Gmail and Outlook now both offer to write your emails for you. If you answer customers every day, that sounds like a gift. It can be. But there is a big difference between AI that writes a draft and AI that sends it.
Below: what AI does well with your email, why it shouldn't send customer replies by itself, what to check before you send, and what to ask about any tool before you let it read your mailbox.
On this page
The short answer
The short answerLet AI help with your email, but don't let it send to customers on its own:
- Let it read, summarise and write drafts. That saves the typing.
- Keep the send button for yourself. AI can get a price, a date or a promise wrong without noticing.
- Before you send, check amounts, dates, promises and tone.
- Before you start, check where your email goes and whether it's used to train models.
1What AI can do with your email
Most AI help for email comes down to four jobs:
- Summarise. A long thread in a few lines, so you know where things stand.
- Find what is being asked. Customers often hide two or three questions in one email.
- Write a draft. A first version of the reply, based on the email and your instructions.
- Rewrite. Shorter, friendlier or more formal, from text you already have.
The big two have this built in. In Gmail, Help me write (part of Gemini) sits in the compose window: you type what you want, click Create, and it writes a draft you can refine with options like Formalize, Friendly and Shorten. You need an eligible Google Workspace or Google AI plan; for personal Google accounts, Google says it is only available in the US. In Outlook, you select the Copilot icon in a new email, choose Draft and describe what you want. Microsoft's own steps end with: "Edit the draft as needed, and then select Send."
In both cases, the AI writes and you send. That is a good split, and the rest of this article explains why.
2Drafting for you or sending for you
"AI that answers your email" can mean two very different things. Either it prepares a reply and waits for you, or it sends the reply without anyone reading it first.
| AI drafts, you send | AI sends by itself | |
|---|---|---|
| Who reads it before the customer | You | Nobody |
| A wrong price or date | You catch it | The customer gets it |
| An email with hidden instructions | You see the draft first | It may act on them |
| What you save | The typing | The typing and the reading |
Sending by itself only saves you the last minute: reading the draft. That minute is exactly where the mistakes get caught.
3Why AI shouldn't send customer replies by itself
It can be wrong and sound sure. A language model writes what sounds likely. If your price or your diary isn't in front of it, it may still write a price or a day. Google's own help page for Gemini in Gmail warns that its features "may suggest inaccurate or inappropriate information". In October 2025 the Dutch consumer authority (ACM) and data protection authority (AP) wrote that chatbots in customer service sometimes give poor or even incorrect answers.
What you send, you have said. To your customer, a reply from your address is your answer. If it says "Thursday works" or "same price as last time", they will hold you to it, whoever typed it.
An email can carry hidden instructions. Anyone can email you, and an email can contain text meant for the AI rather than for you. Security experts call this prompt injection, and OWASP puts it first in its 2025 list of risks for AI applications. In February 2026 the AP warned against AI agents that get full access to your computer, including email, because they are vulnerable to hidden commands in websites, emails and chat messages, and can act without your approval first.
It doesn't know what you'd decide. Can you fit the job in? Will you give the discount? Do you want this customer back? Those are your decisions. An AI can only guess them.
When automatic is fine
Not every automatic email is a problem. If a reply needs no decision and you wrote the text yourself in advance, sending it automatically is fine. An out of office message is the classic example: same text for everyone, written by you, saying when you'll be back. In Gmail you set it up in the settings, see how to set out of office in Gmail.
That isn't AI answering your email. It's a fixed text you wrote once. As soon as a reply contains a price, a date, a yes or a no, a person should read it first.
What to check before you send an AI draft
Reading a draft takes less than a minute if you know what to look for:
- Amounts. Is every price, rate and total one you actually charge?
- Dates and times. Are you really free then? Is the day right for the date?
- Promises. "I'll send it tomorrow", "no extra charge": would you have said that?
- The customer. Right name, right address, right job.
- Every question. Did it answer all of them, and nothing they didn't ask?
- Your voice. Would a regular customer recognise you in it?
A simple rule: if a fact in the draft didn't come from you, check it or take it out. And watch for the stiff phrases AI likes to use:
I hope this email finds you well.
I would be delighted to accommodate your request.
Our standard rates apply.
Thanks for your email.
Thursday works for me.
It's £45, same as last time.
What to ask about the tool itself
An AI tool for email reads your customers' messages: names, addresses, prices, sometimes worse. Before you connect your mailbox or paste an email into a chatbot, find the answers to these:
- Is my email used to train models? For business accounts, Google says Workspace doesn't use customer data to train models without the customer's permission. Free chatbots and personal accounts can have different rules, so check yours.
- Where is it stored, and for how long? Look for the country and the number of days in the privacy policy, not just "secure".
- Who can read it? Staff at the provider, other people on your team, or nobody?
- What may it do in my mailbox? Read only, draft, or also send and delete?
This matters more than it seems. In August 2024 the AP reported several data breaches where employees had put personal data of patients or customers into an AI chatbot, and noted that most companies behind chatbots store everything you type in.
A better split: you decide, AI writes
Most of the risk disappears when you make the decisions and the AI only does the writing. Instead of "reply to this email", tell it what to say: "Yes to Thursday, the price is £45, no weekend work." Now there is nothing left for it to guess. It only turns your answers into a friendly email.
This works in any chatbot too. If that's how you work, see how to write emails with ChatGPT for prompts that give the AI your decisions up front.
That is how Soonbeam is built. It reads each customer email and pulls out the questions. You tap an answer, type a few words or say it out loud, and Soonbeam writes the reply in your style, using your prices and way of working from your knowledge base. Every answer shows where it came from. If Soonbeam isn't sure about an amount or a date, it asks you instead of filling it in, and nothing goes out until you send it.
To read email and find the questions, Soonbeam uses language models from Mistral, a French company. To write replies, it uses Claude from Anthropic, an American company. Training is switched off at both, so your email is never used to train a language model. Soonbeam costs €19 per user per month, excluding VAT, and is invitation-only for now.
Questions
Can AI answer my emails automatically?
Some AI assistants can act in your mailbox without asking you first. For customer email that is a bad trade: a wrong price, date or promise goes straight to the customer. Let AI write the draft, and send it yourself.
Is it safe to let AI read my email?
It depends on the tool. Check where your email is stored, whether it is used to train models, how long the provider keeps it and what the tool is allowed to do in your mailbox. Don't paste customer emails into a free chatbot without checking this first.
Do Gmail and Outlook have AI that writes replies?
Yes. Gmail has Help me write, powered by Gemini, and Outlook has Copilot, which can draft a message for you. Both depend on your plan or account type, and in both you review the draft and send it yourself.
What should I check before sending an AI draft?
Amounts, dates and times, anything you promise, names and details about the customer, whether every question is answered, and whether it sounds like you. If a fact in the draft didn't come from you, check it or take it out.
- Google: Draft emails with Gemini in Gmail
- Google: Generative AI in Google Workspace Privacy Hub
- Microsoft: Draft an email message with Copilot in Outlook
- OWASP: LLM01:2025 Prompt Injection
- AP and ACM: chatbots should not fully replace humans in customer service (2 October 2025)
- AP warns of major security risks with AI agents like OpenClaw (12 February 2026)
- AP: Caution, use of AI chatbot may lead to data breaches (6 August 2024)
Prices, features and menu names of other companies come from their own websites and were correct on the date we checked them. They can change at any time, so check the company's own site before you decide. Soonbeam is not affiliated with these companies; names and trademarks belong to their owners.































