“We received your artwork” and “Your order will be ready on Friday” may appear in the same customer email. One reports a fact. The other commits the business to doing something.
AI can help write either sentence. Your business still needs to decide when each sentence is allowed to leave the building.
The question matters even when a person reviews every draft. A reviewer needs to know whether they are checking the wording or authorizing a price, date or exception. Those decisions should not hide inside a routine request to approve an email.
Look for the promise inside the message
Consider an illustrative print shop using AI to prepare customer replies. A customer asks for five hundred brochures by Friday. The shop has the artwork, a price from last month and a production schedule that changes throughout the day.
A fluent reply could combine those facts into a confident confirmation. But an old price does not establish the current price, and space on the schedule this morning does not mean the job is now booked. Before confirming Friday, check that production has accepted the job for Friday.
Read a sample of your ordinary customer messages and underline every promise. Look beyond obvious guarantees. “We can include that,” “there is no extra charge,” “you can cancel later” and “someone will call today” all create expectations about the business’s next action.
For each type of promise, ask what has to be true before it can be made, and which record shows it is true now.
Give the team a permission table
The print shop could start with the following rules.
| Message content | What authorizes it | What happens if that evidence is missing |
|---|---|---|
| Confirm receipt of artwork | The correct files are present on the customer’s job | Ask an employee to check the files |
| State a price, discount or free extra | An approved, unexpired quote for this exact quantity and specification, including any discount or extra a manager has approved | Send the price question to the estimator and any discount request to a manager |
| Confirm a collection date | Production has accepted the job and recorded that date | Request a scheduling decision; do not present a requested date as confirmed |
| Promise a replacement for faulty work | The service lead has reviewed the job and approved the remedy | Acknowledge the concern and route it for a decision |
The table separates existing decisions from new decisions. Once production has accepted a job and recorded Friday as its collection date, a reply can communicate that decision. If the date is merely the customer’s preference, someone still has to decide whether the shop can deliver it.
These rules apply whether a person sends each reply or a tool sends some on its own. Before relying on a tool, give it a request for a date production has not accepted and check that it does not confirm the date. A written instruction alone does not prove a system will respect the limit.
Treat a changed request as a new decision
The awkward cases often arrive as small edits. The customer replies, “Same as before, except please use the heavier paper.” The previous approval may no longer cover the job, because the new paper could change the cost, availability or production time.
In this example, the shop’s rule is simple: a change to quantity, stock or finishing returns the job for price and scheduling confirmation. The next reply can acknowledge the change and explain that the shop is checking the effect. It cannot reuse Friday’s promise merely because Friday appeared in an earlier message.
The same principle applies when an AI rewrites an approved sentence. Changing “We are checking whether Friday is possible” to “We will have this ready Friday” changes the decision, not just the tone. Reviewers should check the meaning of dates, prices and conditions even when the prose looks smoother.
Make the exception easy to handle
If an approval request contains only “Can we do this?”, the manager has to reopen the conversation and reconstruct the job. A useful request gives the decision-maker the exact promise under consideration and the facts that affect it.
For the heavier-paper request, the internal note might say:
The customer wants five hundred brochures on heavier paper, still for Friday. The previous price and production approval covered standard paper. Please confirm the revised price and collection date before we reply.
The person deciding can approve the new terms, offer an alternative or decline.
Do not put the owner at the end of every exception. Assign the decisions to the people who can make them, within set limits. If only the owner can decide a particular exception, make that narrow dependency visible rather than disguising it as routine email review.
Before expanding AI-assisted customer communication, take ten recent messages and mark the commitments they contain. Use those examples to fill in your first permission table. You will learn more about what to control than you would from checking ten messages for spelling and tone.
