needs a human who stakes their judgment on it, especially before an AI-generated "root cause" gets written into a postmortem as fact.
2. Judgment on novel failure modes with no precedent. Pattern-matching against past incidents is exactly where current models are strong; it's also exactly where it breaks down for the failure that doesn't resemble anything in the training distribution — a new class of cascading failure, a subtle data-corruption bug, a multi-system interaction nobody's seen. That's still an act of hypothesis-driven reasoning under uncertainty that a human has to drive, because there's no pattern to retrieve.
3. Trust in the on-call rotation as the safety net when automation is wrong. Confidently-wrong mitigation is the default failure mode of a probabilistic system — a plausible-sounding fix that's actually masking the real fault (or making it worse) is a normal outcome, not an edge case. The org still needs people who are trusted to override the automation, and that trust is built the slow way, through track record, not granted by default to a tool.

New axioms

1. Distinguishing "resolved" from "silenced" at scale. When mitigation is cheap and instant, an org can rack up a long streak of great MTTR numbers while accumulating a backlog of unresolved root causes it doesn't have visibility into, because each individual incident looks closed. What replaces MTTR (or supplements it) as the metric that catches this before it compounds into a major outage is unsolved.
2. Verifying an AI-generated root-cause narrative before it becomes institutional memory. If synthesis is cheap, a plausible-sounding causal story gets written into the postmortem quickly and confidently — and once it's written down, people stop questioning it. The org needs a cheap, fast way to red-team or falsify that story, or it just accumulates confidently-wrong explanations at the same speed it used to accumulate correctly-slow ones.
3. Recurrence rate as the new lagging indicator is itself getting slower to trigger. If AI mitigation adapts each time — slightly different restart, different reroute, different masking — the same underlying fault can present as a series of "different" incidents that never quite repeats identically, defeating the pattern-matching that used to catch recurrence. What counts as "the same incident happening again" gets harder to detect exactly when it matters more.

Where it breaks

A team keeps MTTR as its headline SLA and lets an AI agent auto-remediate at machine speed — the dashboard goes green, the metric looks best-in-class. Meanwhile nobody's verifying whether the auto-remediation is masking a slow-building structural fault, because the org's real bottleneck — a human confirming the causal story — never got scaled alongside the mitigation. The postmortem, if it happens at all, inherits an AI-drafted root cause that reads as authoritative and gets rubber-stamped, because checking it is exactly the slow, expensive work the team just proved it could skip.

ai-axioms.vercel.app/axioms/engineering/sre-mttr-metric-ai-axioms