An AI tool can write a polite, well-structured reply to a customer in a few seconds. That is the appeal, and also the risk. The text sounds sure of itself whether it is right or wrong, and the customer will not know or care that a machine wrote it. If the reply promises a refund, quotes a wrong delivery date or misreads the question, it is your company that said it.
This article is for an owner or manager of a small company whose staff already use, or are about to use, AI to draft emails, chat answers or support replies. It describes a review that takes a minute or two per message and can be used from tomorrow.
Decide which replies need a review
Not every message carries the same risk. A reply that says "thank you, we received your order" is low risk. A reply about price, a complaint, a contract, a deadline or someone's personal data is not.
A simple rule for a small team:
- Always review anything that mentions money, dates, legal terms, personal data or a complaint.
- Review a sample of routine answers, for example every tenth one, to see whether quality holds.
- Do not let AI send alone to customers until you have weeks of reviewed examples showing that it is safe for that type of message.
If you are still deciding where AI belongs in the first place, our article on choosing a first AI use case is a good starting point.
Step 1: Check facts and numbers
AI tools can invent details that look entirely plausible: a product feature you do not have, an opening time, a price, a name. Check every concrete statement against your own records, not against what the AI says about itself.
A short list to run through:
- Names and addresses. Is it the right customer, product and order?
- Numbers. Prices, quantities, dates and totals. Recalculate anything that was calculated.
- Policies. Does the stated return period, warranty or delivery rule match what you actually offer?
- Links and attachments. Do they open, and are they the right ones?
- Did it answer the question? Compare the reply with the customer's message, not only with itself.
If the AI states something you cannot confirm quickly, remove it or replace it with "we will confirm this by tomorrow". A missing detail is better than a wrong one.
Step 2: Check tone and promises
A reply can be factually correct and still cause trouble. Read it once as the customer would.
- Commitments. Look for words such as "we will", "guaranteed", "always" and "no problem". Each one is a promise. Keep only those you can fulfil.
- Apologies and blame. AI tends to apologise generously or, less often, to sound defensive. Make sure the reply admits only what you are prepared to admit.
- Tone. Does it sound like your company? A reply that is too formal, too cheerful or full of stock phrases weakens trust.
- Sensitive situations. For an angry customer, a bereavement or a health matter, write it yourself or rewrite the draft thoroughly.
- Personal data. Make sure the reply does not repeat or reveal more than the customer needs. If you paste customer details into an AI tool, your company's rules on using AI tools should say what is allowed.
Step 3: Name who approves and keep a record
Without an owner, "someone will check it" often means nobody does.
- Name the approver for each type of reply. For example: the person who wrote the prompt approves routine answers; the manager approves anything about money or complaints.
- Say what approval means. The approver has read the full text, checked the facts and accepts responsibility for sending it.
- Keep a short error log. A shared sheet with four columns is enough: date, what the AI got wrong, who caught it and what changed afterwards.
Once a month, look at the log for patterns. If the same mistake appears three times, change the instruction you give the AI, add the correct fact to your template or move that type of reply back to manual writing. The log also tells you honestly whether the review is still needed at its current level.
A minute-long example
A customer asks whether an order placed today will arrive before Friday. The AI drafts: "Yes, your order will arrive on Thursday, and shipping is free." The reviewer checks the order system: dispatch is tomorrow, delivery usually takes two working days and free shipping applies only above a set order value. The reviewer edits the reply to say that the order is expected on Friday or Monday and states the actual shipping cost. The mistake goes into the log with a note to add the delivery times to the template.
Common mistakes
- Reading only the first lines and trusting the rest
- Checking the style but not the facts
- Letting the person who wrote the prompt approve their own high-risk replies without a second look
- Keeping no record, so the same error returns
- Removing the review after a few good weeks without looking at the log
Checklist
- A list of reply types that always need review
- Facts, numbers and policies compared with your own records
- Every promise and apology read on purpose
- A named approver for each type of reply
- A shared error log, read once a month
- A rule that AI does not send to customers alone
Sources
- Practical guidance based on common quality-control practice in customer communication, offered as a planning aid. The review steps and thresholds are starting points to adapt to your own company. No client results or market figures are claimed in this article.