How Incident.io Builds with AI (w/ CEO Stephen Whitworth)
Deep dive into the current state of AI at incident.io & what it was like building their first AI features.
Stephen Whitworth is the CEO of incident.io, an incident resolution platform powering thousands of engineering teams including Intercom, Netflix, and Etsy.
When something breaks in production, engineers turn to incident.io.
Thousands of teams use it to detect, coordinate, and communicate during outages.
Stephen believes the next era of incident response won’t just notify you when something’s wrong. It’ll help you fix it.
“We’re building agents to go and debug and diagnose and investigate problems... if you had to get a human to do that across thousands of companies and stacks, that’s not something any human has ever done before.”
In our conversation, we deep dive into the current state of AI at incident.io, what it was like building their first AI features, and how their multi-agent system will accelerate the product's value as usage grows.
Here are the highlights from our conversation.
From AI sprinkles to full-stack intelligence
Incident.io’s first steps into AI were modest “AI sprinkles,” as Stephen calls them.
They started with summarization tools: draft updates, channel recaps, the kind of features that save minutes, not hours.
“If you removed GPT from the app.. a lot of it would still work.”
Then came Scribe, an AI note-taker that joins Zoom calls, transcribes incidents, and flags key moments for stakeholders. Next, Chat, a conversational Slack agent that lets you literally @ incident and say:
“Can you write an update for exec stakeholders and pause it till Monday?”
Finally came Investigations - their biggest leap yet. It connects to monitoring, feature flagging, and logging tools to run real root-cause analysis.
All these systems now talk to each other. Scribe provides context for future incidents, and Investigations learns from past ones.
It’s evolved into a cohesive system, not a collection of disparate features.
The gap between demos and reality
Shipping AI isn’t the same as demoing AI.
Stephen said that while it’s easy to put out a demo that “goes viral on Twitter,” making that same system work across thousands of customers is a different challenge entirely.
“You can make something look great for an individual customer, but that doesn’t solve the generalizability gap.”
The hardest part is precision. A false positive in incident management is not only unhelpful, it almost always makes things worse.
“If you get it wrong, you’re making things worse. We need to be a really high-precision but low-recall system to start with. Because if we were 10% right, it would just be actively a lot worse.”
Stephen recommends leaders to favor accuracy over flash, especially when serving enterprise customers. It's what separates operational AI from experimental AI.
Difficulty as a moat
Most founders talk about product-market fit. Stephen talks about problem difficulty and the suffering involved in solving it.
“Difficulty and suffering is a moat. If it’s fundamentally hard, that’s great. I bet on our team - if we can crack it, others will give up along the way.”
Incident management is hard precisely because it touches everything: technical infrastructure, comms, and the most complicated of all - human psychology. Automating that loop goes way beyond logs and alerts. It’s about context, trust, and timing.
The company’s bet is that by owning that complexity early, they’ll be best positioned to scale the next generation of AI-powered operations.
Where the best ideas come from
Many of incident.io’s breakthroughs start internally, because they dogfood their product.
“We’re very heavy users of our own product, and I wouldn’t have it any other way"
That proximity allows for faster loops and more intuition-driven bets.
Customer feedback still matters, but the conviction to build what feels right, even before it’s requested, is a big part of how incident.io stays ahead.
“Everyone’s wanted something that helps them fix things faster for fifty years... the challenge isn’t the idea, it’s getting it to work reliably at scale.”
His thoughts on the future
When we asked Stephen about where he sees the space going, he outlined three eras:
- The first era was where tools like PagerDuty acted as "wake you up" services when something broke.
- The second era is where incident.io's product is today - a tool that helps you communicate and coordinate once something breaks.
- The third era will include tools that help you diagnose and repair the incident itself.
Stephen is very high conviction about incident.io's ability to usher in that third era.
“We’ve built this version of the company to be able to build the next version of the company... the next one helps you fix things alongside you.”
Closing thoughts
One thing that stood out from our conversation with Stephen is his take on defensibility in the AI era. The teams that are poised to build defensible products are the ones willing to tackle the hardest problems (and suffer through them).
He captured this perfectly when we said:
"Difficulty and suffering is a moat"
We had a really great time jamming with Stephen! We’ve been having similar conversations with product and engineering leaders at top engineering orgs (like Vercel, Wix, Vanta, etc) to unpack how they actually build with AI. You can find previous episodes on YouTube.