Should AI send your customer emails?

Not at first, and for most kinds of message it's better to keep a person approving them: let AI write the draft, and have someone click send. Once a narrow, low-risk type of email, like an appointment reminder, has a proven track record, it can reasonably go out on its own.

AI is genuinely good at writing customer emails, and having it draft them saves real time. The question of letting it send them unchecked is worth thinking through carefully, because the reasons for caution are specific rather than general nervousness.

Drafting and sending are two very different jobs

"Automated email" hides a big decision inside one word. If AI drafts, the worst that happens is a bad draft, which somebody deletes in two seconds. If AI sends, the worst that happens is already in your customer's inbox, where it can be quoted, forwarded and kept.

The difference in time saved is smaller than people expect. Nearly all the work is in writing the email, and the approval click takes seconds. What you gain by removing that click is a few minutes of speed, not hours of work. The difference in risk is much bigger.

If you'd like AI drafting replies inside the email system you already use, with a person approving each one before it goes, we're happy to set that up for you.

Four things that go wrong, none of them unusual

A wrong price. AI will quote a figure with total confidence whether or not it came from your price list. If it picked one up from an old email thread, that's the figure your customer now has in writing.

A promise you can't keep. "We can have someone with you on Thursday." Nobody checked Thursday.

The wrong tone for a complaint. AI tends to write in a brisk, cheerful, helpful style. That's the wrong note for someone who's already upset, and it can come across as dismissive, which turns a complaint you could have resolved into a bad review.

A detail that isn't true. This one's easy to underestimate. AI doesn't know when it's guessing, so it doesn't warn you. A believable-sounding warranty term or delivery time arrives in your usual style, from your email address, with your name on it.

A person reading the draft catches all four in seconds. That's the whole case for keeping an approval step.

A tribunal has already decided who's responsible for what your system says

The case is Moffatt v. Air Canada, 2024 BCCRT 149, decided on 14 February 2024 by the Civil Resolution Tribunal in British Columbia. It's a Canadian small claims decision, not a court judgment, and it doesn't bind anyone in the UK. But its reasoning is easy to follow.

In November 2022 Jake Moffatt's grandmother died. Moffatt used the chatbot on Air Canada's website, which said that anyone who had already travelled could submit their ticket for a reduced bereavement fare within 90 days. Air Canada's actual policy page, linked from that very message, said the opposite: the policy didn't apply to requests made after travel was completed. Moffatt booked flights costing $794.98 and $845.38, applied within the window, and was told retroactive applications weren't allowed.

The decision records Air Canada's argument that it "cannot be held liable for information provided by one of its agents, servants, or representatives", and that this included the chatbot. Tribunal member Christopher C. Rivers didn't accept it: "In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website."

He also rejected the idea that customers should have to check one part of a website against another, and ordered Air Canada to pay $812.02 in Canadian dollars: $650.88 in damages, $36.14 in interest and $125 in fees.

Two points to be accurate about. First, it was decided as negligent misrepresentation, not as a breach of contract. In plain terms, the chatbot's words weren't part of the ticket terms, but Air Canada hadn't taken reasonable care that what it told customers was accurate. Second, the decision records that Air Canada "did not provide any information about the nature of its chatbot", so it doesn't tell you whether it was the newer kind of AI that writes fresh text. If anything, that makes the lesson stronger, because a system writing new wording every time is harder to keep accurate than one reading from a script.

The closest UK equivalent is section 2(1) of the Misrepresentation Act 1967, which applies in England and Wales. Where someone enters a contract after being told something untrue and loses out as a result, the person who said it is liable for damages unless they can prove they "had reasonable ground to believe and did believe up to the time the contract was made the facts represented were true". That defence depends on what you believed and why, and it's harder to rely on when nobody at your business read the email before it went out. Where the customer is a consumer with a right to redress under consumer protection law, section 2(4) sends them down that route instead of this one.

If the email is marketing, a second set of rules applies

PECR, the Privacy and Electronic Communications Regulations, are the UK rules on marketing by email, text and phone. They define direct marketing as "the communication (by whatever means) of advertising or marketing material which is directed to particular individuals". Routine service messages, such as appointment reminders, a failed payment or a change to your terms, aren't marketing, and you can send them.

Then comes the sentence that should shape how you set things up: "If your service message has elements that are direct marketing, even if that is not the main purpose of your message, then it will count as direct marketing." The ICO's own example is a gym ringing about a failed payment and mentioning its personal training sessions. The whole call becomes marketing.

AI tends to be friendly and helpful, especially at the end of an email. A cheerful last line offering ten per cent off, or suggesting the customer books their next service, is exactly the kind of thing it adds without being asked. At that point your service message has turned into marketing, and nobody read it.

That matters because the rules for marketing emails are strict when you're writing to individuals: you need their consent, or every condition of what's called a "soft opt-in" met, which is a narrow exception with several conditions set out on that ICO page. Emails to companies, which the rules call corporate subscribers, don't need either, though every marketing email must still say who it's from and give a valid address to opt out.

"The tool did it" isn't a defence either. PECR "applies to anyone who sends or instigates the sending of direct marketing by electronic mail", and the ICO is clear that using someone else's sending platform "does not normally make that service responsible. You remain responsible for complying with PECR."

The fines are set out in law. For anything sent on or after 5 February 2026, Schedule 1 to PECR and section 157 of the Data Protection Act 2018 put breaches of the marketing email, text and call rules, and of the cookie rules, in the higher level: up to £17.5 million or 4% of worldwide annual turnover, whichever is higher. Breaches of PECR rules not on that list are capped at £8.7 million or 2%. Anything sent before that date still falls under the old £500,000 cap. (The ICO's guide to PECR still quotes the old £500,000 figure, so don't be thrown if you see it there.)

Letting AI send on its own, one step at a time

Step one: everything is drafted, nothing is sent automatically. Run it like that for a few weeks and treat the drafts as evidence. How many would you have sent without changes? Which kinds of email are reliably good, and which are wrong in the same way every time? Nobody can tell you that in advance, because it depends on your own information and how tidy it is.

Step two: check a sample, one kind of email at a time. Once one narrow kind of email has a good track record, you can stop reading every one and read a sample instead. Anything that mentions a price, a date, a promise or a refund, sounds like a complaint, or is about a named person still comes back to a human. The system flags the things that could hurt, not just the things it's unsure about.

Step three: sending on its own, but only where a mistake is cheap. Delivery updates. Appointment reminders. "Thanks, we've got your enquiry and someone will be in touch." Messages where putting a mistake right takes one apology and no money.

The number to think about isn't accuracy, it's the cost of the worst email. A system that's 99% right and sends two hundred emails a week sends two wrong ones. For reminders, that's fine. For quotes, it isn't.

If you'd like help setting up those checks, so that emails mentioning prices, dates or complaints always come back to a person, we're happy to build that automation for you. You can see more about what we set up for small businesses on our main page.

What we'd suggest

Start with AI drafting, keep it that way for at least a month, and actually read the drafts rather than skimming and clicking. Then let one routine kind of email go out on its own and see how it goes. The approval step costs less time than people expect, and keeping it for most kinds of email is a perfectly good result, not a failure.

If anyone tells you it's safe to let AI send everything from day one, it's worth asking who would be responsible for putting it right.

Sources

• Moffatt v. Air Canada, 2024 BCCRT 149 - Civil Resolution Tribunal, British Columbia
• Misrepresentation Act 1967, section 2 - legislation.gov.uk
• Identify direct marketing - ICO
• How do we comply with the PECR electronic mail marketing rules? - ICO
• What are PECR? - ICO
• Privacy and Electronic Communications (EC Directive) Regulations 2003, Schedule 1 - legislation.gov.uk
• Data Protection Act 2018, section 157 - legislation.gov.uk
• The Data (Use and Access) Act 2025 (Commencement No. 6 and Transitional and Saving Provisions) Regulations 2026, regulation 11 - legislation.gov.uk

Checked on 30 October 2026. The ICO's guidance is still catching up with the Data (Use and Access) Act 2025, so check the ICO links before relying on them.