Gemini Cloud Assist Investigations:
Solving the “Black Box” Problem in Enterprise AI
- 0→1Pitch to platform pillar
- Next ’24Keynote demo
- AwardPSX UX Trailblazer
- Aug ’25Public preview
Disclaimer
The following case study is personal and does not necessarily represent Google’s positions, strategies or opinions. I have omitted and obfuscated confidential information.

Summary
Overview
Gemini Cloud Assist Investigations started as a pitch deck and grew into a core part of Google Cloud’s AI strategy. It helps engineers figure out what broke instead of staring at dashboards.
The problem
When something breaks, engineers race the clock to fix it. Our tools were great at collecting data and almost no help making sense of it, so engineers had to piece together scattered signals by hand.
What I did
With our product lead, I spotted the opportunity and pitched it through an internal roadshow. I designed the “Blast Radius” MVP for the Next ’24 keynote, then the “Glass Box” model that makes the AI’s reasoning checkable.
Impact
A grassroots pitch became a strategic pillar for Google Cloud, earned the PSX UX Trailblazer award, and shifted how our team designs: from fixed screens to AI that reasons alongside people.
Who I brought togetherA Product Lead, Senior Directors, and the product and engineering teams who built it. A North Star deck turned an idea that wasn’t on any roadmap into a funded product.

The Origin
Friction
In 2022, the engineers who keep Google Cloud running were buried in data. Logs, dashboards and Slack threads were all there, but connecting them during an outage was on them.
The pitch
When generative AI arrived, our product lead and I saw that its real value for troubleshooting was connecting the dots. It wasn’t on any roadmap, so I made a North Star vision deck and we took it to senior directors. We came back with headcount, funding and the flagship demo for the Google Cloud Next ’24 keynote.
The MVP
See it before you fix it
Before AI can help fix a problem, it has to show it can see it. We kept the MVP focused on reliable, high-value signals so engineers’ first experience would be fast and trustworthy.
A map of what’s affected
A map of the failure’s “Blast Radius”, showing every service upstream and downstream of an alert, instead of a list.
Warning signs in one place
Red flags from across the stack, together in one view instead of a dozen dashboards.
Log themes
AI groups similar logs and explains cryptic errors in plain English.
Outcome
This scope became the Next ’24 keynote demo, showed there was real demand for AI in operations, and changed how we measured success:



From Answers to Evidence
A pivot
After Next ’24, we realized the MVP was built for the happy path. To really cut time to resolution, the AI needed to suggest what might be wrong, not just summarize. That raised a hard question: why would an expert engineer trust a machine’s best guess?
Designing for trust
Instead of one “magic answer,” I designed a Glass Box model where the AI acts as a collaborator, not an oracle.
A · Probable root cause
The AI offers ranked theories, like a database running out of connections, so the expert can weigh them and decide.
B · Show the evidence
We called it “Observations”: every insight links back to the exact log lines, metric spikes or config changes behind it.
Engineers don’t have to trust the AI blindly. They can check its work.
Try it yourself
You’re on call. Fix it.
It’s 14:02. Checkout is slow and customers can’t place orders. See how the same incident plays out when the AI hands you an answer, and when it shows its evidence.
Interactive concept demo. The incident, services and data are fictional, and this is not Google product UI.
The key decision
One confident answer, or a ranked set of possibilities?
The options
Most AI assistants give one answer and move on. We could have done the same, or shown several ranked theories with the evidence behind each.
What we chose and why
Ranked hypotheses, each anchored to Observations. Expert SREs are paid to be skeptical. An answer they can’t check is an answer they won’t act on during an outage.
The trade-off
More to read than a one-line answer. Ranking and filtering had to do real work, so the most likely cause and the most relevant evidence always come first.

The Vision
Getting ready for agents
Engineers will soon spend less time doing the work and more time supervising AI agents that do it.
I’m now leading design standards for “Human-on-the-Loop” interactions: how people can safely hand work to AI in systems where mistakes aren’t an option.
Impact
What changed
Investigations launched in public preview in August 2025. Outcome metrics are confidential, and I’m happy to walk through them in conversation.
Strategy
Found the opportunity with our product lead, wrote the North Star and ran the roadshow.
Design
Designed the Blast Radius MVP and the Glass Box model.
Leadership
Won senior director support, headcount and funding, then guided the team through several big pivots.
Standards
Now defining Human-on-the-Loop patterns for agents.
In the Wild
Learn more about Gemini Cloud Assist ↗ · Public coverage of this work.
What people are saying…
Our success at Google Cloud Next can be directly attributed to Ashley’s thoughtful and innovative designs that incorporated & balanced UXR feedback, product and engineering requirements effectively. Thank you Ashley for all your hardwork, leading the team through several key pivots and always maintaining a can-do positive attitude!
Congratulations on the successful launch of Gemini Cloud Assist at Next ’24! It would not have happened without all your hard work and focus!