PagerDuty’s AI chief says outage response is becoming fully autonomous
A top AI executive warns self-acting systems can fail unpredictably, as most employees use AI tools without approval
AI agents are taking over the detection, triage, diagnosis and repair of technology outages, pushing incident response toward full autonomy at a moment when disruptions are becoming a near-daily cost of doing business online. The change removes engineers from much of the work that once demanded round-the-clock attention.
Systems increasingly investigate and resolve problems on their own before a human is even called in. The stakes are high, since a single hour of downtime can now cost a large business millions of dollars.
“We’re getting to a point where these diagnoses become so good that you’re able to suggest very good recommendations in order to solve these issues,” João Freitas, Chief AI Officer of PagerDuty, told TechJournal.uk in an interview. “We’re starting to see that it’s possible to have completely autonomous incident management.”
“You could have the agent immediately start to investigate what’s wrong,” he said. “When a human gets into the issue, or gets into the on-call, it already has all the context, the triage and the investigation prepared.”
PagerDuty has launched multiple agents over the past two years, covering tasks within an incident as well as work such as managing on-call shifts and verifying that event pipelines are configured correctly. The company had already run automation workflows before generative AI arrived, built in part through acquisitions in 2021 and 2022.
“Agents tend to be non-deterministic, so the combination of the powerful agents that help you through this lifecycle of the incident and the deterministic workflows is quite interesting in terms of value for the customer,” he said.
“Agents can help you not only detect, triage and get you faster to the right diagnosis, but they can also help write the post-incident review for you,” he said. “You capture much more information, and that information can be fed again into the system, so that you are continuously learning from your problems.”
“Our goal is for teams to get more proactive and preventive in the time they spend on operational issues,” he said. “That time should instead go to innovation and to things that are valuable for the business.”
He leads artificial intelligence and engineering at PagerDuty, listed on the New York Stock Exchange as PD.
The company began as a 2009 spin-off from Amazon focused narrowly on incident response, and now serves more than 35,000 organizations worldwide.
When agents go wrong
Autonomy cuts both ways. The same systems capable of resolving problems without a human are also capable of making decisions nobody asked for, and by their nature, they do not fail the way older software does.
“AI also tends to fail a little bit differently from a traditional system,” Freitas said. “It is non-deterministic. If you have an agent that is talking to another agent and this agent makes a mistake, then it propagates to the other agent.”
He said errors compound quickly in agent-to-agent systems, because each agent pulls data from multiple sources and acts on it. A single wrong step early in the chain can escalate into a much larger business risk before anyone notices.
A second failure mode is hallucination, in which an agent simply invents an incorrect number that someone then acts on.
“Imagine you have an AI that is interacting with your customers, and your customer says, ‘I’m not sure if I want to cancel my trip,’ and the AI goes, ‘Okay, I’ve canceled your trip,’” Freitas said. “The AI made something that the user was not asking for.”
“When AI fails, we catch that as early as possible and correct it and make sure the system is back working again,” he said. “There are some discussions on how we can make sure that AI is doing things according to the policies of the enterprises, and that the agents are accessing the right data.”
The governance gap he described is already visible inside companies.
A PagerDuty survey of 1,250 office professionals across the United States, the United Kingdom, Australia, and Japan found that two-thirds have used AI tools at work that they believed were not permitted under company policy.
Eighty-eight percent said they had shared work-related information with public AI tools such as ChatGPT, Claude or Gemini, including emails, meeting notes and, in nearly a third of cases, financial or confidential company data.
About 75% of respondents said they would likely look for a new job that offered better AI skills development, a figure that climbed to 80% at companies with $1 billion or more in revenue. Most also said they trusted their own AI judgment over their employer’s technology team.
PagerDuty Chief Technology Officer Tim Armandpour said that once more than 30% of employees are feeding confidential data into public models, shadow AI becomes a serious enterprise liability.
He said the answer is not to slow adoption, but to redirect that energy into platforms that offer governance and control.
A decade of downtime
That original incident-response mission has expanded considerably since.
PagerDuty went public in 2019. Nearly half of the Fortune 500 and about two-thirds of the Fortune 100 use its platform.
“Today we address the entire incident management space, and also what Gartner calls AIOps (AI for IT Operations),” Freitas said. “We’re focused on automating everything that is operations.”
PagerDuty acquired several automation companies in 2021 and 2022, then layered machine learning onto event processing for noise reduction and alert grouping, before generative AI agents arrived on top of both, he said.
None of that infrastructure is short of work. Freitas said major outages are no longer rare events.
“I’d say almost every day. Every day you have incidents,” he said. “There was the big CrowdStrike incident that happened in 2024 that shut down airports and several systems.”
He said that when AWS, Azure, or Cloudflare goes down, typically 20% to 30% of the internet starts to complain because those providers carry workloads for many other companies. A single outage can ripple through customers on multiple clouds at once.
PagerDuty’s research found that 8% of organizations lose more than $1 million an hour during outages, with more than half saying the damage also extends to brand reputation.
“Almost 70% of incidents are caused by changes,” Freitas said. Someone typically changes something that later causes a problem, and in distributed systems, where one service connects to another, the impact can take time to trace.
That complexity is at the center of PagerDuty’s pitch against homegrown alternatives, including the growing number of small teams building their own AI agents for incident response.
“I don’t think they will be our competitors,” Freitas said. “In order to have a reliable system, you want the system to operate and be always on. If you’re developing something yourself, you need to maintain it and guarantee the reliability.”
“It’s not scalable for you to put people just looking at the systems,” he said. “That’s not how software works.”
Freitas said a homemade alternative carries its own risk, since if the system meant to catch an incident fails during that incident, nobody gets notified. Building one also pulls engineers into work that isn't a company's core business, at a scale where a single team can be responsible for thousands of services that it cannot watch by hand.
He said companies that adopt the platform typically see reliability improve within a short setup period, without having to build the system themselves. PagerDuty’s next phase is pushing agents further into diagnosis and mitigation, narrowing the gap between an alert firing and a problem actually being fixed. Outages at major cloud and AI providers are likely to keep testing that progress.