Reimagining AI Assistance
Six AI features, built independently over three years. No shared strategy, no coherent system. This is how we changed that.
Role
Lead Designer
Company
Oracle
Year
2025
The redesigned system — AI detects intent from the query and routes to the right surface. The canvas stays uncluttered until assistance is needed.
The question I couldn't stop asking
Why do six AI features add up to zero strategy?
01 — The Problem
Six AI features. Zero coherent strategy. Analysts couldn't find what they needed — or know it existed.
Challenges
- ↳
6 disconnected AI features built independently — not designed to work together.
- ↳
Pro-grade complexity in a BI tool already constrained by screen real estate.
- ↳
Assistance felt bolted-on: not discoverable, not core to the experience.
Goals
- ◎
Consolidate 6 patterns into a cohesive, harmonious system.
- ◎
Orchestrate AI around user tasks and intent — not a single generic solution.
- ◎
Preserve and improve all existing functionality. No feature loss.
03 — The Hypothesis
The question
One panel that does everything, or intent-based routing to the right pattern?
The reasoning
The research made it clear: a generic assistant creates cognitive overhead because users have to translate their specific need into a chatbot format. Routing by intent — Answer, Explore, Create — means the right interaction surface appears before the user has to ask for it. The AI meets the analyst in their workflow, not the other way around.
The 3 intent modes
Answer
Specific question → direct, contained response
Explore
Open-ended → dedicated space for deeper analysis
Create & Edit
Manipulation → inline, contextual assistance
04 — The Work
Designing the orchestration system
01
Discover
Pulled 40+ natural language queries from design reviews, observed sessions, and engineering discussions. What struck me immediately: the same user need would arrive three different ways depending on which feature they happened to find first.
02
Categorize
The first taxonomy had five modes. It collapsed to three when I realized 'format changes' and 'structural builds' were the same intent — manipulation. That simplification unlocked the entire design logic.
03
Orchestrate
Designed the routing layer that detects intent and surfaces the right pattern. The harder problem: making 'consolidation' legible to stakeholders who assumed it meant cutting features.
The 6 existing patterns — before consolidation
Today, those needs are solved through 6 different features and patterns.
01
Assistant
A side panel that overlays the data content. Always visible, even when not needed.
02
Standalone Ask
A dedicated page for NL queries. Separated from the data, breaking user flow.
03
Contextual Insights
AI-generated insights surfaced inline on dashboards. Powerful but inconsistently placed.
04
Targeted Commands
Point at any element to trigger AI on it. Useful, but yet another siloed pattern.
05
Explain
One-click explanation of data points and anomalies. Useful but treated as an isolated feature.
06
Auto Insights
Proactive pattern detection surfaced automatically. Often ignored — no clear entry point.
05 — The Design
From feature sprawl to intent-based orchestration
The problem wasn't any one feature — it was the absence of a system. Each AI capability had been built independently, with no shared logic about when it should appear or why. Analysts were forced to navigate between surfaces, often not knowing which one to use.
The intent taxonomy changed the frame. Once every analyst query mapped to one of three modes — Answer, Explore, or Create & Edit — the design logic became clear: don't surface everything always. Route to the right pattern based on what the analyst is actually trying to do.
01
Meet analysts in their workflow
AI assistance should appear where work is already happening — not require a detour to a separate surface.
02
Match the surface to the intent
A factual question deserves a quick inline answer. An open-ended exploration deserves dedicated space. The surface should reflect the task.
03
Consolidate, don't cut
Every existing AI feature maps to one of the three modes. Nothing is lost — everything is reorganized into a coherent system.
Six disconnected AI surfaces — no shared model, no discoverability
A persistent chat panel competes with the canvas — visible even when the analyst isn't using AI.
No signal to guide users toward the right surface for their specific task.
Features feel bolted on. Each one a separate interaction model, a separate entry point.
Intent-based orchestration — one system, three modes, right surface every time
AI detects intent from the query and routes to the right pattern — Answer, Explore, or Create & Edit.
The interface stays clean until AI is needed. Assistance surfaces in context, not on top of the work.
One coherent model across all AI features. Analysts learn it once and it works everywhere.
06 — The Outcome
2
New patterns added to the roadmap
Both net-new — not retrofits of what already existed.
6→3
AI patterns consolidated into intent modes
Every existing feature preserved — reorganized, not removed.
3
Analytics apps aligned to one model
The intent taxonomy became the shared reference across the suite.
- ◎
Strategy approved and two patterns formally added to the product roadmap.
- ◎
Established a shared interaction model now used as a reference across Oracle Analytics — subsequent features are spec'd against it.
- ◎
The '3 modes' frame became the vocabulary PMs, engineers, and designers use when talking about AI in the product.
07 — Next
Next project
People Leader Scenario Modeling