■ HIGH RISK ■ Technology
Not replaced, but thinned out and reshaped. The toil half of SRE — dashboards, runbooks, capacity math, first-pass triage — is exactly what machines are good at, while the judgment call of whether to fail over a region at 3am stays with a human who can be held responsible.
“AI can monitor dashboards and auto-scale. But at 3am when everything's on fire? They'll still page you.”
Our AI replacement risk score — how we score jobs
An SRE's week is a strange blend of software engineering and emergency medicine. You write Terraform and controllers, argue about error budgets with product teams who want to ship anyway, tune alert thresholds so the pager stops crying wolf, run load tests before Black Friday, and write the postmortem nobody wants to read. Then a certificate expires at 2:47am and you spend ninety minutes chasing a cascading failure through four services, none of whose owners are awake.
The automatable slice is large and growing. Autoscaling and self-healing controllers already handle capacity events without a human. Anomaly detection over metrics and traces beats hand-tuned static thresholds. LLM agents wired into observability stacks can correlate a latency spike with a deploy, summarize the blast radius, draft the incident timeline, and propose the rollback command — the grunt work that used to eat the first twenty minutes of every incident. Runbook execution, config drift remediation, and dependency upgrades are increasingly scripted rather than staffed. Our risk score of 52 reflects that: a real squeeze on headcount per service, not an extinction event.
What resists is the part where the model is wrong and the stakes are real. Novel outages are novel precisely because they fall outside the training distribution — the failure mode nobody encoded, the third-party provider degrading in a way its own status page denies, the fix that risks data loss. Someone must decide to shed traffic, accept the customer impact, and own it in the review. Equally human: negotiating reliability targets with engineering leaders, designing systems so they degrade gracefully in the first place, and pushing back on architectures that look fine in a diagram and terrible under load. The SREs who get squeezed are the ones acting as human alert routers. The ones who design for failure get more leverage, not less.
Automatability: our editorial assessment of current and near-term AI capability
Pressure is already visible: teams that used to staff a follow-the-sun rotation now run leaner because automated remediation handles the routine third of incidents. Expect the squeeze to sharpen by around 2030, when AI incident copilots are standard in observability platforms and per-service headcount drops further. Total elimination isn't on the table — regulated and high-revenue systems will keep humans accountable for failover decisions well beyond that.
Yes, with a caveat. Demand for people who can design reliable distributed systems remains strong, but the entry-level path of watching dashboards and following runbooks is evaporating. Come in with real software engineering ability — you'll be expected to write the automation rather than execute manual steps that a controller now handles for free.
It can run the first ten minutes of a familiar one: correlate the alert with a recent deploy, summarize impact, suggest a rollback. What it can't do is decide to accept customer data loss to restore service, or coordinate five teams and a vendor at 3am. Novel failures are, by definition, outside what the model has seen.
It will shrink them. Auto-remediation already resolves a meaningful share of routine pages before a human wakes up, and that share keeps growing. But the pages that survive are the nasty ones, so on-call gets less frequent and more intense. Expect smaller rotations covering more services rather than no rotation at all.
Systems design under failure, cost-and-reliability tradeoffs, and how to evaluate an AI agent's proposed fix before it runs in production. Also worth investing in incident command skills and writing — the work that survives is increasingly about judgment and communication, and both show up in postmortems that change how the org builds.