Why we always generate three drafts, never one
A short essay on why "the right answer" is the wrong product surface for support, and what we picked instead.
Alex Rivera
Co-founder & CEO · Mar 30, 2026 · 5 min read
Every AI-writing product has to make the same decision: one suggestion, or many? We picked three, and I want to explain why — because it's load-bearing for the whole product.
One draft is a verdict
When you generate one draft, you're telling the agent "this is the answer." Agents either copy it (and learn nothing) or fight it (and feel slower than typing from scratch). Neither is the relationship we want with the tool.
Worse, one draft hides the model's uncertainty. The model doesn't know if the customer wants a concise refund acknowledgement or an empathetic apology with shipping options. Picking one and hiding the alternative is the LLM equivalent of a confident hallucination.
Three is the smallest number that creates choice
Two drafts feels like A/B. Three feels like a menu. We use three deliberately:
- Balanced — the safe middle. Closes the ticket without overcommitting.
- Concise — the one-liner. For customers who just need a yes/no.
- Detailed — context, options, next steps. For when the customer is frustrated or confused.
Almost every reply could be answered well by one of those three. The agent's job stops being "write a reply" and becomes "pick the right mode for this customer, then nudge it." That's a much faster cognitive task.
What the data says
Across our customer base, the balanced draft gets selected 54% of the time, concise 28%, detailed 18%. The interesting number is the edit rate: balanced drafts get edited 31% of the time before sending, concise 14%, detailed 47%. Detailed drafts are doing the most heavy lifting precisely because they're the longest and have the most room to be wrong — but agents still prefer starting from them on hard tickets, because trimming is faster than expanding.
What we tried before three
Early QuickPly generated five drafts. Agents spent more time choosing than they would have writing. We dropped to two — paralysis went away, but so did the sense of having options; teams told us it felt like a forced binary. Three was the number where the choice felt useful but not exhausting.
The deeper lesson: when you're designing AI tools, the model isn't the product. The interaction is. Three drafts isn't an LLM decision — it's a UX decision, and it's the single most important one we've made.