AI Reply Generator: What It Does Well and Where It Embarrasses You
AI reply generators are good enough now that using one is not embarrassing. But they fail in a specific, predictable way, and knowing the failure mode is the difference between “saved me an hour” and “sent a promise I cannot keep”.
What it is genuinely good at
- First drafts. Getting from blank page to something editable in three seconds.
- Tone shifts. “Make this warmer / shorter / less defensive” works remarkably well.
- Rewriting your own confused sentence. If you half-know what you mean, it usually finishes the thought cleanly.
- Translation and formality. Going from terse to polite without sounding stiff.
Where it quietly fails
- Specifics. Dates, prices, order numbers, policy details. It will invent a plausible one.
- Commitments. “We'll ship by Friday” is a promise, not a sentence. Never let a model make promises on your behalf.
- Admitting fault. Models hedge. “We apologise for any inconvenience” is exactly the phrase that makes a complaint worse.
- Anything legal, medical or safety-related. No.
- Your voice. Left alone, it writes like a polite corporation, which is the opposite of what most people want.
The workflow that avoids the mess
- Write the answer yourself, badly. Ten seconds, no polish. It contains the facts.
- Ask the model to rewrite that. Not to answer from scratch.
- Give it three to ten of your own past replies as style examples.
- Check every number and name before sending. Always.
Why examples beat instructions
“Be friendly and concise” produces generic friendliness, because every model has seen that instruction a million times. Three real lines you wrote last week produce something that sounds like you — rhythm, length, punctuation habits included. Instructions describe; examples demonstrate.
This is exactly what ReplyTray does when you ask it to draft: it reads your saved replies first, uses them as the style sample, then writes. With nothing saved, it can only be generic — which is a feature, not a bug, because it pushes you to build the library. The longer version of this idea.