Einstein Validator
Designing trust, transparency and control into AI that turns customer emails into CRM orders.
The stakes
Multi-item orders arrived buried in conversational emails, and specialists re-keyed every one into Salesforce by hand.
It was slow, repetitive and error-prone, and errors were costly: bad CRM data flows straight into inventory, billing and fulfillment. That made full end-to-end AI automation a business risk, not just a shortcut.
The call I made
AI proposes. People decide.
Rather than chase full automation, I designed the product around a human-in-the-loop workflow: the AI does the extraction, and no record reaches Salesforce without explicit human approval.
See it work
The assistant lives inside the account record, so the seller never leaves Salesforce to validate an order.
Three decisions that build AI trust
No new form for the customer
Customers keep ordering the way they already do, in a plain email. The AI does the work of turning that message into structure, so nobody on either side has to change habits for the system.
Show exactly what the AI understood
The assistant plays back the customer, the account and every item with its quantity in the same terms as the original email, so a seller can check it against the source at a glance. Values the AI is less sure of are flagged for a closer look.
Nothing ships without a yes, and every yes can be undone
Explicit confirmation, one-click undo and a plain-language summary keep the seller in charge. Only then does the verified order land in Salesforce, item for item.
How I led the team
Owned the roadmap
As product owner, defined the roadmap and balanced technical constraints with user needs.
Shaped the architecture
Defined the interaction model and product architecture with 3 technical architects, and ran the design sprints.
Pitched to leadership
Crafted and delivered the executive pitch to the leadership judging panel.
Outcomes
Where the model goes next
Dynamic thresholds. Auto-commit routine reorders above 99% confidence; flag complex ones.
Inline disambiguation. Suggest matches when an email’s product reference is ambiguous.
Guided recovery. Clear fallbacks when order creation fails on inventory locks or account holds.
“Janie’s ability to coordinate, encourage, and motivate a busy team of people was a key success factor that enabled the team to complete the solution within the competition timeframes.”
