An exploration of how AI can support enterprise data analysts — from discovering fragmented context to generating recommendations, while keeping humans in control of the final decision.
AI synthesizes documentation & fragmented legacy schemas upfront.
AI generates candidate field mappings backed by concrete evidence.
Analyst inspects data samples, spec citations, and confidence ratings.
Analyst overrides edge cases and formally approves mappings into pipeline.
Background execution runs without blocking analyst context switches.
The brief was open-ended: “What could AI do for enterprise data integration?”
It would have been easy to immediately start designing a conversational chatbot or an automated mapping wizard. But jumping straight to features would have solved the wrong problem.
Enterprise source-to-target mapping is not simply drawing lines between columns. It is an arduous process of reconstructing fragmented context across legacy schemas, missing business definitions, scattered documentation, and high regulatory audit risk.
“Don't start with ‘What AI feature should we build?’ Start with: ‘Where is the analyst actually spending cognitive effort?’”
AI could parse incoming tables and automatically draw mapping lines between source and target columns.
Analysts spent 80% of their effort acting as data detectives—hunting down cryptic abbreviations (TX_CAT_V2), wikis, and PDFs before linking a single field.
Instead of “Where should AI appear in the UI?”, I asked: “How can AI help the analyst understand the data before mapping begins?”
TX_CAT_V2) with zero self-describing documentation.
Enterprise schemas contain tens of thousands of columns. Querying everything at once produces untrusted results. Working with PMs across Ingestion and Catalog squads, I bounded the AI's scope into hierarchical domain containers and established clear boundaries between deterministic rules and AI inference.
Legacy tools treated integration as a simple line-drawing interface. But analysts spent 80% of their effort hunting through PDFs and deciphering cryptic columns before linking a single field. I redesigned the sequence so that AI-driven document discovery and glossary synthesis occur upfront, transforming the analyst into a high-leveraged reviewer.
Manual investigation of cryptic fields and scattered PDFs.
AI extracts domain glossary; analyst verifies recommendations.
AI recommendations cannot simply appear as an opaque “magic” result. In enterprise compliance, an unverified link creates major regulatory risk. I designed an inspectable interaction model where every AI recommendation provides supporting evidence, data samples, and spec citations—giving analysts total confidence to refine and approve.
AI processing on large enterprise schemas takes minutes or hours to run. Forcing an analyst to sit through a blocking loading spinner destroys productivity. I architected an asynchronous, non-blocking model and prototyped multi-product context switching in Figma and Codex—allowing analysts to dispatch tasks, continue other work, and return precisely where they left off.
The outcome of this exploration was not a standalone chatbot or a decorative widget bolted onto existing screens. It was a comprehensive product model connecting the four decisions into a cohesive, non-disruptive workflow:
Synthesizing unstructured specs, parsing thousands of candidate columns, decoding naming patterns, and running heavy jobs asynchronously in the background.
Inspecting sample data evidence, validating semantic meaning, overriding ambiguous edge cases, and taking formal audit ownership of the final integration.
“I didn't start by asking: ‘What can AI do?’ I started by asking: ‘What does the analyst need help with?’ That led to: ‘Where should AI participate — and where must the human remain in control?’ And finally: ‘How can AI continue working without making the analyst stop working?’”
Defined the bounded domain containment model and established the work allocation framework separating rules from inference.
Created the 4-stage inspection model and evidence drawer, establishing compliance-ready AI interaction patterns.
Proved asynchronous multi-product switching and return-to-context behaviors with working interactive prototypes.