Sonar — Data Dictionary with AI

Company: Sonar — B2B SaaS, Salesforce change management
Role: Senior UX Designer — sole designer: research, UX/UI, prototyping
Duration: ~6 months to first launch

average increase in deal size as a paid add-on: +40%
customer adoption — success target was 30%: 55%
from undocumented field to AI-written docs: 1 click

Production telemetry — Autodoc generating live documentation in customer orgs

Autodoc turned Salesforce documentation from a dreaded spreadsheet chore into a one-click, AI-generated system that stays current on its own. It shipped as a paid add-on, lifted average deal size 40%, and hit 55% adoption against a 30% target.

The Challenge

As a Salesforce org grows, one team's "small change" quietly breaks another team's process. The defense is documentation — and in practice that means sprawling spreadsheets that are outdated the week they're written. Customers told us upkeep was painful; what they couldn't tell us was how bad it had gotten, because nobody wanted to open the spreadsheet.

The objective: make documentation something the system does for you, not something admins owe the system.

Example of a messy spreadsheet that customers were using before Autodoc

The Solution: Autodoc

Interviews hinted at the problem, but the route revealed itself when I asked customers to open their actual spreadsheets on calls. A patchwork of stale tabs, half-filled columns, and tribal knowledge — that artifact reframed the work: the dictionary couldn't just be editable, it had to be self-filling.

I designed the data dictionary on top of Sonar's live metadata tracking, so entries update as the org changes. Then the key move: Autodoc, a one-click action that uses AI to draft context-aware help text and field descriptions, turning a blank-page chore into a review task.

Key Decisions & Rationale

This was 2023 — customers were curious about AI and trusted none of it. The design had to earn trust, not assume it: skeleton loaders make the AI's "thinking" legible, and every suggestion requires an explicit Accept so the admin stays the author of record. I de-risked the interaction patterns with high-fidelity prototypes, tested with customers and engineering before we committed a sprint to it.

Outcomes

Autodoc shipped as a paid add-on and became a deal-maker: a 40% average increase in deal size, 55% adoption across the customer base, and the manual-documentation pain point effectively retired for adopting teams.

"Our company is actually cutting software right now and we still pushed through the data dictionary add-on."

John at Alteryx

Reflection

Accuracy alone didn't buy trust — control and transparency did, and we only learned how much hand-holding the first AI interaction needs by watching people hesitate. The accept-every-suggestion model that built trust at launch becomes friction at scale; bulk actions and proactive suggestions were the obvious next climb, and personalization beyond generic context was still unexplored when I left the rope.