Support AI for Xbox

Designing a conversational support experience for Xbox, integrating a Support Virtual Agent (SVA) across chat, search, and entry points.

UX designer Jan 2024 – Jan 2025 Shipped

Support Virtual Agent

About this case study

The project

The process

  1. Hi. I can answer questions about this case study, and only this one. Pick a topic, or ask in your own words.

A demo agent built for this case study, not the Xbox product.

Overview Rebuilding a scripted menu as an AI agent

I took a menu buried in the footer and made it the first place most people go for help.

UX designer

I owned conversational flows, escalation logic and accessibility.

12 months

Discovery through to ship

9 people

2 designers, 1 researcher, 4 engineers and 2 PMs

Figma

Tools

Problem Fragmented, reactive and expensive

Help meant a menu-driven agent, hand-offs that lost the conversation, and a heavy reliance on people.

The legacy Xbox Contact us page. The virtual agent sits in a card below the fold, behind an Ask a question button.
Before: the agent lived in a card on Contact us.

What was broken

  • Menu-driven virtual agent

    Outdated static scripts that couldn’t adapt to user intent.

  • Inconsistent hand-offs

    Created friction and dropped context at every transfer.

  • Low satisfaction, high cost

    Unsustainable over-reliance on human support.

What we set out to do

  • Reduce support volume

    Intelligent and adaptive self-service.

  • Make the SVA the front door for help

    Resolve issues directly, or hand off cleanly when needed.

  • Cut maintenance cost

    A system that learns over time, instead of constant manual updates.

Solution Answers, not menus

Deliver adaptive answers where gamers were already looking.

Building the core experience.

  • The SVA answering how to reset an Xbox One console in four steps, with a Learn more link and thumbs up and down

    AI-powered chatbot

    Generative AI replaced the static scripts, so the SVA responds to what someone actually asked instead of routing them through a menu.

  • The SVA explains how to investigate a charge you don’t recognize, offers three support options (Do it yourself, Chat with us, and Request a call with a five-minute wait), then links three articles

    Escalation hand-offs

    When it can’t help, the user is handed off to a human agent. Too early and the SVA is just a doorway. Too late and people loop. I designed the escalation logic around that.

Impact SVA as the front door for help

Ninety days after launch, the numbers held.

Resolved means the chat ended without a hand-off or a thumbs down. The two systems measured success differently, so these numbers show direction, not proof. But the direction was unambiguous.

38% 12%

Escalated to a human agentDown 26 points

Legacy agent sessions, then SVA conversations.

30% 69%

Resolved without reaching a human agentUp about 39 points

The legacy figure is roughly 30% of sessions.

Discovery Framing the problem right

With generative AI, people expect a chat to understand them. However, most support bots in the market were still running scripts.

Heuristic review

Of the heuristic reviews I ran, Amazon and Zoom each showed something different worth learning from. Amazon used context, like your orders and your last chat. Zoom was easy to find and clear about what it covered. Both relied on preset options, and neither could adapt to what someone actually asked.

Audit of Amazon's customer service chat, four screens and seven findings: it opens with one question, offers common issues and your real orders, and resumes chats, but its empathy reads as scripted
Audit of Zoom's support chatbot, four screens and eight findings: easy to find and clear about itself, but it answered off topic, pushed marketing options and locked live chat
The heuristic audits. Select one to open it full size.

Then I turned the audit on our own agent

The same heuristic review, applied to the legacy agent we owned.

Two things were worth keeping: the legacy agent set expectations about its capabilities before the user invested in the conversation, and its clarifying questions offered actionable options rather than open-ended prompts.

Audit of the legacy Xbox virtual agent, four screens and seven findings: easy to spot, it set expectations and offered clarifying options, but its icon misled, it ignored that you came from Xbox Support and it never acknowledged feedback
The same audit, on the legacy Xbox virtual agent.

What I had to solve

  • Buried entry point

    Reaching the legacy agent meant scrolling to the footer or finding an article first.

  • Hand-offs lost context

    A human agent picked up with none of the conversation, so users had to repeat themselves, which caused frustration.

  • Feedback went nowhere

    Ratings were collected and never acknowledged, so there was no reason to leave a second one.

Beyond support The bar had already moved

None of these were support tools, but I studied them because they were reshaping what people expected from an AI conversation.

  1. Trust and control

    Microsoft Copilot Microsoft

    Cited responses let users verify AI reasoning. Trust comes from showing the work, not just the conclusion. A design principle I carried into Xbox.

  2. Exploration without lock-in

    Google Gemini Google

    Draft-based thinking let users explore options without feeling committed. It informed how we framed AI suggestions in support flows.

  3. Simplicity at depth

    ChatGPT OpenAI

    A single input field carrying deep capability, with no menu of options in front of it. It set the baseline for what our chat entry should look like.

Research Users preferred personality.

With a UX researcher, I concept tested chat with and without embodiment. 20 participants, split evenly. It’s why we landed on a restrained presence.

4.1vs3.2

Average rating out of 5, embodied vs not. Directional, small sample.

  • I do like that it is giving a sentiment… it is an affirmation that it is following along with my thought process.

    Participant P06

  • I’m a fan. He seemed chipper. Good voice. Pleasant.

    Participant P17

8secvs3sec

Roughly 8 seconds a reply, against the 3 seconds P16 expected.

  • I expect chatbots to be quick and helpful: responses shouldn’t take more than 3 seconds.

    Participant P16

  • Any interaction should feel simple, magical, and embody the gaming experience.

    Participant P01

Define How present should a support agent be?

An embodied agent could extend beyond support to Minecraft and Halo, a real reason to explore character. But people reach support already frustrated. I framed it as a spectrum we could test, so the team could decide with evidence.

The opportunity

The case for character

Gaming runs on characters. One embodied agent could carry from support into Minecraft and Halo.

The context

The case for restraint

A bot saying “I totally understand” reads as performative to someone already frustrated.

The spectrum

Each layer runs from none to full. The dot is where we set it.

Visual

Static, abstract avatarFull animated character

Aural

UI sounds onlyVoice, gender, accent, tone

Textual

Neutral, functional languageCharacter voice, humor, informality

Where we landed: a restrained presence, with character as a secondary experience.

The presence infographic: visual, aural and textual layers mapped from a static avatar through presence and embodiment to singularity
The infographic I made to give the team shared terms: visual, aural and textual presence, from a static avatar to full embodiment. Select to open full size.

Approachable and reliable, calibrated to what the system could deliver.

Define Escalation path

Every path ends at an answer or the support options, and none reaches support without a topic.

The escalation flowchart from the dev spec. Three ways in lead to high or low confidence. Confident questions get an answer with links; unsure ones get topic presets, then a question in the person’s own words, one more round of presets, and finally category, topic and platform dropdowns. Every branch ends at an answer or the support options.
Open the full chart

The whole flow

Every branch the agent can take, from the greeting to the support options. Each stop zooms into one part of it.

Three ways in

A question, a question with a request for a live agent, or a request for a live agent alone. The greeting already carries browse data from SXC, Xbox Support, so the agent knows what the person was looking at. The agent sorts each first message by how confident it is.

When it’s confident

The agent answers with links to the source. Asking for a human agent after that brings up the support options; a question that already asked for one gets the answer and the options together.

When it isn’t

Instead of guessing, the agent offers three to five topics, picked from what the person browsed and said so far, plus “Something else”.

In their own words

“Something else”, or “live agent” again, gets a plain question with no presets. If the agent is still unsure, it offers one more round of topics. That loop runs once.

The full list

If the agent still can’t tell, it shows every category, then topic, then platform, and from there the support options.

Design Four ways in

The agent came out of the Contact us page and onto the pages people already visit, so help starts wherever they are.

Implement Iterating toward the shipped experience

The legacy agent opened in a pop-up and asked people to pick from a list. The shipped SVA answers in a full chat, cites its source and asks for feedback on every answer.

I wrote the spec engineering built from: the escalation flowchart, each component’s anatomy, six scenarios and layouts from 360 to 1920 pixels.

The legacy Contact us page with the virtual agent pop-up open, answering a subscription question with a list of options
The shipped Support Virtual Agent: an abstract animated orb on the left and the chat on the right, answering how to reset a console

After: a full chat beside an abstract animated presence, with a source and feedback on every answer.