Daisy

Overview

Daisy is an AI-powered assistant web software designed to streamline daily tasks and capture insights for streaming companies.

The Welcome screen showcases available AI chatbot domains for Warner Bros Discovery channels and clients, including CNN, HBO MAX, HGTV, Paris Olympics, COSMOS, and Nielsen Media Ratings. Clicking on a card starts a chat with the AI Copilot for that specific domain. The sidebar keeps a chronological record of all conversations, as well as the ability to save the most twop three most recent domains used.

Daisy is an AI-powered assistant web software designed to streamline daily tasks and capture insights for streaming companies. It integrates with various platforms to provide a seamless user experience, enhancing productivity and decision-making.

AI System Model

Daisy is a domain-scoped AI copilot designed to support analysts through guided retrieval, generation, and review.

User intent is captured via prompts or inline questions, scoped by explicit domain selection, and returned as a reviewable draft.

The system performs no autonomous actions; all outputs require explicit human validation.

Context Grounding & Retrieval Strategy

Enforced retrieval boundaries through explicit domain selection as a UX constraint.

Prevented cross-domain leakage and hallucination by limiting retrieval to approved datasets per domain.

RAG behavior was defined at the interface level rather than left to implicit model context.

Problem

Analysts and stakeholders needed faster ways to query and summarize domain-specific information across multiple business areas.

Existing workflows were fragmented across tools, making it difficult to keep context, reuse prompts, and standardize outputs.

Goals

Reduce time-to-answer for common questions with guided prompts and domain selection.
Keep conversations organized and recoverable through history and pinned domains.
Maintain a UI that feels familiar to the client ecosystem while staying accessible and scalable.

Process

Step 01
Mapped user flows for domain selection → prompting → conversation → prompt management.
Step 02
Designed key screens within existing design system and accessibility constraints.
Step 03
Iterated on prompt cards, conversation actions, and admin tools based on stakeholder feedback.

Failure States & Trust Preservation

Failures are surfaced as first-class UX states, distinguishing between system, network, and input issues.

User questions and context persist across failures, enabling retry or refinement without re-entry.

Explicit uncertainty reinforces trust and keeps users in control.

AI UX Principles Applied

Preserve user intent across retries and failures.,Make AI state and uncertainty visible.,Treat AI output as a draft requiring human review.,Favor refinement over replacement.

AI Design Framework Ownership

Defined and owned the AI interaction framework governing prompting, generation, review, failure handling, and governance.,Established reusable UX patterns for conversational state, refinement loops, error recovery, and human validation.,This framework scaled across domains and teams as the standard model for AI-assisted analysis.

Intentional Constraints & Non-Automation

Daisy was explicitly designed as a copilot, not an autonomous agent; it cannot execute actions, modify data, or trigger workflows.

All AI output remains advisory, reviewable, and reversible by the user.

Prompt Governance & Control Plane

Administrators manage prompt creation, versioning, and retirement.

Governance ensures consistent AI behavior across domains while balancing flexibility with enterprise control.

Outcomes

Enabled faster, repeatable analysis through domain-scoped prompting and reusable conversations.
Established a review-first conversational UX suitable for enterprise decision-making.
Delivered a governed prompt system that scaled safely across multiple business domains.

UI Walkthrough

A screen-by-screen breakdown of intent, interaction, and design decisions.

Domain Selection

Preselected Chat Interface

The active chat interface shows the selected domain and metadata and four suggested prompt cards with the most asked questions. From here, the user can click on a prompt card to start a chat with the AI Copilot for that specific question or type in a custom question below in the askbar provided. The sidebar is still available and can be used to navigate to different domains or to clear the chat history. Once a conversation is started, the domain will be added to the sidebar as well as to the recent domains list. The user can also choose to pin any of the most recent domains to the top of the list.

Conversational Chat Interface

The chat interface uses a conversational UI with question and draft responses facing each other in a chat bubble format. This allows the user to quickly distinguish between the user's question and the AI's response. The avatars are displayed on the left and right of the chat bubble, respectively. The user's question is in a blue bubble, while the AI's response is in a green bubble. The user is also able to copy, like, dislike, and refresh the response for review, as well as edit the question if needed and regenerate it.

Inline Question

The inline question is a feature that allows the user to ask a question directly in the chat interface without having to click on a prompt card or use the askbar. This is useful for when the user wants to ask a question that is not on the prompt card or when the user wants to ask a question that is not related to the domain. For productivty purposes, it simplifies the process of asking a question and allows the user to focus on the conversation rather than navigating the interface or typing the question.

Generating Response

The generating response animation is displayed when the AI is generating a response. This is useful to keep the user informed that the AI is working on the question and to avoid the user from thinking the AI is stuck or not working. This animation is displayed for a maximum of 10 seconds and then the response is displayed. A pulsating skeleton loader is used to indicate that the AI is working on the question.

Failed Response

The failed response text is displayed when the AI is unable to generate a response in a descriptive manner rather than a generic AI uncertainty or failure state message (i.e. "Network Interruption, Please Try Again", "Server Error, Please Try Again", "AI Processing Error, Please Try Again", "We need more details to answer your question, please edit your question and try again"). This is useful to keep the user informed that the AI is unable to generate a response and to encourage the user to try again as well as to inform the user of the type of error that occurred. Failure state preserves user context and allows retry without losing the original question.

Settings - My Prompts Management

The administrative settings interface is for admins only and uses a minified verison of the sidebar since it doesn't interact with the chat interface. The My Prompts section allows the admin to manage the prompts for the AI Copilot among other settings not shown in this screenshot to protect the confidentiality of the client. The user can add, edit, and delete prompts, as well as view the history of prompts used. The user can also export the prompts to a JSON file for backup purposes.

Prompt Governance

The prompt governance section allows the admin to manage the prompts for the AI Copilot. The user can add, edit, and delete prompts, as well as view the history of prompts used. The user can also export the prompts to a JSON file for backup purposes. The user can also see the prompt's usage metrics and the prompt's performance metrics.

Key AI UX Tradeoffs

Speed vs trust: prioritized reviewability over instant execution.,Flexibility vs governance: constrained prompts to reduce risk.,Autonomy vs control: positioned Daisy as a copilot, not an agent.

Learnings

Explicit context boundaries reduce both cognitive load and model error.
Treating AI output as a draft materially increases user trust and adoption.