Using AI to Draft Customer Emails: What to Check Before Sending

Laptop open on a small shop counter beside a paper order ledger and a coffee mug, with soft daylight from a nearby window.

An AI tool can turn three rough lines into a finished reply to a customer within seconds, and that speed is exactly why the draft deserves a slower read than a human-written one. The U.S. Small Business Administration, in its guidance on AI for small business, advises to “have another person review all AI products” before they reach anyone outside the company. Behind that short instruction sit three separate risks: the text can be wrong, the tone can miss, and the customer details typed into the tool can end up somewhere unintended.

Video: Google for Education

What AI Drafting Tools Are Used For in Customer Email

Replies, follow-ups and rewrites

The most common use is a first draft: a reply to a pricing question, a follow-up after a delivery problem, a polite refusal of a refund request outside the return window. The video above, from Google for Education, is titled “Quickly draft professional emails using Gemini in Gmail” and illustrates the use. Vendor demonstrations show what a tool can do under ideal conditions; none of the sources reviewed for this article measured how often drafts need correction, so no accuracy or time-saving figures appear here.

Where the SBA draws the line

The SBA’s page, last updated February 14, 2025, pairs its advice on human review with other cautions. It flags that outputs may raise intellectual property problems because the tools draw on web content, and it warns about phishing campaigns that may have been written with AI. The second warning matters for inbound mail as much as outbound: a customer email asking for an urgent payment change or a gift-card purchase deserves the same suspicion whether or not a person wrote it.

Where Drafts Go Wrong

Confident errors

NIST, the federal standards agency, published a companion profile for generative AI in July 2024. Its list of risks includes confabulation, described as systems that “generate and confidently present erroneous or false content in response to prompts”. In a customer email, a confabulated sentence might be an invented return window, a shipping date the shop never promised, or a discount that does not exist. The tool has no way of knowing the shop’s actual policy unless it is pasted into the prompt.

The NIST framework itself is voluntary. The agency describes AI RMF 1.0, released January 26, 2023, as a framework intended for voluntary use, with the generative AI profile arriving on July 26, 2024.

Tone that fits nobody in particular

A draft can read smoothly and still sound like a stranger wrote it. A regular customer who has bought for years notices when a reply drops the first-name familiarity, or when an apology for a late order sounds like a template. The SBA page notes that recipients may treat AI-generated messages as spam, which makes a generic voice a deliverability and relationship problem rather than only a matter of taste.

Details the tool cannot know

Order numbers, names, dates and amounts are the parts most worth checking against the actual record. The tool produces plausible text; the order system holds the facts.

Hands marking a printed email draft with a red pen on a wooden desk next to a calculator and paper receipts.

What Four Vendors Say About Customer Data

The privacy terms differ by product tier, and one brand can have different rules for its business and personal products. Each statement below reflects the vendor page as fetched on November 8, 2026, and the pages change.

OpenAI

OpenAI’s help page on how data improves model performance says that content from individual services such as ChatGPT may be used for training unless the user opts out. Content from ChatGPT Team, ChatGPT Enterprise and the API Platform is not used for training, according to the same page. The page carries no specific date, only a relative “updated” note. A separate Data Controls FAQ describes the “Improve the model for everyone” setting, which stops new conversations from being used for training when turned off, and Temporary Chats, which are deleted from OpenAI’s systems within 30 days.

Anthropic

Anthropic’s privacy center page for commercial products, dated August 18, 2026, states that inputs and outputs from commercial products are not used for training by default. The exception is feedback: chats submitted through a thumbs up or thumbs down button may be used. A separate article, dated March 16, 2026, covers consumer plans, where training use depends on a setting the user chooses.

Google

Google’s Generative AI in Google Workspace Privacy Hub, last updated October 6, 2026, says that Workspace does not use customer data to train models without the customer’s prior permission. The same page says Gemini retrieves only content that a user is already permitted to see.

Microsoft

Microsoft’s Copilot data page, showing a last-updated date of November 3, 2026, states that prompts, responses and data accessed through Microsoft Graph aren’t used to train foundation LLMs. It also says Copilot surfaces only organizational data the individual user has at least view permission for, and that administrators are responsible for getting those permissions right in services such as SharePoint.

What the pattern means for a customer list

Business and workspace tiers commit, on the pages above, to not training on customer content by default. Free and personal accounts follow different rules, and a staff member pasting a customer’s name, address and complaint into a personal account falls under the personal-account terms. The SBA’s instruction is blunt on this point: “Try not to feed any sensitive data or proprietary information.” A draft can usually be written with a placeholder such as “[CUSTOMER]” and the real details added afterward in the email program.

Terms Change, and the FTC Has Said Why That Matters

The FTC’s Office of Technology addressed the data question in a January 2024 blog post. It describes AI model companies as having a “continuous appetite for data”, which can pull against promises made to customers. The post says that quietly changing terms of service, or “burying a disclosure behind hyperlinks, in legalese, or in fine print”, can draw FTC enforcement when it results in customer data being used for new purposes without clear notice and consent.

That guidance is aimed at the AI companies, not at the shops using their products. For a shop, the practical consequence is dated: a data setting confirmed in one month is not guaranteed the next. Re-reading the vendor’s data page when a plan is renewed, or when a new product feature is switched on, keeps the file current.

Back office with a locked filing cabinet and a monitor showing a blurred spreadsheet under a desk lamp.

Marketing Email Written by AI Is Still Marketing Email

The FTC’s CAN-SPAM compliance guide for business, originally published in August 2023 and modified in July 2025, applies to all commercial email, with no exception for business-to-business messages. A promotional email drafted by AI has to carry the same elements as one typed by hand. The guide lists accurate header information, a subject line that reflects the content, clear disclosure that the message is an advertisement, a valid physical postal address, and a clear way to opt out. Opt-out requests must be honored within 10 business days.

Penalties are stated per message. The page gives a figure of up to $53,088 for each email in violation, and it adds that hiring another company to send the email does not transfer legal responsibility.

A subject line written to maximize opens can slide into misleading territory, whether a person or a tool wrote it. The compliance question is whether the line matches what the email contains.

Transactional and relationship messages, such as an order confirmation, are treated differently, but the guide says those exemptions are read narrowly. A message that starts as a shipping update and ends with a sale on new stock risks being classed as commercial.

Two people reviewing a printed page together at a restaurant host stand with a pen and a reservation book beside them.

A Check Before Sending

Facts

Every number, date, policy statement and promise in the draft gets compared against the shop’s own records: the order system, the posted return policy, the price list. Anything the tool added that was not in the prompt is treated as unverified until matched.

Data

The draft was written in an account covered by business-tier terms, or the real customer details were left out of the prompt. A quick look at the vendor’s current data page, with the date noted, answers the first question.

Voice and fit

The email reads the way the shop writes to this particular customer. A second person reads anything sensitive, such as a complaint, a refund denial or a response to a legal threat, because the SBA recommends review by another person for all AI output.

Compliance

If the message promotes a product or sale, the CAN-SPAM elements are present and working: sender information, a truthful subject line, a postal address and an unsubscribe method that has been tested.

Related posts on this site cover the groundwork for the email itself, including writing a welcome email for new subscribers, collecting email addresses at a counter and three free email marketing tools.

For informational purposes only — consult a licensed attorney, tax professional or financial advisor.

Sources