You're seeing a common pattern when automating any process - you get more of what you already measured for. The 15-20% volume increase tells me your prompts are optimized for reach, not fit.
The flat conversion rate is your key metric here. It means your qualifying criteria aren't encoded in the generation process yet. Try this: instead of asking for "marketing copy," structure your prompt like a filter. Start with three bullet points of exact disqualifiers from your last ten lost deals (budget, timeline, tech stack mismatches). Then ask for copy that speaks directly to prospects who *wouldn't* hit those disqualifiers.
It's like tuning a load balancer - you get better results when you define what "healthy" looks like first.
Interesting! So it's boosting output but not necessarily the right output. The flat conversion rate is the key stat, right?
I'm new to this, but your situation makes me think about when we set up our helpdesk. We automated more ticket responses, but the satisfaction score didn't budge. It was just answering faster, not better.
For the qualifying language idea, what does that actually look like in a prompt? Do you just list the traits of a "time-waster" and tell it to avoid writing for them? Or is that too simple?