Turned messy order emails into verified CRM records. AI extracts, a human confirms.
Role
Lead Product Designer & Product Owner
Timeline
Team
Cross-functional Team of 5 (2 Designers, 3 Technical Architects)
Outcome
3rd Place Winner · Salesforce Innovation Competition
Core Design Pattern
Human-in-the-Loop (AI Extracts → User Verifies → CRM Executes)
The Challenge: Unstructured Emails & Manual CRM Re-Entry
Sales reps and order specialists routinely received multi-item customer orders buried inside unstructured, conversational emails. Manually parsing these emails, cross-referencing product SKUs, and re-entering order records into Salesforce was repetitive, slow, and highly vulnerable to data-entry errors.
Because inaccurate CRM data directly impacts inventory, billing, and customer fulfillment, full end-to-end AI automation posed a business risk. The solution required a workflow that maximized speed through AI while preserving human control over the final decision.
The Core Design Question
How might we leverage generative AI to automate unstructured data entry without sacrificing accuracy, transparency, or human control?
The 3-Tier AI Trust Framework
To build user trust and prevent silent AI errors, I designed the interaction model around three core pillars of AI ergonomics:
1. Predictable Extraction Transparency
UX Pattern: Side-by-side visual mapping comparing raw email text directly alongside AI-populated CRM fields so users can verify sources instantly.
2. Confidence Disambiguation & Smart Highlighting
UX Pattern: Visual indicators highlighting fields where the AI model has high vs. low confidence, drawing immediate human attention to ambiguous quantities or SKUs.
3. Reversible Execution & Human Control
UX Pattern: Inline field correction and 1-click batch confirmation. No record is created in Salesforce without explicit human approval.
1. Ingest (Email Processing 📧): Customer sends an order request via email. Einstein Validator automatically parses the message text, identifying line items, quantities, delivery addresses, and account names.
2. Verify (Human Review 🤖): The AI populates a structured order preview page. The seller reviews extracted values, checks highlighted low-confidence fields, and makes inline edits if necessary—cutting manual entry time from minutes to seconds.
3. Execute (CRM Commit 📦): With one click, the verified order is created directly in Salesforce CRM, triggering downstream fulfillment and notifying key stakeholders.
VERIFIED OUTCOME
Product Leadership & Impact
0→1 Product Ownership & Facilitation
As Product Owner and Lead Designer, I led a cross-functional team of 5 (2 Product Designers and 3 Technical Architects). I defined the product roadmap, aligned technical constraints with user needs, facilitated design sprints, and crafted the executive pitch presentation delivered to the leadership judging panel.
Key Outcomes & Recognition
PROJECTED VALUE — NOT A MEASURED RESULT
🏆
3rd Place Award — Awarded 3rd place out of company-wide submissions in Salesforce’s innovation competition.
⚡
80%+ Reduction in Entry Time — Projected to reduce order creation time from minutes of manual data entry to a 10-second verification click.
🔒
Zero Silent Errors — Preserved 100% human oversight, eliminating the risk of incorrect AI order execution.
PROPOSED FUTURE AI ROADMAP — NOT COMMITTED WORK
If extended beyond the competition MVP, the human-in-the-loop interaction model would expand to include:
1. Dynamic Confidence Thresholds
Auto-committing 99%+ confidence routine reorders while flagging complex custom orders.
2. Inline AI Disambiguation
Prompting the seller with smart suggestions when an email contains ambiguous product references.
3. Automated Exception & Fallback Flows
Providing guided recovery if CRM order creation fails due to inventory locks or account holds.
