Result
Time spent regenerating drafts fell from ~15 minutes to under 2 per email, measured in production telemetry.
Company
6sense
Year
2025
Role
Product Designer III · with 1 PM, 2 data scientists and 4 engineers · 4 weeks
The problem
6sense's AI Writer drafts sales emails from account intent signals and CRM data. The writing was good, but sales reps couldn't control it:
No comparable drafts. Each regeneration replaced the last one, so a better earlier draft was lost.
No editing in place. Reps copied drafts into a notepad, edited them, then pasted them into their email.
No way to steer it. Reps couldn't give instructions or adjust a single draft.
The first draft was never the one reps used, so they kept clicking regenerate, losing each draft as the next appeared. From the first regenerate click to leaving AI Writer took around 15 minutes per email. I framed the core problem as trust, not output quality: reps won't send what they can't control, check or recover.
In Image: v1 AI Writer, annotated with the three gaps
Four decisions that built trust
1. Keep every draft. All generations stay in a scrollable thread, so reps can compare, go back and choose instead of regenerating blindly.
In Video: The draft thread with several generations
2. Steer instead of regenerate. Reps edit drafts inline or type a prompt to shape the next one, inside AI Writer. No more notepad, and no more hoping the next draft is better.
In Image: Inline editing and the prompt input
3. Let the system explain itself. Generation counters, preference indicators and like/dislike feedback show what the AI is doing. Like/dislike also feeds a signal back to Data Science to improve the model.
In Image: System explaining itself
4. Design for failure. Wrong prompts, conflicting settings, running out of credits and empty outputs each get a clear state, and reps can check where AI-written content came from. Trust grew when failures were acknowledged, not hidden.
In Image: Edge cases and error states
The final experience
Set up once. Reps choose insights, tone, email framework and formatting. Settings are saved across sessions.
Generate in one click from the contact's profile, using intent signals, CRM data and saved preferences.
Refine and send. Edit inline, prompt or regenerate, compare drafts in the thread, then copy the final version.
What I ruled out
Full automation. Auto-sending AI emails reduced confidence. Reps need to own what goes out under their name.
A single “best” draft. One top-ranked guess felt arbitrary. Reps needed to compare and go back.
A separate writing tool. It would break the sales workflow. The writer had to live where the work happens, on the contact's profile.
Trade-offs
Thread vs simplicity. The thread added interface weight, but it removed the fear of losing a good draft.
Open prompts vs guided input. Free prompts gave control, so they needed guardrails for irrelevant or conflicting instructions.
Transparency vs clutter. Counters and indicators built trust only if they stayed lightweight.
Results
Time spent regenerating fell from ~15 minutes to under 2 per email, from a rep's first regenerate click to leaving AI Writer, measured in production telemetry before and after the release.
The same release also updated the AI model and adjusted the email frameworks, so this is the result of the whole release, not the design alone. The design's part was replacing blind regenerating with steering: keeping every draft, editing inline and prompting.
Used in sales teams' daily prospecting.
What I learned
A better model alone doesn't stop people regenerating. Reps needed to steer, compare and recover drafts before they could trust what the AI gave them.
