100% of my customer replies are now AI-drafted first, edited by me second. It’s cut the time I spend on communication quite a bit, and somehow made my replies sound more like me, not less.
I’m a small business owner running a very lean team. I’m not an AI expert. I’ve been trying to increase AI use in my business for a while, and I haven’t managed to fully automate anything yet. I also don’t know how my AI usage compares to other businesses, and a lot of what’s out there right now suggests reliability is a real, growing concern. (That’s worth its own post.) So take this as one practitioner’s experiment, not a playbook: here’s exactly what I did, what worked, and what didn’t.
The experiment: what changed
This month I moved almost all my pre-sales and post-sales replies to an AI + human process. AI drafts the reply, I review and send it.
How it was done
- I uploaded around 300 FAQs to ChatGPT: real answers I’ve written and actually use, covering pre-sales, post-sales, and various use cases.
- I added our policies and general website information as plain text.
- I paste in customer queries and tickets exactly as they arrive, no cleanup.
The replies I get back are the most relevant and usable I’ve had from AI so far, and they sound like me.
Did it work?
Yes, on the main thing I cared about: time. This task now takes noticeably less, and the quality has stayed high, in some cases better, since it’s consistent and pulls from the same base of good answers every time.
What didn’t work
It misses about 20-30% of the time, mostly on policy parts. One reminder usually fixes it. I still read and edit every reply before sending, so nothing goes out unchecked. This isn’t a “set and forget” system yet, it’s a faster first draft. But despite this minor failure, it is currently saving me a lot of time & mental effort.
Learnings, what’s next
Right now my prompts live in Google Docs and I paste them in every time I need one. Next step: separate chat threads or projects for each use case; one for product descriptions with the template already loaded, one each for pre-sales and post-sales with context already set up. The idea is that each task gets a ready-made home instead of me re-explaining it every time.
Is this technically an agent? Not yet; an agent acts on its own, mine still waits for me to hit send. But it’s a step toward one, and I’ll write about it as I get further.
If you want to try this
- Pull together your best real answers and upload them as context.
- Paste customer messages in raw, don’t clean them up first.
- Keep a human check in the loop until you know exactly where it fails.
Leave a Reply