Is it safe to let an AI worker send email in your name?

Yes, on three conditions: the AI worker says it is AI in what it sends, routine messages go only in wording you have approved in advance, and everything else waits for a person's explicit approval. An email sent in your name is your firm's email. The rules on marketing and on company correspondence apply to it whoever, or whatever, wrote it, and the reply comes back to you.

The email is yours

When an administrator sends a reminder from the accounts mailbox, nobody thinks the administrator is the sender. The firm is. The same holds for an AI worker. The customer who receives it holds your firm to what it says. So the question is not "what might the AI write?" but "which of our emails would we let a new member of staff send without showing us first?"

What the law says about business email

These are duties as the regulator and the legislation state them, and they apply across the UK. Which of your messages they catch, and how you comply, is for you and your own advisers.

• Marketing to individuals needs consent. Under the Privacy and Electronic Communications Regulations, the Information Commission's Office (ICO) says you must not send marketing email to individuals unless they have specifically consented to it. The exception is the soft opt-in: an existing customer who bought or negotiated to buy from you may be sent marketing about similar products, if they were given a clear chance to opt out when their details were collected and are given one in every message since.

• Sole traders count as individuals. So do some partnerships. A corporate body, such as a limited company or a limited liability partnership, can be emailed without that consent, and the ICO recommends keeping a list of those who object.

• No disguise, and a way out. A marketing email must not conceal who sent it, and must give a valid contact address for opting out.

• A company names itself. A company must disclose its registered name on its business letters and on all other forms of its business correspondence. GOV.UK adds that business letters, order forms and websites show the registered number, the registered office address, where the company is registered and the fact that it is limited.

None of this is new: it already governs every email your people send. An AI worker sends the same kind of message many times, so the wording and the list are worth getting right once.

How approval works

An AI worker has a written specification of its job. For email, the specification starts from one rule: a message waits for a person's explicit approval. Your firm can choose one exception to it.

• Routine messages in wording you approved in advance. Requests, reminders and confirmations: a form outstanding, an appointment booked, an enquiry received. If you choose, you approve the wording, the intervals and who receives them, and from then on those messages go without a person approving each one. The choice is yours, and you can withdraw it.

• Everything else waits, every time. Anything written for one person's circumstances, anything that commits the firm, any complaint, and anything the worker has no approved wording for. It is drafted, shown to a named person, and sent only on their explicit approval.

Every message, sent or held, is logged with what prompted it. And the worker says it is AI in what it sends: your firm's name, your signature block, and a plain line that the message comes from an AI assistant working for you. Anyone who asks for a person is passed to one, with the thread.

What an AI worker does with a mailbox

The jobs that fit are high in volume and low in judgement: sorting a shared inbox and answering from the record where the record holds the answer and you have approved the wording, following up quotes that went quiet, as drafts for the person who owns the customer, chasing documents, confirming bookings. For first contact, the consent rules above decide who may be written to at all, which is why the list for outbound prospecting is approved by a person before a message is drafted.

What stays with a named person

• Anything that commits the firm: a price, a discount, a payment plan, a deadline, an offer, a notice.

• Anything written for one person's circumstances, advice of any kind, and any reply that disputes a fact.

• Complaints.

• Who may be sent marketing: the consent record and the do-not-contact list are yours, and a person owns them. The worker follows the list and logs every request to opt out.

• The wording itself. A change to an approved message is approved again.

What depends on your own policies and on the contract

Which mailbox the worker sends from, what the signature and footer say and what tone your firm takes with a late payer are your policies. How your approvals are recorded, how long the log is kept and what happens to the mailbox access when the work ends are for the contract. On data, our statement is one sentence: Your data is stored in the UK, and is never used to train AI models.

What to ask before it sends a single email

• Does every message say it comes from an AI?

• Which messages go without a person seeing them, and who approved that wording?

• What happens when someone replies "stop" or asks for a person?

• Can I read everything it sent last week, and why it sent it?

In your mailbox the worker would draft each message for a person's approval, or send routine reminders and confirmations in wording you approved in advance if you choose, and your people would see only the messages that need them. Is your inbox where the day goes? Talk to our agent.