2026

Insight Flow: Conversational BI Platform

Target Audience

Executives and managers at medium-sized companies (B2B SaaS)

Problem

Executives and managers constantly ask analysts questions such as: Why has revenue dropped? Which customers are churning? Why has revenue grown in Germany but not in France? Which marketing channels deliver the best ROI?

Preparing an answer involves:

Finding the right dashboard
Filtering
SQL queries
Follow-up questions

Objective

To create an AI interface that will allow business users to explore data on their own through dialogue

My achievements

Identify the positioning and key features of the product
Design a usage flow and information architecture
Create a UI and prototype
User testing on synthetic personas
Design backlog management

Tools used

Claude Sonet 5
Figma Make
Chat GPT-5.5

Competitor Analysis

Insight: The three competitors are solving different but partially overlapping tasks:
• Power BI Copilot → access to data within the Microsoft ecosystem
• Tableau Pulse → KPI monitoring and proactive alerts
• ThoughtSpot → data search and self-exploration

None of them systematically solves the issue of trust: "Why should I believe this answer?" — the answers are given without layers of evidence, certainty, and connection to the decision.

Positioning

Conversational BI for Decision Confidence. It «Ask your data», а «Understand why the answer is true».

Principle 1: Evidence-first AI

Each response is accompanied by the data sources used, the metrics applied, the contribution of various factors, and visual confirmation—there's no need to search for evidence separately (unlike Power BI Copilot, where you have to trust the model's "black box")

Principle 2: Confidence Layer

Each response includes an explicit confidence level (High/Mid/Low), the number of datasets used, and the percentage of data completeness—something none of the three competitors explicitly provide.

Principle 3: Decision Mode

Instead of simply stating a fact ("sales fell by 12%"), the product structures its response as Problem → Cause → Impact → Recommended action—this is closer to a recommendation system for decision-making than to a dashboard or data search engine.

The principle of nesting within the dialogue

Each AI response has its own local architecture according to the trust layers of the concept

Personas

Each role has its own criteria of trust. The 5 levels embedded in the archetect serve the criteria of trust of each

Response Architecture

The answer is never displayed 'raw' — there's always an initial self-assessment

If a problem is found, the system first tries to fix it itself. Only if this does not help is the reason disclosed to the user. If there is a lack of data, it explains exactly what is incomplete and what risk this poses. If there is a technical error, it suggests waiting or receiving a notification. The product always explains why confidence is not 'High', rather than simply marking the answer with an icon

Prototyping

ASK section

New dialogue · Active dialogue · Evidence panels

Scenario
Question → Answer → Evidence → Certainty → Action
The user receives a response and immediately sees the specific dashboard containing the settings, the information from which was used to generate the response.

Key feature
After generating a response, before displaying it to the user, the system automatically analyzes the response's accuracy. The self-assessment result is shown to the user along with the response.
If the accuracy is average or low, the system attempts to correct the result. If it is impossible to generate a more accurate result:

Due to insufficient data, the system displays a self-assessment rating of "average" or "low" and explains the reason and associated risks.

Due to a technical failure, the system displays preliminary data and asks the user to make the request later or to request a notification when the source server is able to return a response.

History

All Threads
Folders
Topics (auto-tags: Sales/Marketing/Finance)
Saved
Shared

Organizing and reusing past investigations. The 'Saved' folder is the default option for opening the 'History'

Data Sources

List of Connections and Status
Add a Source (Connection Wizard)
Semantic Layer / Metrics Dictionary
Source Access Rights
Synchronization Logs
Ensures transparency and control over where data comes from and how metrics are calculated
Allows the administrator to configure the semantics

Insights Hub

Proactive Insights Feed
Metrics Subscriptions (watchlist)
Anomalies (similar to SpotIQ)

Key Feature
Analyzes dashboards within the user's role-specific area of interest.

In the event of critical anomalies, it highlights an indicator to attract the user's attention to that section.
The rest of the time, it highlights curious patterns and serves an educational purpose by showing other ways to analyze the data.

Share Center

Logging of sent analytical data
Tracking of subsequent use
Viewing statuses

Converting an AI response into a document for external communication. Allows the user to track subsequent activity based on the data that was sent.

Settings & Admin

Users and Roles
Integrations (CRM, Ads, DWH, etc.)
Audit Log
Billing

Company-wide Platform Management

Validation

The testing used 3 AI-generated characters representing the user segments of the conversational AI assistant. These characters covered motivation, proficiency and usage scenarios

Link: Testing methodology and results in Coda

Summary

All the previous steps helped me to

Identify competitive advantages for launching the product in the B2B market and stand out from the competition
Create the interface for the Conversational BI Platform
Pre-validate the mock-ups using synthetic personas and make adjustments based on the results of the initial testing
Create a backlog design for further product development

Key insights

By using synthetic personas generated on Claude Sonet 5, preliminary testing was carried out and the results analysed in 3 hours at no cost. Yet this alone enabled us to identify risks and refine the mock-ups. Consequently, testing with real users will be more focused on personal and role-based experiences
I refined the UI, optimised the UX interaction, added supporting screens and responsive design to the existing mock-ups generated by Figma Make. This saved me several hours of work
I used Claude Sonet 5 to prepare a prompt for Figma Make, containing all the requirements and stylistic preferences.
I used Claude Sonet 5 and ChatGPT-5.5 for market analysis and to define positioning and key features

Throughout my work, I adhered to the GenAI principles

Breaking down the response into layers with supporting evidence helped to achieve the principle of Progressive Disclosure
Following a brief initial response, subsequent layers corroborate the first, which aligns with the principle of Trust & Verification
At the end of the response, the Conversational BI Platform suggests possible avenues for the dialogue to take within the context of the topic specified by the user. This corresponds to the Human-in-the-Loop principle
The Interactive approach is achieved when the system cannot provide a High Confidence response and interacts with a human to jointly determine how to proceed