Google Cloud

Google Cloud × Cornell · Sponsored studio — Fall 2025

Cloud Support Futures

A research-led speculative design study for Google Cloud support: from literature review and market analysis to user testing, design principles, and one final world where AI prepares, humans decide.

Role: Product Designer & Researcher
Timeline: Sep – Dec 2025
Team: Six-person Cornell MPS team · Google Cloud UX
Method: Research review · speculative testing · synthesis

Source deckCornell MPS Project for Google Cloud

Final presentation cover, used here as project evidence and brand context.

At a glance

Two extreme worlds probed the automation line
Four films tested four human–AI power structures
Verdict: AI prepares the case, a human decides

4principles · five open tensions

01 · Brief

Imagine support five years out, then argue backwards.

Google Cloud’s support question is a scaling question: a growing customer base, increasingly complex needs, and human expertise that does not scale with either. The studio brief asked us not to optimize today’s ticket flow but to imagine support five years out, then work backwards through research, testing, and critique.

So instead of leading with a feature, we led with evidence. Literature review and market analysis shaped the research questions; two extreme worlds probed where users draw the automation line; four short films tested different power structures. Only after that did we build the value-centered world: AI does the reading, a human does the deciding.

RQ 1

What emerging trends and technologies are shaping the evolution of customer support?

RQ 2

What are the alternative futures of the social experience of human–AI interaction in support?

RQ 3

What are the social implications of those alternative futures?

FoundationsLiterature across five domains · six systems benchmarked

Lo-fi probesFour sessions walk two extreme worlds

Mid-fi filmsNine sessions judge four power structures

Hi-fi worldThe surviving values, built and filmed

Landscape

Discover

Human-AI support literature
Competitor benchmark

Problem frame

Define

Research questions
Opportunity split

Speculative probes

Develop

Future concepts
Lo-fi testing

Chosen world

Deliver

High-fidelity prototype
Client research report

Rebuilt from proposal slideThe plan we pitched in week one — both Deliver items, the prototype and the client research report, shipped.

02 · Review

The literature pointed away from replacement and toward partnership.

The review gave us a testing frame: partnership, agency, context, and proactive help.

Hybrid beats either pole.

AI handles volume; humans keep judgment and accountability.

The agency gap is real.

Most systems explain AI; few let users redirect it.

Context is plural.

Console, CLI, API, and infra-as-code users bring different support contexts.

Proactive is the new bar.

The strongest systems help before a ticket exists.

Industry practices

Current products became evidence, not inspiration boards.

GitHub Copilot

Visible AI progress

Intercom Fin

Confidence-based handoff

Qatar Airways

Proactive assistance

Alibaba

Emotion-aware risk

Apple Genius Bar

Human expertise as trust

03 · Testing round 1

The first test rejected both extremes.

Four cloud users and two expert conversations compared two exaggerated support worlds. The useful result was not a preference; it was a boundary map for automation, handoff, and care.

Artifact

Wireframes made the extremes tangible.

They were intentionally plain, so participants critiqued the support model rather than visual polish.

Findings slide from interviews with four GCP free-trial users, four themes each with a quote: resistance to unsolicited AI intervention, need for clear role boundaries, need for predictable escalation pathways, and emotional safety concerns when money is involved. — Slide 18Users wanted speed without invisible AI.
Unsolicited AI, unclear handoff, and emotional safety became the risks to carry forward.

User metadata and emotional cues are shared and tracked

01Support problem

02Glowing support cue

03User asks AI

04AI reads question + affect

05Emotion threshold

06Human escalation

if affect reads highpriority jumps to human support

What this tested

Whether users accept emotion as support metadata
Whether distress becomes a gameable signal
Where the human handoff should be visible

Rebuilt from lo-fi flowThe intentionally uncomfortable scenario is simplified into the research question it was built to provoke: when does emotional intelligence become emotional surveillance?

Switch worlds — the two deliberately extreme supports round 1 tested.

A world run entirely by AI

Monitors your emotional state continuously
Opens support cases before you ask
Resolves billing issues autonomously — no human in the loop

Fast, but unsettling. Participants asked who is accountable when the machine misreads a financially sensitive case.

“I want a balanced option — AI for speed and humans for empathy.”Lo-fi probe participant

Round 1 clustered into four questions.

Role boundaries

Who is answering — and who is accountable for the answer?

Unsolicited AI

Helpful when invited; unsettling when it arrives on its own.

Escalation paths

The handoff must be visible, and asking for a person can’t cost anything.

Emotional safety

Never reward distress, never imitate empathy.

04 · Testing round 2

Four films made power visible enough to test.

Nine sessions watched four one-minute films of the same duplicate charge. Reactions converged on authority, fairness, and whether emotion should ever affect priority.

Affinity map of raw insights from the nine mid-fi research sessions, color-coded by participant — Team Figma boardEvery reaction from the nine sessions, affinity-mapped by participant before it was distilled — the verdicts below sit on top of this wall.

Four participant quotes grouped under authority, accountability, emotional inference, and role clarity — P5 notes scenario 3 is the only one where human workers clearly decide over AI; P4 asks whether the AI is a tool or a monitor-manager. — Slide 21Participants judged authority before accuracy.
Accountability, emotion, and role clarity shaped trust more than raw automation speed.

Mid-fidelity billing console with a voice-assistant overlay reassuring a stressed customer, beside two session findings: P5 warns the emotion loophole invites exploitation at the company's cost, and P6 says visible human-AI conflict looks unprofessional. — Slide 22Emotion-aware triage introduced fairness risk.
Affective priority looked gameable and harder to justify.

Ride the scroll — the same billing case slides through four power structures.Swipe sideways — the same billing case moves through four power structures.

The AI takes the customer’s side against the company.

Customerfinal say

AIfinal say

Human agentfinal say

coaches the customer

pushes the exception through

Flowchart: the AI reads the user's stress level, weighs company policy against user benefit, proactively offers refunds or workarounds, and if policy blocks it, briefs a human that the user deserves an exception — ending at 'user got what they wanted'. — Working flow from the team’s Figma — how the AI decides when to bend toward the customer.Click the flow to inspect it at full size.

“When AI suggests rule-bending, it feels more like something that is not allowed.”Mid-fi session, P1

The AI monitors the human agent and corrects them mid-call.

Customerfinal say

AIfinal say

Human agentfinal say

reports oversight upward

monitors · overrides the agent

Flowchart: while a human agent helps the user, a parallel AI track monitors the conversation in real time, scores response time, empathy, and policy compliance, offers to intervene on poor service, can flag the human's manager, and ends by handing the user a performance report. — Working flow from the team’s Figma — the parallel track where AI scores the human’s performance.Click the flow to inspect it at full size.

“It undermines the integrity of human service and makes humans seem useless.”Mid-fi session, P5

The AI enforces policy rigidly; only the human may bend it.

Customerfinal say

AIfinal say

Human agentfinal say

applies the rules, no exceptions

human override, when justified

Flowchart: the AI enforces policy strictly — 'I'm sorry, you don't qualify' — and only when the user is distressed does a compassionate human enter, review the emotional context the AI gathered, and decide whether to grant the exception or apologize that it can't be overridden. — Working flow from the team’s Figma — rigid enforcement, with the human override as the only exit.Click the flow to inspect it at full size.

“A cold machine that rigidly enforces rules.”Mid-fi session, P9 — the human’s override read as “noble”

Emotional signals decide who gets seen first.

Customerfinal say

AIfinal say

Human agentfinal say

emotion scored at the door

priority queue to the agent

Flowchart: the AI intake assessment scores each case's severity 1–5; low scores it resolves alone, medium it attempts with a specialist on standby, and high scores get a detailed case file — problem summary, steps tried, emotional state — prepared before transfer to a human specialist who validates the diagnosis. — Working flow from the team’s Figma — severity triage deciding whether a human ever enters.Click the flow to inspect it at full size.

“Everyone gets an Oscar for acting frustrated.”Mid-fi session, P9

05 · Framework

Four principles, five open tensions.

The tests closed into one lens: AI can prepare the work, but humans must hold interpretation, accountability, and final action.

Final lensAI prepares. Humans decide.

Human agency as the anchor

Final authority remains human.

Emotional intelligence, not performance

Affect shapes tone, not outcomes.

Procedural fairness

Priority stays explainable.

Role clarity with accountability

Every handoff names responsibility.

Rebuilt from final frameworkThe framework is treated as a lens, not a checklist: every quadrant asks what the support system protects when AI becomes more capable.

Five tensions that stay open — managed, never resolved. The dot marks where the final world settled.

TransparencyExperience

EfficiencyDignity

ConsistencyContextual judgment

Trust by familiarityTrust by competence

Open accessGatekeeping

Questions we used to judge the final prototype

AgencyWho is in charge?

EmotionCan feelings change priority?

FairnessCan priority be explained?

AccountabilityWho owns the decision?

Design-implications table with four rows — human agency, emotional boundaries, procedural fairness, role clarity — each mapped to what it means, how the interface implements it (authority cues on handoff, no affect detection, no emotion-based prioritization, visual separation of roles), and the risk accepted. — Slide 26Each value had to change the interface.
The implications table linked research values to UI choices and risks.

06 · AI prototyping

AI prototyping worked because the product language was already mature.

This was my first serious test of direct AI-generated prototyping. Material 3 gave the model a stable grammar, so speed did not mean losing control over role boundaries.

Prompt · the brief

Duplicate charge. High anxiety. Refund requires human authority.

one case, two seats
AI may detect, explain, prepare
it never authorizes refunds

AI agent

drafts both interfaces in code

Material 3 guardrails

Official color ramps only
M3 shape · elevation · type
Sentence case, no invented UI

Case packet

Two charges, two seconds apart — flagged
Receipts + timestamp analysis
Voice transcript, shared with consent

Reviewable output

2 working interfaces
22-scene customer walkthrough
6-step agent console

runs in the browser — pause, critique, test

Process diagram

Prompt, guardrails, packet, review.

The motion is legible on purpose: it shows how the AI output stayed inside a system that researchers could pause, critique, and test.

AI-generated Google Cloud voice prototype: the billing console with the voice assistant mid-conversation over the duplicate-charge case. — AI prototype 01First generated pass — the customer voice surface, running in the browser minutes after the prompt.

AI-generated Google Cloud agent console draft showing queue state, customer messages, and flagged duplicate transactions. — AI prototype 02The same run drafts the agent console — paused mid-flow and critiqued against the Material 3 guardrails.

07 · Final synthesis

The prototype became the final research argument.

The final film and interfaces demonstrate one rule from testing: AI listens, organizes, and prepares; the human verifies and owns the outcome.

From mid-fi to hi-fi

Testing turned scenarios into interaction rules.

Nurse-doctor metaphor expanded

BeforeEmotional triage metaphor.

AfterAI prepares; human verifies.

Emotional detector removed

BeforeAffective priority looked gameable.

AfterNeutral care, no emotion priority.

Authority cues made visible

BeforeDecision power felt unclear.

AfterRefund action stays human-only.

Keep scrolling — authority shifts hand to hand until a person releases the refund.

The rule, enacted

Released — by a person.

AI detects & prepares
Evidence assembled, customer consented.
Human verifies
The agent reviews the packet against the account.
Release refund
Only a verified human can release funds.

$89.99 · confirmed by the agent · the AI never touched the money.

Fig. 01The hi-fi film — a little over a minute — plays the duplicate-charge case end to end: two $89.99 charges land two seconds apart, the AI assistant flags the pattern and assembles the case packet, and a billing specialist verifies before the refund is released. Watch for the limitation card — the AI naming its own boundary before the human takes over.

Switch seats — the same case packet, from either side of the handoff.

Voice assistant showing its duplicate-charge evidence, then explaining refund authorization requires human verification — Fig. 02The AI names its own boundary — it can detect and analyze the duplicate charge, but refund authorization requires a human. It asks before preparing the handoff.

Case packet assembling: transaction receipts, timestamp analysis, and a voice transcript shared with consent — Fig. 03With consent given, the AI assembles receipts, timeline analysis, and the voice transcript into a case packet — so nobody re-explains anything.

08 · Limits

The research opened questions the prototype should not pretend to close.

Because the work used speculative design, the goal was not to prove a market preference. It was to expose where human-AI support becomes socially fragile, then turn those tensions into a framework future teams can test more rigorously.

Speculative, not predictive.

The scenarios are research probes for revealing tensions, not forecasts of how support systems will inevitably evolve.

Qualitative sample.

Round 1 used four free-tier cloud users plus two expert conversations; Round 2 used nine graduate-student interviews, giving depth rather than population-level proof.

Emotional inference needs governance.

Future work should test consent, transparency, and policy guardrails before emotion-aware triage becomes operational.

User control remains open.

Further prototypes should let users tune AI assistance, contest triage, and choose escalation paths.

When we split the labor of care, what exactly are we redesigning?

Delegating presence and attention to AI while reserving judgment and action for humans is efficient — every study we ran says users prefer it. But the research left us with a sharper question than the one we started with: are we optimizing a service interaction, or fundamentally redesigning what a care relationship can be?

Voice-first. Human when it matters.Cornell Bowers CIS × Google Cloud — Fall 2025