top of page
CASE STUDY

Google Cloud Platform

Simplifying Cloud Management with Active Assist

My small team built an AI tool that allowed Spotify to remove 9.5 million unused permissions on the first weekend and gave Google Cloud a new way to earn customer trust.

UX Team

1 UX Lead | 1 UX Designer (me) | 1 UX Researcher

Timeline

12 months

Collaborators

Partnered daily with PM and Engineers

Cloudy Sky

Shifting from Technical Accuracy to Compelling Invitations

When I joined the Active Assist team, Google Cloud’s AI was already capable of generating thousands of optimization suggestions—but in all of the alpha testing, users simply weren’t acting on them. The recommendations were technically accurate, but overwhelming, confusing, and risky to apply. These customers were ignoring the a powerful tool built to help them.

Our job was to turn that around.

By reframing the project around reducing user stress, uncertainty, and cognitive load, I helped the team pivot from “here’s what’s correct” to “here’s what’s helpful.” We bundled insights by goal, modeled outcomes visually, made actions reversible, and prioritized human-readable wins.

The results? Across a single weekend with select Beta users — some of Google Cloud’s most valuable enterprise customers — we saw:

Millions of unused permissions removed 

Massive cost savings applied with confidence

Rave feedback from customers like the New York Times

More than five years later, the core design structure I created is still in use — and continues to help Google Cloud grow its reputation for thoughtful, trustworthy tools in a fiercely competitive space.

Our Challenge

Reframing AI optimization into meaningful wins for real users — and helping engineers see why UX matters

By the time Google Cloud ramped up its platform efforts, competitors had already secured massive market share. Google’s technical infrastructure was impressive — but breaking through would require a reputation for clarity, confidence, and care.

The stakes were high. Cloud environments are complicated, error-prone, and constantly changing.

Industry-wide

Over

30%

of cloud users' time was spent on network troubleshooting

Over

90%

of cloud permissions were over-granted

Over

70%

of virtual machines were over-provisioned

Cloud environments were asking people to make decisions that computers would be better suited to calculate.

Google’s Active Assist team had begun addressing this problem using AI — surfacing thousands of recommendations about security, cost, and performance.

The algorithm was powerful. 

But the interface was a firehose.

When I joined, the product consisted of ungrouped, overwhelming lists of machine-generated suggestions. Internally, engineers saw the recommendations as a checklist of “correct” system changes. So why wouldn’t users just act on them?

But our initial user studies showed the reality: these recommendations increased stress, confusion, and anxiety about breaking things.

Reframing the Problem

Early on I proposed a shift in focus: this wasn’t a data visualization challenge — it was a trust-building challenge.

Our users weren’t systems. They were people.
Security teams

Worried that removing unused permissions would trigger a deluge of access requests the next day.

Scientists

Needed computing power but dreaded justifying a budget increase.

Sysadmins

Were terrified of clicking “apply” and breaking production systems.

They didn’t need a list of what was “right.” They needed reassurance, clarity, and outcomes they could be confident in.

Designing for Action, Not Just Accuracy

To shift the product focus from “accurate” to actionable, I introduced a series of design pivots:

A   B

Show, don’t tell

We added side-by-side visualizations that modeled the outcomes of implementing recommendations vs. doing nothing — showing predicted cost savings, security reductions, or performance gains.

Reversible changes

I advocated for “undo” features wherever technically feasible, so users could take action without fear.

Bundled recommendations 

We grouped like recommendations to reduce thousands of individual recommendations to hundreds of more broad recommendations.

Recommendations in context

I pushed to place recommendations where users need them—such as billing insights directly on the billing page—to reduce friction and streamline decision-making.

Card layout

I advocated for introducing a new card pattern to the platform's design system as an alternative to massive tables.

Bringing consumer grade design to a very utilitarian B2B space

One of my first proposals felt relatively innocent to me. I mocked up the bundles of the recommendations as cards. The problem was that the only precedent for cards in the Google Cloud Platform was in the marketplace where third party software was listed.

And those cards were incredibly basic.

Engineers on the team initially viewed UX as unnecessary overhead. One PM warned me that pushing a new card design would be blocked by Google Cloud’s design system. But I built prototypes, modeled user flows, and ran testing sessions showing a 28% increase in user action rates. That opened the door.

Scannability

Because we had a full-time UX designer, I was able to craft a series of studies where we were able to compare time-to-click, click-through rate, and misclick rate between tables and cards.

Then I got zero in on the ideal card layout for a scannable, actionable presentation of the recommendations. 

Adding Insights to Recommendations

I wanted our users to feel like they were in the driver's seat. This lead me to propose that we pivot from recommendations alone to insights coupled with recommendations.

"Remove 500 permissions inactive for over 6 months." 

This was how the the recommendations were initially presented. It was concise, but bossy.

"500 permissions have been inactive for over 6 months. Remove them?"

I rewrote copy to pivot to our interface making an observation (insight) about the permissions followed by a suggestion (recommendation) to remove them, and framed as a question. 

I wanted the pattern to feel like "I've noticed the following thing..." coupled with "Would it be helpful if we..."

434898_36fbd1c7edfe493f97bd865a06e617a9~mv2.gif
Abstract Design

Like a lot of designers, I usually work at pretty high fidelity. But when it came to reaching out to internal partners about our design patterns, I found that sample copy kept tripping them up, blocking out the forrest because a few trees didn't look right. 

Switching to an abstracted representation of the interface, similar to this one, allowed us to keep the conversation at a high level, considering overall patterns and placement of the new features.

The Power of Icons

Simply adding an icon to any recommendations that showed up on a table multiplied the rate of user interaction.

Cloud.jpg
Cloud-Icons.jpg

SIDE NOTE

Outside the Box Solution

As we came closer to launching, we needed about a dozen stakeholders to craft the final copy for the recommendations specific to their departments. Many stakeholders came back to us flustered about what copy we needed and where it would appear in the platform.

ewfweew_edited.png

I quickly used Google Sheets (Google's spreadsheet editor) to mock up a form where stakeholders could fill in all needed fields.

THE SOLUTION

Active Assist’s Recommendation System

The final product was a powerful yet intuitive recommendation system with three main components:

Small-Overview-Hub.webp

Overview Hub

The home page, which displays all available recommendations in a card-based format for easy scanning.

Small-Overview-List.webp

List View

Clicking a recommendation card displays a searchable and sortable table detailing all affected resources and the specific impact on each one.

Small-Overview-Detail.webp

Detail Panel

Each recommendation also has an accompanying panel where users can dive deeper, customize recommended actions, and view the predicted impact of their decisions.

HARD NUMBERS

Increasing Usage 10 fold

When I came on this project, only about 5% of alpha test users who saw recommendations worded in utilitarian ways on a massive table were comfortable enacting those recommendations. 

When we went to launch, we were seeing over half of our alpha test users comfortable acting on the now in-context, thoughtfully worded, and carefully laid out recommendations. 

THE LAUNCH

Trust Earned in a Weekend

The Beta launch weekend was carefully orchestrated with a small group of Google Cloud’s most important enterprise customers — including teams from Spotify, the New York Times, and others.

The stakes were high.

But the impact was immediate:
Security teams

Spotify alone removed 9.5 million unused permissions that weekned

Scientists

IT teams saved thousands of hours in troubleshooting and cost optimization

Sysadmins

Users praised the tool not only for its intelligence — but for its clarity

“I have been dreaming about this feature since 2 years ago.”

— Beta User

“This is by far one of the best things we’ve seen in Google Cloud.”

— New York Times' Cloud Team

“Networking security insights without us having to specify anything is magic. Real opportunity for Google to differentiate.”

— Beta User

Long-Term Impact

The core structure I designed is still in place, five years later. The same Recommendation tab. The same card layout. The same integration into users’ day-to-day cloud workflows.

Active Assist was a valuable contributor to Google Cloud's ambitious growth plan, with revenue more than doubling from ~$6 billion the year we began to over $13 billion in 2020, the year it launched. Active Assist remains a key part of that growth story, helping users take confident action in a world of overwhelming complexity.

In a Forrester Total Economic Impact study of Google Cloud customers, companies using Active Assist reduced internal GCP consumption by 40%, saving $2.5M over three years. Those numbers confirm the value of our UX work in enabling clearer insights, trustable recommendations, and sustainable cost management.

Forrester.png
Square,_Inc._logo.svg.png

"Active Assist’s Policy Troubleshooter is going to make my life so much easier. I can't begin to tell you how I've suffered from generic error messages in the past. Policy Troubleshooter is exactly what we need to quickly find, understand, and fix policy misconfigurations. " ​

Paul Friedman

Sr. Security Engineer

Square

Uber_logo_2018.svg.png

"[Active Assist's] IAM Recommender does all the heavy lifting. It was very easy to auto-apply all the recommendations we got." ​

Abhi Yadav

Sr. Cloud Security Engineer

Uber

Video Overview

I wrote, illustrated, animated, and narrated this video which was shared with the Cloud team at launch to bring everyone up to speed with the basics of how Active Assist's Recommendations functioned. It does a good job covering the basics.

PRAISE

“Tucker Is awesome to work with, both a top-notch designer and human.”

-Robert Lasker

UX LEAD ON THE PROJECT

bottom of page