Project Case Study
Chatbot for Operational Data (Plain-English Querying)
A conversational assistant embedded into operational systems, enabling users to query approved datasets in plain English—without exposing raw databases or bypassing governance controls.
Context
Operational users often need quick answers from structured systems (time entries, workflow status, client metrics, utilisation, etc.). Traditional dashboards require navigation and filtering knowledge. SQL requires technical expertise. The objective was a safer middle ground: conversational access within defined boundaries.
Problem
- UI friction: users struggle to find the right report or filter combination.
- Dependency on analysts: simple data questions create reporting backlog.
- Security risk: unrestricted conversational AI could overexpose sensitive datasets.
- Trust concerns: users need confidence in both accuracy and scope control.
Approach
- Controlled dataset layer: exposed only pre-approved, scoped views—not raw tables.
- Intent classification: determined question type (metric, comparison, filter, status).
- Conversation memory (bounded): retained short session context while preventing cross-session leakage.
- Policy enforcement: role-based filtering before any data query is executed.
- UI integration: embedded chatbot inside application dashboard with contextual awareness.
How It Differs from English → SQL Reporting
- Focuses on approved operational views, not dynamic free-form SQL generation.
- Optimised for conversational flow rather than complex analytics.
- Prioritises guardrails and scope limitation over flexibility.
- Designed for embedded, day-to-day workflow usage.
Conversation Patterns
- “How many hours did I log this week?”
- “Show open workflows assigned to my team.”
- “What’s the current utilisation for Client X?”
- “Compare this month’s billing to last month.”
Design Principle
The chatbot is not a database replacement. It is a governed conversational interface layered on top of approved operational metrics.
Next Enhancements
- Context-aware prompts based on the screen user is viewing.
- Smart follow-up suggestions (“Would you like a breakdown by matter?”).
- Confidence indicators explaining scope and assumptions.
- Escalation pathway to full analytics mode when deeper reporting is needed.