GoodData Cloud's support hub answers questions in plain language, with sources, so people resolve them in seconds instead of filing a ticket. I led the UX, designing for trust, scannable answers, and a graceful handoff to humans.
Overview
Customers were waiting hours for answers that lived in our docs. We shipped an assistant that resolved tier-one questions and escalated the rest without friction.
It lives in the support hub as one widget with two ways in: search, for people who want to find a doc, and Ask AI, for people who just want the answer.
Seconds
Time to answer, down from a ticket cycle
Two-in-one
Search and Ask AI in one place
Cited
Every answer linked to its sources
The problem
Documentation-heavy products carry a quiet tax: the answer exists, but people can't find it, so they file a ticket anyway.
A list of links doesn't help, it just hands the reading and the deciding back to the user. And that friction is exactly where people give up.
"The job wasn't to add a chatbot. It was to deflect answerable tickets without eroding trust."
What I owned
I led the UX design and rollout end to end, designing with support, not just for it.
Co-design sessions with support engineers grounded the work. We treated their handoff notes as design material.
Interaction & flow
The two-way entry, the handoff between search and Ask AI, and the answer experience.
Research
Intent research on when people want to search versus ask, plus competitive analysis.
Trust & accessibility
Citations, escalation cues, keyboard nav, focus states, and contrast as a baseline.
Scope & rollout
Deciding what the assistant should answer, how to message what it couldn't, and shipping it.
Two intents, one entry
Users arrived with two goals: some wanted the answer, others wanted to find and read the doc. Plain search makes the answer-seekers do the synthesizing; a plain chatbot frustrates the people who just want the source.
So we built one entry that holds both, and let intent pick the tool instead of making people hunt for the right feature.
The search-to-ask handoff
People reach for search first because it's familiar, but their query is often really a question an answer would serve better. Without a bridge, they bounce between modes or give up.
So a quiet prompt sits atop the results, offering to turn the query they just typed into an instant AI answer, with nothing to retype.
Tradeoff
To avoid pulling people away from the source docs, we kept the full search results visible underneath rather than replacing them.
Answers built for trust
Source-grounded answers with inline citations, and a visible confidence signal that doubles as the escalate-to-human trigger. The answer arrives framed as evidence, not opinion.
Default answers were too long to scan, so we tuned them toward tables and short, copyable snippets, and moved the experience from a popup into a calmer drawer, far easier to read, especially on mobile.
"An answer you can't trace is an answer you won't act on."
Knowing what to answer
An assistant is only as good as what it's allowed to read, and answering everything would make every answer worse. So we scoped it to where it could be confident: Cloud and Cloud Native docs, plus the developer references and SDKs.
We deliberately left out the legacy product, whose older structure would have muddied answers, and pointed those users somewhere better instead.
Tradeoff
A smaller, accurate scope beats a broad, unreliable one. Some users fall outside it, accepted on purpose to protect trust, and handled with clear messaging about what the assistant covers.
Outcome
Deflection landed high within weeks. CSAT for assistant-handled tickets exceeded the human baseline.
Reflection
We invested as much in the handoff as in the answer. That is where the trust lived.
