The Pager Still Rings: Inside the 2 A.M. Rituals AI Observability Was Supposed to Kill

There's a particular flavor of irony in modern software operations: a startup sells AI-driven anomaly detection to enterprise customers, then its own engineers get woken up at 2:14 a.m. by a PagerDuty alert that the AI saw coming three hours ago.

It happened last quarter at a fintech shop in Austin. The anomaly detector flagged a slow memory leak on a payments worker at 11 p.m. Nobody acted — the alert looked 'low confidence,' whatever that means — and by 2 a.m. the service was down and a senior engineer was squinting at a laptop in the dark, muttering at both the bot and himself.

If AI observability is so good, why is the on-call rotation still a rite of passage?

The Tools Got Smarter. The Rituals Didn't.

Talk to engineering leads at US startups and you hear the same story on repeat. They've bought the modern stack — anomaly detection, predictive alerting, LLM-powered incident summaries. And they still run rotations that would look familiar to a sysadmin in 2009.

The reasons are less technical than cultural. On-call isn't just a monitoring strategy; it's a accountability structure. When something breaks, a named human owns it. An AI model that predicts a failure 80% of the time is genuinely impressive. It's also genuinely useless for the 20% where it stays quiet, and nobody wants to be the team that explained to customers that the model didn't feel confident.

'We trust the AI to tell us something is wrong,' one platform engineer in Denver told me. 'We don't trust it to tell us nothing is wrong.'

That asymmetry is the whole ballgame. False negatives are catastrophic and public; false positives are annoying and private. Humans are wired — and orgs are structured — to over-index on the catastrophe.

Alert Fatigue Meets Alert Distrust

Here's the uncomfortable part: many teams aren't ignoring their AI monitoring out of skepticism. They're ignoring it out of exhaustion.

Early anomaly detection tools earned a reputation for crying wolf. Thresholds were too twitchy, 'low confidence' alerts arrived in droves, and within a month engineers had muted the channel. Now the industry's better models are drowning in the distrust earned by the worse ones. It's the Boy Who Cried 5xx.

The result is a weird hybrid workflow that nobody designed: AI watches everything, triages most things, and escalates the scary stuff to a human — who then double-checks the AI's work manually, because the last three escalations were noise.

So the human is still in the loop at 2 a.m. Not because the AI can't do the job, but because the organization hasn't built enough trust to let it. Trust in ops, like trust anywhere, is earned in incidents and lost in false alarms.

The Comfort of a Human on the Hook

There's also a quieter reason the pager survives: it's comforting.

Founders like knowing a person is accountable. Engineers like knowing a person can override the machine. Customers like hearing 'our senior engineer was on it immediately' more than 'our model detected the anomaly and auto-remediated.' The on-call rotation is the software industry's version of a smoke detector with a face.

And honestly? That's not entirely irrational. AI observability is great at pattern recognition across millions of metrics. It's still mediocre at judgment calls — is this degradation acceptable during a product launch? Do we roll back or ride it out? Those decisions carry business context that no model has fully ingested yet.

What Would It Take to Actually Retire the 2 A.M. Page?

A few things, and none are model improvements:

• Track the AI's record publicly. Teams should know their detector's precision and recall the way they know their uptime. If it catches 95% of incidents with a 2% false-positive rate, trust follows.
• Give the AI authority, not just a voice. Auto-remediation for well-understood failure modes — restart the worker, fail over the region — with humans reviewing after the fact. Most 2 a.m. pages are routine; let the routine be routine.
• Redefine on-call as oversight. The rotation doesn't die; it changes shape. Fewer pages, higher stakes, and a standing audit of what the machines handled while everyone slept.

Until then, American dev teams will keep living in the gap between what their tools can predict and what their cultures will permit. The AI sees the failure coming. The human still wakes up for it.

Somewhere tonight, an anomaly detector in Seattle flagged a problem at 11 p.m. and a engineer's phone buzzed at 2 a.m. anyway. The tech works. The trust hasn't shipped yet.