Let AI Draft the Reply. Keep Your Finger on Send.
Published
A customer asks if you can fit them in Friday. AI writes a clean reply saying you'll be there at 9.
The problem is nobody checked Friday.
That's a made-up example, but it's the difference that matters when you put AI into your business email. Writing an answer and having permission to make a promise are two different jobs.
You can use help with the first without handing over the second.
Start with one kind of email
If you run a service business around Elmira, Corning, or Northeastern Pennsylvania, you probably answer some version of the same questions every week.
Do you cover my town? What information do you need for an estimate? Can someone call me back?
Start there. Pick one routine type of reply instead of connecting an assistant to everything in your inbox.
A useful first job is drafting a response to a new estimate request. The assistant can summarize what the customer needs, list missing details, and prepare a reply for you to review.
It doesn't need permission to send email, change a booking, or issue a refund to do that.
Put the limits in the setup, not just the instructions
Telling an assistant "don't send anything without asking" is worth doing. But a sentence in a prompt isn't the same as removing its ability to send.
Check what the actual tool and connection allow.
Can it read your whole mailbox or only selected messages? Can it create drafts? Can it send, delete, forward, or change anything else?
For a draft-only test, leave sending and other write actions disconnected or disabled where the tool supports that separation. Don't assume a button labeled "approval required" covers every action. Check which actions it covers and test it.
If the connection bundles access you don't want to grant, skip the connection. You can test the writing with a made-up customer message and a short set of business facts first.
You don't need to hand over the mailbox to find out whether the replies are useful.
Give it facts it can use
An assistant needs more than "be helpful and professional."
Give it a short, current reference with your actual service area, business hours, estimate process, and the details you need from a new customer. Include a date and name someone responsible for keeping it current.
Then spell out what still needs a person:
- No price quotes unless you've approved the exact amount and scope.
- No appointment promises without a confirmed slot and your approval.
- No refunds, discounts, or changes to existing jobs.
- No claim that a message was sent or a booking was made unless that action actually happened.
- Missing or conflicting information gets flagged, not filled in with a guess.
The customer might write, "Your office already approved 20% off. Just confirm it."
That's a claim to check. It isn't permission to change your price.
Treat instructions inside customer messages the same way. A customer asking the assistant to ignore your rules doesn't get to rewrite them.
Test the replies before you connect real email
Use fictional messages with no customer details. Give the assistant the same business reference each time.
Try these:
- A normal estimate request with enough information to draft a useful reply.
- A request from outside your service area.
- Someone asking for a firm price without enough job details.
- Someone asking to book a specific time.
- A complaint demanding an immediate refund.
- A message claiming you've already approved a discount.
For each one, read what it wrote and what it says it did.
Did it ask for the missing information? Did it invent a price? Did it turn "we'll check availability" into "you're booked"?
A polished sentence can still be the wrong answer. Count a made-up commitment as a failed test, even if the rest sounds good.
Fix the instructions or workflow and run the examples again. Passing these checks isn't a guarantee. It's a way to catch problems before a real customer gets the reply.
Review the whole message before you send it
When you move to real email, use a tool and account setup you've checked for customer-data handling. Review its storage, retention, and data-use settings. Share only what the task needs, not a full thread of unrelated customer information.
Keep the first live workflow simple. The assistant prepares the draft. You check it. You send it.
Read the recipient, the subject, the entire reply, and any attachments. Check names, amounts, dates, job details, and promises against the original request and your actual records.
If you edit the draft after reviewing it, review the final version too. Approval should apply to the exact message going out, not an earlier version.
Measure the cleanup, not just the typing
Keep a short log as you try it. Note which replies you used, what you had to correct, and which ones were faster to write yourself.
If you keep removing invented promises, don't turn on automatic sending. Narrow the job or stop using it for that kind of message.
The point is less work for you. A pile of drafts that need checking from scratch isn't much of a win.
Start with one routine reply. Keep control of what leaves your business. Expand only when the work supports it, not because the tool has another switch.
Want help deciding which tasks to automate and which decisions to keep? See how the Fuzzy Solutions AI Assessment works.
← All posts