KoltrixReviewing AI-drafted replies: a 60-second checklist
A quick checklist for reviewing AI-drafted email replies before you send: recipients, facts, promises, tone, links and the things a model simply can't know.
On this page(13 sections)
- The checklist at a glance
- 1. Right recipients and right thread
- 2. Answers the question actually asked
- 3. Facts and numbers are correct
- 4. No promises you haven't decided to make
- 5. Links and references are real
- 6. Attachments mentioned are attached
- 7. Nothing the model couldn't know
- 8. Tone fits this person and this moment
- 9. No leftover placeholders or odd phrasing
- Common catches in practice
- Making review fast
- Key takeaways
An AI draft that's almost entirely right is a trap, because it reads well enough that you stop looking for the one sentence that's wrong. The fix isn't to distrust every draft; it's to review the same few things every time, quickly, so the errors that matter don't slip through.
Here's a checklist that takes about a minute per reply once it's habit. Each item includes the kind of mistake it catches.
The checklist at a glance
- Right recipients and right thread
- Answers the question actually asked
- Facts and numbers are correct
- No promises you haven't decided to make
- Links and references are real
- Attachments mentioned are attached
- Nothing the model couldn't know
- Tone fits this person and this moment
- No leftover placeholders or odd phrasing
Go top to bottom. The first items catch the most expensive mistakes.
1. Right recipients and right thread
Look at To, Cc and Bcc before you read the body. Mistakes here are the hardest to undo.
Catches: a reply-all that includes the customer on an internal discussion; replying to the colleague who forwarded an email instead of the customer; a draft attached to a similar but different thread.
Quick check: Is everyone on this email someone who should see every word of it?
2. Answers the question actually asked
Models sometimes answer the question they expected rather than the one in the email. A customer asking "Can I move my account to a different email address?" might get a well-written explanation of how to change a display name.
Quick check: Reread the customer's last message, then the first two sentences of the draft. Do they match?
3. Facts and numbers are correct
Dates, prices, plan names, limits, version numbers, order numbers: anything specific deserves a glance. Models are fluent with numbers whether or not the numbers are right.
Catches: "Your renewal is on March 3" when it's March 13; a limit or price that's outdated; the wrong plan name.
Quick check: For every number or proper noun, ask: do I know this is true, or does it just sound right?
4. No promises you haven't decided to make
This is the item most worth slowing down for. Watch for future-tense commitments:
- "We'll refund the full amount."
- "This will be fixed in the next release."
- "I've extended your trial."
- "We can add that feature for you."
Each of these is a business decision. If it's a decision you'd make, fine. If not, rewrite it to what you can actually commit to: "I've passed this to our billing team and will update you by Thursday."
5. Links and references are real
AI drafts sometimes include help-center links that look right but don't exist, or reference a doc by a plausible-sounding title.
Quick check: Hover or click every link. If you can't verify it in a few seconds, remove it or replace it with one you know.
6. Attachments mentioned are attached
"I've attached the invoice" with no invoice attached is a classic, and models write it easily because it's a common phrase.
Quick check: Search the draft for "attached," "attachment," "see the file." If it's mentioned, confirm it's there.
7. Nothing the model couldn't know
The model knows what's in the thread and whatever context your tool provides. It doesn't know about the call you had yesterday, the exception your manager approved, or the fact that this customer is on a custom contract.
Catches: a standard policy answer for a customer who was promised something different; a cheerful reply to someone whose issue was discussed internally as serious.
Quick check: Is there anything I know about this customer or situation that isn't in the email thread? If so, does the draft account for it?
8. Tone fits this person and this moment
The words can be fine while the tone is wrong. A customer on their third frustrated email doesn't want "Thanks so much for reaching out!" A long-time customer may find a formal template cold.
Quick check: Read the customer's message and the draft's opening line together. Would a thoughtful person who knew this customer open that way?
9. No leftover placeholders or odd phrasing
Last pass for things like "[Customer name]," a sign-off with the wrong person's name, repeated sentences, or phrases nobody on your team would say.
Common catches in practice
| What the draft said | What was wrong | Fixed version |
|---|---|---|
| "Your plan includes unlimited seats." | Plan has a seat limit | "Your plan includes [N] seats; you can add more in Billing." |
| "We'll have this fixed by tomorrow." | No one has committed to a date | "Our engineers are investigating. I'll update you by tomorrow afternoon." |
| "See our guide: help.example.com/export-data" | Page doesn't exist | Link to the real export article |
| "Happy to help with that!" | Customer's third complaint about the same bug | "I'm sorry this is still happening. Here's where it stands." |
| "I've attached the receipt." | Nothing attached | Attach it, or "You can download the receipt from Billing → History." |
Making review fast
The checklist is fast because most items take a glance, and most drafts pass. A few habits help:
- Review in the same order every time. Consistency is what makes it quick.
- Fix the draft, don't just approve it. If you change something, that's the review working.
- Notice patterns. If drafts keep inventing the same wrong fact, update your snippets, style guide or the context your tool uses.
- Don't review when rushed. If you have thirty seconds, leave the draft for later rather than skimming and sending.
In Koltrix, an AI draft stays a draft until someone sends it; nothing goes out without a human click. That design only protects you if the click comes after a real look, which is what this checklist is for.
Key takeaways
- Review the same items in the same order every time; most drafts will pass quickly.
- Check recipients first, because those mistakes are the hardest to undo.
- Treat every number, link and future-tense promise as something to verify.
- Ask what you know that the model couldn't, and make sure the draft reflects it.
- Feed recurring mistakes back into your snippets and style guide.
Start with Koltrix
Your domain, one inbox, and an API that sends.
A team inbox where AI sorts and drafts (nothing is sent without your click), plus the transactional API and SMTP relay your product sends with. 7 days free, no card.
