AI agents can take the first page, but humans still own the outage
In 2026, AI agents are moving into IT service operations faster than the controls around them. Downtime is expensive, with high-impact outages at a median cost of $2 million per hour in New Relic's 2025 Observability Forecast, so teams are looking to AI for triage, routing and first response. The open question is no longer whether agents can help on call, but who checks them when they act.
What do outages and incidents cost in 2026?
Enough to justify serious automation. New Relic's survey of 1,700 IT and engineering professionals, run in April and May 2025, puts the annual median cost of high-impact outages at $76 million per business (New Relic). PagerDuty's 2024 survey of 500 IT leaders estimated downtime at $4,537 per minute and an average resolution time of 175 minutes, so each incident can cost nearly $794,000 (PagerDuty). Volumes were rising too: 59% of those IT leaders said customer-impacting incidents had increased, by an average of 43% over 12 months among that group (PagerDuty). Both are vendor surveys built on respondents' own estimates, so read them as orders of magnitude.
Independent data points the same way with more nuance. The Uptime Institute's Annual Outage Analysis 2025 found outage frequency and severity declining for the fourth consecutive year, yet 54% of respondents said their most recent significant outage cost more than $100,000, and one in five said more than $1 million. Fewer outages, but each one still hurts.
Visibility changes the bill. New Relic respondents with full-stack observability reported high-impact outages averaging $1 million per hour, against $2 million for those without, and fewer of them had such outages every week (New Relic). That is a correlation in a vendor survey, but it explains why ops teams invest in context before they invest in automation.
High-impact outages with and without full-stack observability
| Criterion | Measure | With full-stack observability | Without |
|---|---|---|---|
| Cost per hour of a high-impact outage | $1M | $2M | |
| High-impact outages at least weekly | 23% | 40% |
Source: New Relic, 2025 Observability Forecast (n=1,700, April to May 2025; vendor)
How far has AI gotten into IT operations?
Further than most engineers expect. Use of AI monitoring capabilities rose from 42% in 2024 to 54% in 2025, and 48% of organizations are increasing investment in AIOps and machine learning capabilities (New Relic). In PagerDuty's April 2025 survey of 1,000 IT and business executives, 51% of companies were already using AI agents, and 86% expected to be operational with agents by 2027 (PagerDuty). By September 2025, 75% of companies in a separate PagerDuty survey of 1,500 executives had deployed more than one agent (PagerDuty).
The service desk shows early results. In SolarWinds' 2025 report, a vendor study, average incident resolution time fell from 27.42 hours to 22.55 hours after generative AI features were enabled (SolarWinds). In its 2026 survey of 844 IT professionals (also vendor research), 84% said AI met or exceeded ROI expectations, but only 23% said it significantly exceeded them (SolarWinds). Useful, then, but rarely transformative yet.
Who supervises the agents?
Here executives and engineers disagree. In PagerDuty's September 2025 survey, 81% of executives said they trust AI agents to act on the company's behalf during a crisis such as a service outage (PagerDuty). Engineers are more careful. In a Stack Overflow pulse survey of about 1,100 technologists, published in May 2026, 60% block agents from making unapproved system changes and 63% rarely or never let agents run fully on autopilot.
Production is where developers draw the line. In the Stack Overflow Developer Survey 2025, 75.8% said they do not plan to use AI for deployment and monitoring, the most resisted task in the survey. The 2026 results show only 20% use AI to deploy, operate or troubleshoot production systems. The people closest to the pager want agents that gather context and propose, with a human who approves.
Executive views on AI agents and AI safeguards, 2025
- Need better procedures to detect AI errors or failures
- 85%
- Had at least one AI-related outage
- 84%
- Trust AI agents to act during a crisis
- 81%
- Deployed more than one AI agent
- 75%
- Test AI-generated code consistently through formal processes
- 39%
What are the main challenges with AI agents in service operations?
Five stand out in the recent data.
- AI tools cause outages too. 84% of companies in PagerDuty's 2025 resilience survey had at least one AI-related outage, and 85% of executives said they need better procedures to detect errors or failures in AI tools (PagerDuty).
- Testing is informal. In the same survey, 85% test AI-generated code, but only 39% do so consistently through formal processes (PagerDuty). Runbooks that an agent can execute deserve at least the same rigor.
- AI adds work before it removes it. 52% of SolarWinds respondents say their workload increased after adopting AI, and 47% spend time reviewing and validating AI-generated content (SolarWinds).
- Detection still depends on customers. 41% of leaders learn about service interruptions through customer complaints, tickets or manual checks, and engineers spend 33% of their time fighting fires (New Relic). An agent cannot triage what nobody detects.
- Manual process is the expensive default. Organizations with five or more manual incident-response processes reported $30.4 million in annual costs of customer-facing outages, against $16.8 million for those with five or more fully automated (PagerDuty). Uptime adds that staff failing to follow procedures has become a bigger cause of outages than the year before, which argues for procedures that are executable and tested.
What does good look like, and how do you start?
Good agentic operations look boring: agents do the reading, people do the deciding, and every action leaves a trail.
- Start read-only. Let an agent enrich alerts and tickets with logs, metrics, recent deployments and similar past incidents. Measure time to first useful response.
- Route and deduplicate next. Ticket classification and routing are low-risk, high-volume and easy to check against human decisions.
- Turn runbooks into tested actions. Parameterize each routine fix, test it in staging, and expose it to the agent as a proposal, not a command.
- Require approval for writes. Restarts, config changes and rollbacks go through a named person, with a full log and a rollback path.
- Watch the agent like a service. Track its suggestions, approvals, overrides and mistakes, and treat an agent error as an incident with a review.
- Expand autonomy only with evidence. Move an action to auto-approve when its record shows it is safe, and keep the switch to take it back.
How can SDCLabs help with agentic ServiceOps?
SDCLabs builds agents into the tools your operations team already runs, such as your service desk, monitoring and chat. AI agents draft integrations, runbook actions and tests; senior engineers review and decide what ships; automated tests verify every change. Our founders created the open-source testing frameworks Codeception and CodeceptJS, so every runbook an agent can execute gets a test before it reaches production.
If your on-call team is drowning in alerts and tickets, see our Agentic ServiceOps service or tell us how your operations run today. Related: our cloud operations and managed services and our report on AI across the SDLC.
Sources checked on 2026-11-17.
Frequently asked questions
How much does IT downtime cost?
In New Relic's 2025 Observability Forecast, high-impact outages had a median cost of $2 million per hour and $76 million per year. These are vendor survey estimates from 1,700 respondents.
How widely is AI used in IT operations?
Use is growing. In New Relic's 2025 Observability Forecast, use of AI monitoring capabilities rose from 42% in 2024 to 54% in 2025, and 48% of organizations are increasing investment in AIOps and machine learning.
Should AI agents make production changes on their own?
Most engineers say not without approval. In a Stack Overflow pulse survey, 60% block agents from making unapproved system changes and 63% rarely or never let agents run on autopilot.
Do AI tools cause outages?
Often, by companies' own account: 84% of companies in PagerDuty's 2025 resilience survey had at least one AI-related outage, and only 39% consistently test AI-generated code through formal processes.
