How-to · Research & support
How to set up AI help for customer email
A careful rollout for using AI on customer email: write the answers first, limit what the tool can send, and keep a person on anything that costs money.
On this page
Customer email is where a confident wrong answer becomes a refund, a review, or a regulator’s screenshot. You can still use AI here. You use it as a draft on top of answers you wrote, not as a personality with your logo. Last reviewed October 2026.
Write ten answers before you connect a tool
Open last month’s inbox. Copy the ten questions that showed up more than once. Write the answer you want sent, including the conditions: the return window, the country you ship to, the thing you will not promise. Put those answers in a doc your team can edit. That doc is the help center, even if it is not a help center product yet.
If you cannot write the answer, the tool cannot send it. “It depends” is a fine answer when you say what it depends on.
Choose a tool that matches the inbox you have
A small site can start with Tidio or Crisp, which put chat and a shared inbox on the site and publish free plans. A team that already lives in tickets can look at Intercom or Zendesk, which are paid suites. The support shortlist explains that split. Do not buy the suite to avoid writing the ten answers.
Turn on drafts, not autonomy
For the first two weeks, the tool may suggest a reply. A person sends it. Configure whatever limit the product offers so the suggestion can only draw on your ten answers. If the product cannot be limited, do not enable auto-send. Use the suggestion box and paste nothing you did not check.
Each day, mark suggestions as kept, edited, or rejected. If you are editing every one, the answers are wrong or the tool is ignoring them. Fix the answers. Do not “give it more personality.”
Never automate these on week one
- Refunds and exceptions to a written policy.
- Anything about legal, medical, or account security.
- A message that would admit fault.
- A message to someone who is angry. Speed is not the goal on that thread.
A human can still use a draft to get the tone calmer. The human sends it.
Tell the truth about who is replying
Do not sign a bot with a colleague’s name and a stock photo. If a message is automated, say it is automated and show how to reach a person. Office hours belong in the first reply, not in a footer nobody sees after the bot has looped.
Review on a Friday
Once a week, read a sample of threads. Update the ten answers when the product or the policy changed. Remove any automation that sent a sentence you would not say on the phone. That Friday review is the operating system. The software is a participant, not the owner.
When the ten answers have been stable and mostly sent unchanged, you can consider letting the tool send those ten and only those ten. Expand one question at a time. The privacy policy of your own business, not Umee’s, is what you owe the customer about their email. Read your tool’s data terms before you connect the inbox.
Questions this guide answers
Should the AI send email by itself?
Not on the first week, and not for refunds, legal questions, or anything you have not written an approved answer for.
Which inbox should I start with?
The one you already answer. A lighter widget such as Tidio or Crisp is enough for a small site. A suite is for a team that is already losing track of tickets.
What do I tell customers?
If a reply was drafted with software, you do not have to write an essay. You do have to make sure the reply is true, and you should not pretend a bot is a named employee.