Together with Inbal Tal (Senior Product Manager) we led the product, and design from the earliest research through the shipped experience (0 to 1). My responsibilities included: Product strategy, user research, information architecture, UX and UI, AI interaction design, trust and explainability, design system, prototyping, engineering collaboration, and product vision.
Note: Some of the screens shown here span different stages of the process, sketches, research prototypes, usability tests, MVP, and vision work, which is why the product appears under a few different names throughout, CRM Update, Auto-Update, AI Suggestions, among others.
Manual work
Updating CRM after meetings required significant effort. Most users postponed it until later, and many never got around to completing it at all.
Lost context
Important information stayed inside transcripts, personal notes, or people’s memory. CRM records often ended up with only partial, incomplete information.
Low trust
Users appreciated AI-generated summaries but hesitated to let AI update customer records automatically. Trust mattered more to them than automation.
Survey highlights
Sales methodology templates topped requests at 87%, followed by auto-logging to CRM at 73% and auto-transcription at 69%.
“I love this, it’s exactly what I’ve been asking for.”
“If it’s working, this could be a game changer for us, honestly.”
“We struggle taking notes, so just having this run in the background helps a lot.”
Users told us they needed to see where a suggestion came from, “if I can click and see where it came from, I’d trust it more.” We responded with source citations, every suggested update links to a timestamped moment in the original interaction.
Others worried about blanket trust, “we’d need to constantly review that everything is correct.” We built field-level approval instead of an accept-all, so users could trust each suggestion on its own terms.
One participant flagged the workflow itself, “it should have been a pop-up right after the call, not a new page.” We redesigned the review flow to surface inline, right after the interaction, instead of sending users somewhere else.
Another wanted control over what AI could touch, “I’d want it to add context, not replace what’s already there.” We made append and merge the default, overwriting only on explicit intent.
And one comment cut to the core of it, “if I make a mistake, can I undo it?” That became undo and version history on every AI-applied change.
Trust before automation
Users should always remain in control.
Show the source
Every suggestion traces back to where it came from.
Stay inside existing workflows
No additional tools or tabs.
Preserve customer context
Add to what’s there, don’t silently replace it.
Recovery is a feature
The ability to undo is what makes automation safe to trust.
Field-level granularity
Approve Deal Stage without touching Deal Value.
Show the source
Every suggestion traces to a specific moment in the source interaction.
Preview before commit
Test AI behavior before applying it permanently.
Audit trail
Managers see which updates were AI-generated and who approved them.