AI Email Assistant: What to Hand Over and What to Keep

Most AI-written emails don’t fail on grammar. They fail because the model has to guess. Type “write a polite reply turning down the supplier” and you get three paragraphs of courtesy, no substance, and at least one sentence you would never say out loud.

That is fixable with a template, two minutes of prep and one rule about what you never delegate.

Which emails are worth delegating

The rule of thumb is simple. The more a reply is built from facts you already have, and the less it negotiates, the better a fit it is.

Good candidates: scheduling back-and-forth, chasing an unpaid invoice for the first time, summarising a 40-message thread before a call, polite declines, and answers to questions customers ask every week. Poor candidates: quotes with prices, contract terminations, complaints with legal weight, anything involving HR.

There is one category where AI helps more than people expect. The awkward email you’ve been avoiding for three days. Getting a first draft on screen lowers the barrier, even if you rewrite half of it.

Criterion	AI draft works	Write it yourself
Scheduling, quick follow-ups
Summarising long threads
First payment reminder		partly
Quote with prices	partly
Termination, disciplinary, legal

Our own assessment. "Partly" means the AI provides structure, you provide numbers and commitments.

The prompt template that turns a draft into an email

Every useful email prompt answers four questions. Who is writing to whom, what happened, what should the email achieve, and how should it sound. Then add a word limit, because models drift long.

Here is a reusable version:

You are drafting an email on my behalf.
Sender: me, owner of a small bookkeeping practice in Leeds
Recipient: Priya, client since 2022, we're on first-name terms
Situation: She asked for her year-end figures by Friday. I need until
Tuesday because her bank hasn't sent the March statement yet.
Goal: agree a new deadline and ask her to chase the bank.
Tone: friendly, direct. No warm-up sentence.
Length: 80 words maximum.
Give me two versions: one neutral, one slightly warmer.

The “Situation” line does the heavy lifting. It holds the facts the model cannot know. Leave it vague and the assistant invents a plausible reason for the delay, which then sits in your sent folder with your name under it.

Two versions are better than one. You spot the right tone faster, and you’ll often take the opening from one and the close from the other.

Teaching it your voice

The default AI email opens with “I hope you’re doing well” and signs off with “Please don’t hesitate to reach out.” Nobody actually writes like that.

The fix is a style sample. Paste three real emails you’re happy with and add: “This is how I write. Match the sentence length, greeting and sign-off. Ignore the content.” If you use a custom assistant with saved instructions, store the sample there once. Our piece on setting up your own assistant with fixed instructions shows how that works in the main tools.

Banning specific phrases helps too, but Anthropic’s prompting guide makes a better point: tell the model what to do rather than what to avoid. “Open with the reason for the email” beats “don’t be wordy.”

Outlook, Gmail or a chat window?

There are three routes. The real difference is how much of your mailbox the AI can see.

Microsoft lists three core Copilot functions in Outlook: drafting emails, summarising threads and coaching on clarity and tone. What you get depends on the licence. The full add-on reads your mail plus calendar entries and files automatically; the basic tier works only with content that is open or uploaded. Microsoft’s Copilot overview breaks down the tiers.

Gmail offers “Help me write” for drafts plus AI-suggested replies. According to Google’s help page, the personalised version that pulls details from earlier emails needs an eligible Workspace or Google AI plan and is English-only for now.

The third route is a standalone chat assistant such as ChatGPT or Claude. You copy text across manually. Clunkier, yes, but it works with any mail client and gives you full control over the prompt. For freelancers without a Microsoft 365 business licence it is often the practical answer. Our ChatGPT vs Claude comparison covers which suits what.

What never goes out unchecked

Here’s the rule we’d recommend: every number, date and commitment in an AI draft gets checked against the source. No exceptions.

Language models fill gaps with plausible content. If the thread said “end of the month”, the draft may say “by the 30th” in a month with 31 days. Small errors, but they land in a client’s inbox. The mechanics behind this are covered in why AI tools make things up.

• Names, titles and level of formality correct?
• Every figure and date compared with the original?
• No promise you didn't mean to make?
• Stock opening and closing lines deleted?
• No client data pasted into a personal AI account?
• Would you sign this exactly as written?

Privacy: what you can paste into a prompt

Forwarded client emails nearly always contain personal data: names, email addresses, often phone numbers or order details. Under the GDPR (and the UK GDPR), pasting that into a personal AI account means passing it to a third party without a clear basis.

Two habits cover most of the risk. Swap names and contact details for placeholders like [client] or [company] before pasting; the model doesn’t need them to write a good reply. And if client data is a daily thing, move to a business plan with a data processing agreement where inputs aren’t used for training by default. Our guide to ChatGPT and the GDPR explains what that means in practice.

How much time it actually saves

There is little reliable data on time saved per email. Vendor figures are marketing. You can do an honest estimate yourself, though.

Writing it yourself, procrastination included

20 min

Filling in the template

3 min

Reading two versions, merging, checking

5 min

Saved per email

about 12 min

Assumptions, not measurements. Plug in your own times.

For routine replies you’d type in two minutes anyway, the detour through an AI barely pays. Saved templates are faster: Gmail has them under advanced settings, Outlook has “My Templates”. Ten standard replies written once will cover a surprising share of your inbox.

The maths shifts when a message needs thought. Tricky tone, an angry customer, saying no to someone you like. That is where a draft saves the time you’d otherwise spend staring at a blank reply window.

When the inbox outgrows copy and paste

Some businesses get dozens of near-identical enquiries a day. At that point you no longer need a writing aid; you need a system that sorts incoming mail, proposes answers from a knowledge base and leaves only the approval to a person. Off-the-shelf tools get you part of the way. For the rest, a studio like OnlineBuilders builds assistants that plug into a company’s own systems.

For everyone else, the next step is smaller. Take the email that has sat longest in your inbox, put the facts into the template above and ask for two versions. More tasks worth handing over are collected under everyday automation.

Questions people ask

Can people tell when an email was written by AI?

Often, yes. Long warm-up sentences, stock phrases like "I hope this email finds you well" and suspiciously tidy structure give it away. Feeding the assistant samples of your own writing and trimming the draft removes most of the tells.

Is it safe to paste client emails into ChatGPT?

Not by default. Names, addresses and order details are personal data, and in the EU and UK that brings GDPR into play. A business plan with a data processing agreement is far easier to justify than a free personal account. Replacing names with placeholders is the quick fix.

Can an AI answer my inbox automatically?

Add-on tools can do it, but it only makes sense for narrow cases like acknowledging receipt. Anything involving prices, deadlines or commitments should get a human sign-off.

Which AI writes the best emails?

For everyday messages the major assistants are close. The prompt, meaning the facts and the style guidance you supply, makes a bigger difference than the model.