Write and reply to work emails with AI
Six work emails nobody wants to write, what to ask for in each, where AI gets the tone wrong, and how to stop the reply reading like software.
It is twenty to five on a Thursday. Near the top of your inbox is a message you have opened four times today and closed four times, and now you also need to acknowledge the delay. Below it: a request you have to turn down, a chase you have already sent twice, and a thread of forty messages you joined at message thirty-four under the word "thoughts?".
None of that is really a writing problem. What slows you down is deciding what you want to happen next and how much you are willing to say out loud. Once those two are settled, the email takes ninety seconds.
Six situations follow, each in the same shape: what to paste, what to ask for, where it goes wrong, a verdict.
Platform: Web, iOS, Android
AI models: 60 from 18 providers, 21 of them on the free plan
Free plan: 50 messages per 24 hours
Price: Free, or Pro at $20.99/mo
Before you paste someone else's email
Work email is other people's words. Strip anything you would not want retyped elsewhere: salary figures, client identifiers, the name of someone not on the thread.
On the privacy question, the honest version is narrow. MultiChats does not train on your conversations. What the provider behind a given model does varies, and six of the 60 carry a notice on the card: the OpenRouter-routed entries say they are "Routed via OpenRouter to third-party providers that may briefly retain requests; they do not train on your data", while Grok's says outright "xAI may use your conversations to train Grok models." The other 54 say nothing.
The reply you have been avoiding
It has sat for six days. Every draft you start opens with an apology, and the apology makes the delay louder.
Paste. The original message, plus one line about why you actually did not answer. Write the unflattering version. It is context, and it does not appear in the email.
Ask for. "Reply in under 120 words. Acknowledge the delay once, in a single clause, then answer the question that was asked. No apology paragraph."
Where it goes wrong. Models grovel. Left to their own devices they spend two sentences on regret before reaching the answer, which is exactly the shape you were avoiding writing yourself.
Verdict. Worth it. The hard part was starting, and the model has no ego invested in the six days.
Saying no without burning the relationship
You cannot take this on, and you will need this person in three months. The refusal is easy. The relationship is the constraint.
Paste. The request, your real reason, and the alternative you are prepared to offer: a later date, a smaller slice of the work, or the name of someone better placed.
Ask for. "Decline in under 100 words. Make it unmistakably a no by the second sentence. Offer the alternative below. Give the reason once and do not repeat it."
Where it goes wrong. Ask for warmth and you get padding. The reader finishes a cushioned no thinking it was a maybe, then comes back next week. Models also invent flattery to soften the blow, so cut any compliment you did not supply.
Verdict. Strong, on one condition: you supply the alternative. A model asked to invent one will offer something you cannot deliver.
The third chase
Two polite nudges, no reply, and now you need the thing. The third chase has a tone risk running both ways. Too soft and it gets ignored again. Too sharp and you pay for it later.
Paste. Your previous two messages, the date you need it by, and what actually happens if it slips.
Ask for. "Write a third follow-up, five sentences maximum. State the date, state the consequence plainly and without threatening anyone, and offer one easy way to reply, including a no. Do not use 'as per my previous email', 'just circling back' or 'gentle reminder'."
Where it goes wrong. Passive aggression wearing a lanyard. Those three phrases are everywhere in the training data, so they arrive by default, and everyone reads them the same way.
Verdict. This is where a model genuinely earns its place. Firmness without heat is a register most people cannot hit while irritated.
Delivering bad news
The date slipped, the budget went, the answer is no and it was not your decision. Someone is going to be disappointed, and they will read this email more than once.
Paste. The facts in the order you learned them, and a clear line about which of them you are allowed to share.
Ask for. "Put the news in the first two sentences. Then what happens next, with a date. Then what I am doing about it. Apologise once at most, and do not use the word 'unfortunately'."
Where it goes wrong. Manufactured sympathy, the single most detectable thing AI does to an email. "I completely understand how frustrating this must be", from a colleague who has never written that sentence in their life, reads as a script, and the reader stops trusting the rest.
Verdict. Use it for structure and running order, which is most of the difficulty. Then write the sentence that carries the sympathy in your own words, or leave it out.
The introduction you were asked to make
Two people, one of whom asked you for this a week ago. Your job is a sentence on each, then getting out of the way.
Paste. What each person does, how you know them, and the specific reason they should talk now. "She's great" gives the model nothing, so it will fill the space itself.
Ask for. "Write two emails. First, a short check to each side asking whether they are happy to be introduced. Second, the introduction itself: one sentence per person, one on why now, then hand over. Under 90 words."
Where it goes wrong. Superlatives you never supplied. Give it a bare job title and you get "one of the sharpest people I know", going out under your name to someone who will remember it.
Verdict. Reliable, and the word limit is doing most of the work. Long introductions are the ones nobody acts on.
The long thread you were added to late
Forty messages, three arguments running in parallel, and you joined at message thirty-four.
Paste. The whole thread, oldest message first. Even a long email thread runs to a few thousand words, comfortably inside what the free models hold.
Ask for. "Read this thread and give me four things: what has been decided, what is still open, who is waiting on whom, and the two points I could add that nobody has made yet."
Where it goes wrong. Models are agreeable. Ask one what you should say and it will summarise the room's consensus back to you and present it as your view. The last clause of that prompt is the one that produces anything worth sending.
Verdict. The best use on this page. Read the thread yourself afterwards if a decision hangs on it.
Making it read like you rather than like software
A colleague who has read a hundred of your emails will spot a generated one, and the tells are consistent.
Length is the first. You write sixty words, the model writes two hundred. Register is the second: you open with "Hi", it opens with "Dear", you sign off "Cheers", it signs off "Kind regards". Then the stock furniture, "I wanted to reach out", "I hope this email finds you well", "please do not hesitate", plus punctuation you never use, particularly the long dash.
Three fixes, in ascending order of how long they last.
Per email: paste three emails from your own sent folder and tell it to match their length and register while ignoring what they say. It works, and it stops working the moment you open a new chat.
Once, for everything: put your house style in Personal preferences. On the web it is in Settings, under Account. In the mobile app it lives in your profile under Personalization, in the field labelled "Anything else the AI should know". You get 5,000 characters, it is free, and it applies to every new conversation without you doing anything. Worth putting in it: your job title, the greeting and sign-off you actually use, the phrases you never want to see, your default length. (Memories, which pick things up from your chats on their own, are the paid version of the same idea.)
Every time: the final edit is yours. Read the draft aloud, and anywhere you stumble the sentence is not yours. Replace one line with something only you would write, and the whole message passes.
When to ask a second model
Underneath any reply sit two actions that look alike and do opposite things. One regenerates, replacing the draft where it stands. The other branches, reopening the request in a separate thread so the first draft survives. Read the label before you tap: on the web the branch action is Try Again, while in the mobile app Try again replaces and Branch here keeps. Either way you can change model, and there are 60 to choose from.
Worth doing on the delicate ones. Two models handed identical facts will lead with different sentences, and in email the opening line is most of the tone. Read both openings and keep one.
Frequently asked questions
Can I do all of this on the free plan?
Yes. 21 of the 60 models are on the free plan, and 50 messages per 24 hours is more email than most people write in a week. Personal preferences is free too. Memories is the paid one.
Will people be able to tell I used AI?
If you send the first draft, often yes, and length gives it away before anything else. The three fixes above close the gap, and the last one matters most.
Which model should I use for email?
Most of the 60 will produce a competent work email. The differences show up in register and in the opening sentence, so run the delicate ones through two and keep whichever sounds closer to you.
Can it read the thread straight out of my inbox?
No. MultiChats does not read your mail, so you paste the thread in yourself and decide exactly what leaves your inbox. Going the other way is easier: on the web, a reply that comes back as a draft carries a Send via Gmail button, with Open in Mail as the alternative. Both open a compose window already filled in and neither sends anything for you. On mobile you copy the draft across.
Sixty models, one subscription, on web, iOS and Android. Free to start, Pro at $20.99/mo.