Ask Alteryx: one AI assistant for a whole analytics platform, piloted in Designer Desktop and built to work across all of Alteryx One.
Alteryx One is an analytics platform made up of several products: Designer for building data workflows on the desktop, Designer Cloud in the browser, and Auto Insights for exploring the results. Each had grown up on its own, but AI needed to show up the same way in all of them. The answer was Ask Alteryx, a single conversational assistant designed as a chat panel that sits beside the work rather than replacing it. People ask a question or describe a goal in plain language, and the assistant suggests the next tool, places and configures it on the canvas, summarises complex workflows, and helps troubleshoot when something breaks. Its decisions came from what customers were actually asking for, and its design system wasn't tied to any one product. Piloted in Designer Desktop, foundational to the rest of Alteryx One.
Decisions started from evidence, not guesses. Beta data and metrics showed how people already used AI in the product, and customer feature requests showed what they wished it did. Interviews filled in the why, and the requests became the roadmap.
The conversation was designed before the interface, with the prompts written alongside engineering. Every answer was shaped into a small, useful moment, a suggestion, an explanation, a tool placed on the canvas, so a response always led somewhere.
Early token usage was high. Teaching the assistant Alteryx’s own tools meant it needed less context, and only workflow structure and metadata reached the model. Alteryx’s benchmarks back the approach: up to 93% fewer tokens and 85% faster answers.
Built as a micro-frontend: one self-contained panel any Alteryx One app can host, with a component library of its own. Shell, content and behaviours were all documented in light and dark mode, with accessibility in from the start.
Components already had hover, active and disabled states, but nothing said when AI was available, working, had contributed or could explain itself. AI states layered on top gave Alteryx One one consistent signal for each.
Launched first in Designer Desktop, with trust built in: sources on every answer, canvas changes that can be restored, and clear data controls. When Designer moved under Alteryx One, the same panel shipped across the platform.
Decisions started from evidence, not guesses. Beta data and metrics showed how people already used AI in the product, and customer feature requests showed what they wished it did. Interviews filled in the why, and the requests became the roadmap.
The conversation was designed before the interface, with the prompts written alongside engineering. Every answer was shaped into a small, useful moment, a suggestion, an explanation, a tool placed on the canvas, so a response always led somewhere.
Early token usage was high. Teaching the assistant Alteryx’s own tools meant it needed less context, and only workflow structure and metadata reached the model. Alteryx’s benchmarks back the approach: up to 93% fewer tokens and 85% faster answers.
Built as a micro-frontend: one self-contained panel any Alteryx One app can host, with a component library of its own. Shell, content and behaviours were all documented in light and dark mode, with accessibility in from the start.
Components already had hover, active and disabled states, but nothing said when AI was available, working, had contributed or could explain itself. AI states layered on top gave Alteryx One one consistent signal for each.
Launched first in Designer Desktop, with trust built in: sources on every answer, canvas changes that can be restored, and clear data controls. When Designer moved under Alteryx One, the same panel shipped across the platform.
Ask Alteryx: one conversational assistant for a whole analytics platform, shaped by what customers asked for, kept lean on tokens, and built on its own design system. Piloted in Designer Desktop, foundational to Alteryx One.
Making AI feel native in several products that each look and work differently.
Beta data, metrics and customer feature requests decided what the assistant did first.
Prompts designed hand in hand with engineering, with token usage brought down so it stays affordable at scale.
A micro-frontend chat panel with its own component library, themed for light and dark, built to sit inside any Alteryx One app.
Designing for AI with mostly static tools meant guessing how real answers would feel. Today, the model can be part of the design process itself.
thenStatic screens of ideal, hand-written responses.
nowLive prototypes calling a real model, so the design is tested against what the assistant actually says.
thenReviewing sample prompts and replies by hand.
nowSets of real customer prompts run against each change, to catch bad answers and token spikes before release.
thenComponents rebuilt and checked by hand in every product.
nowThe system as structured tokens and specs an agent can read and apply in each product, with design QA done in parallel.
