Start with drafts that a person sends. Let AI send on its own only for a narrow kind of email where a wrong answer is cheap to correct, the facts come from one stable source, and you have seen those drafts go out unchanged for weeks. Keep refunds, complaints, legal questions and anything that commits money or a date with a person, whatever the tool can do.
What is the difference between a draft and auto-send?
A draft is a reply the tool writes and leaves for you. Nothing reaches the customer until a person reads it and presses send. Auto-send means the tool sends the reply itself, usually only for emails that match rules you set, such as a sender, a subject or a topic.
The writing is the same in both cases. What changes is who catches a mistake: with drafts, you do, before the customer sees it. With auto-send, the customer does.
What does auto-send risk?
Every wrong reply costs something. The question is how much, and whether you can take it back.
| Risk | Example | How hard to undo |
|---|---|---|
| Wrong fact | “Delivery takes 2 days” when it takes 5 | Easy: send a correction |
| Promise you can’t keep | “We’ve refunded your order” when nobody has | Hard: the customer acts on it |
| Money or legal commitment | Agreeing to a discount, a contract change or a deadline | Hard: may bind you |
| Wrong tone at a bad moment | A breezy reply to a complaint about a damaged gift | Hard: trust is already lost |
| Reply to the wrong kind of mail | Answering a supplier’s invoice as if it were a customer question | Easy, but embarrassing |
| Leaking information | Mentioning another customer’s order in a reply | Very hard |
The first and fifth rows are what auto-send is usually about. The others are why most emails should stay with a person.
Which emails can be sent without a person?
Ask five questions about a kind of email. Only if the answer to all five is yes is it a candidate.
- Is the answer the same for everyone who asks? “What are your opening hours?” qualifies. “Why was I charged twice?” does not.
- Does the answer come from one source you keep current? A help article or a page on your website, not memory.
- Is nothing committed? No money, no dates you do not control, no exceptions to policy.
- Is a wrong answer cheap to correct? One follow-up email fixes it, and nobody is harmed in between.
- Have the drafts proven it? For this kind of email, you have sent the drafts unchanged week after week.
Typical candidates for a small business: opening hours, where to find an invoice, how to reset a password, whether you ship to a given country, and “did you get my email?” messages.
Which emails should always stay with a person?
Some emails need judgement, carry risk, or matter too much to the relationship. Keep a written list, and keep these out of any auto-send rule:
- Refunds, cancellations, disputes and anything about money owed
- Complaints, and any email where the customer is upset
- Legal, privacy or data requests, including requests to delete personal data
- Questions about a specific order, account or invoice that need a lookup
- Anything from a large or long-standing customer
- Press, partnership and supplier emails
- Anything the tool has not seen before
A tool that sends on its own should also skip anything that does not match your rules instead of guessing.
How do you know a kind of email is ready?
Count, per kind of email, how many drafts you sent unchanged and how many you edited. Set your own bar in advance, for example: every draft of this kind sent unchanged for four weeks in a row, over enough emails that you would have noticed a problem.
Read the edits, not only the count. A small wording change is a different signal from a corrected fact. One corrected fact in a kind of email means it is not ready yet, however good the rest looks.
Keep checking after you switch it on. A price change, a new policy or a new product can make yesterday’s correct answer wrong today.
How do you run an auto-send pilot safely?
Treat it as an experiment with an exit.
- Pick one kind of email that passed the five questions above.
- Write a narrow rule: a specific subject or topic, not “all general questions”.
- Keep a copy of everything sent, in a label or folder you read daily.
- Read every sent reply for the first two weeks. Correct any mistake with a follow-up email the same day.
- Decide in advance what stops the pilot: for example, one wrong fact or one upset customer.
- Make sure you can switch it off in one step, and know who does it when you are away.
Widen the rule only after the pilot is quiet, one kind of email at a time.
Should customers know a reply was sent automatically?
If a person read and sent the reply, it is that person’s reply, whatever wrote the first version. If nobody read it, being open about that is the honest choice, and it sets the right expectation: a short line saying the answer was sent automatically and that a reply to it reaches a person. Signing an unread reply with a named person’s name invites exactly the misunderstanding you want to avoid.
How does ReplyPen handle this?
ReplyPen writes Drafts in your own Gmail or Outlook thread, and today you send every one yourself. Sending on its own, for kinds of email you choose with a Rule, is planned for a later version and will be off until you switch it on. What exists now is the evidence you would need to decide: ReplyPen’s Reports show how many Drafts went out unchanged and how many you edited, week by week, and each Draft’s Check note shows which help pages and past replies the answer came from.
For how AI reply tools work overall, and what to check before choosing one, read What is an AI email reply tool and how does one work?. For the kinds of mistakes that make a draft unsafe to send, see Why AI answers sound right and still get the facts wrong.
