Almost every tool that helps you write email produces usable prose. The difference is made before and after. Before: what the tool knows about your case, meaning whether it has the deal, the price and the last conversation, or only your prompt. After: who presses send. Five questions separate these products in everyday use, and none of them is "how well does it write". The built-in features in Outlook, Gmail and Apple Mail are drafting aids per their makers. Tools with write access to your mailbox can do more, and that is where the second look pays off: what goes out without asking doesn't save time, it moves the risk. Everything below comes from the vendors' own pages, retrieved 20 September 2026.
For my own office I ended up with a configured role rather than a feature inside a mail client, because context mattered more to me than phrasing. How I build roles is collected on AI Employees; who does what in my business is in My AI Workforce. This post is about writing; the tool question for sorting your inbox is answered in AI Email Triage: Agent or Tool?.
The five questions that decide it in practice
Get questions one and three wrong and you find out late and expensively. Get question two wrong and you find out immediately.
What the tools do in September 2026, according to their makers
The table shows why "which tool writes best" is the wrong question. Between the first row and the second-to-last there is no quality gap in sentence construction, there is a gap in access. And access hangs on a license: what sits in your tier is stated by the vendor, not by the product name.
Context: the part no tool invents
A new-client email is four or five facts that live somewhere: the right name, what was last discussed, the link, the price, the deadline. Writing it takes two minutes, collecting the facts takes twenty. That makes context the most expensive of the five questions.
Microsoft describes Copilot as connecting language models to organizational data through Microsoft Graph: documents, mail, calendar, chats, contacts, and only as far as the individual user has at least view permissions. A chat window with no connection knows none of that until you paste it in. Not a shortcoming, a division of labor.
The point where both routes are equally weak: what is written down nowhere, the tool doesn't know either. One that writes something anyway produces a fluent email with a wrong number in it.
Tone: learned or generic
This is where the promises differ. Microsoft lists a function in its Copilot overview for Outlook that is meant to make drafts sound like the user. Apple writes on its product page that in Messages and Mail, Siri AI can match your writing style, punctuation and tone. Superhuman advertises auto drafts in its Business tier. These are vendor claims, not measured results: those pages don't say what the style is derived from or how the result was checked.
What actually makes the difference is independent of the tool: whether real emails of your own exist as material, or only an adjective in the prompt. "Casual but professional" returns the average of other people's mail. Uncomfortable side effect: the better the voice fits, the less likely you are to notice on a quick read that the price inside is wrong. How to make your tone describable in six lines is in Can AI Write Emails in My Voice?.
Draft or send: the point that carries everything
The built-in features of the three big vendors are drafting aids by their own descriptions. Microsoft puts it in its privacy documentation: the Copilot capabilities provide useful drafts and summaries and give you a chance to review what was generated rather than fully automating these tasks. In all three cases you tap send yourself.
It's different for a tool with write access to the mailbox. Anthropic describes two stages for the Microsoft 365 connector: reading mail, calendar, files and Teams messages, and only with write tools explicitly enabled, creating drafts, sending and forwarding. That requires extra permissions and an administrator to switch it on, and it matters more than any phrasing feature.
My own line on it, and this is a setup decision, not a product property: sending stays with the human. Not because the drafts are bad, but because you can't rely on a recall: delay and recall options exist, but they only work under narrow conditions. A wrong price in a draft costs thirty seconds, the same price in a sent email costs you the negotiation. What following up without pressure looks like is in Follow-Up Email Templates, No Pressure.
Data and price
The two belong together, because the cheapest option is usually the one already paid for. The prices are the vendors' list prices, retrieved 20 September 2026, in the currency each page shows. The data statements beside them apply per product and tier, not to the vendor as a whole.
Often the feature is already paid for: anyone on Google Workspace has the Gemini assistant in Gmail from the Starter tier. But every data promise hangs on the product and the tier: what holds for a business tenant doesn't automatically hold for a personal account of the same brand.
Which of the other people's names, amounts and dealings in your mailbox belong in a tool is covered in AI and Privacy: What the AI Gets to See. If you're torn between Microsoft's assistant and Claude, the full comparison is in Copilot vs Claude: What Works at the Office.
What none of these tools takes off your hands
The second half of the headline. Four things stay with you whichever tool you pick:
- The message. Whether you accept, decline or hold your price is a decision. A tool phrases it, it doesn't make it.
- The facts. Name, link, price, deadline come from a source. If one is missing, asking is the only fix.
- The timing. Whether today is the day for the third follow-up isn't in any database.
- The responsibility. What sits with the recipient sits there with your name on it.
If you want to cut the effort further, you don't need better phrasing, you need better templates and source files: the collection in Business Email Templates That Work, how a business email holds together in Writing Business Emails: With or Without AI.
How it runs in my business
Here that job belongs to Sonja, my AI assistant for business email. She isn't a product from the table above, she is a configured role: connected to my Microsoft 365 mailbox by device-code login, with no password in a file. Her tone comes from 50 of my real sent emails. She checks every fact against the source file, leaves the text as a draft in the mailbox, and sends only after an explicit word from me in the individual case.
In numbers: for a campaign of 15 new-client emails in July 2026 I had sent eight myself, the remaining seven drafts were in the mailbox within two minutes, every demo link in them verified with HTTP 200 the same day, and one oddity went back as a flag. I did the sending. The whole chain is on Sonja, who writes the email in my voice. What I was missing wasn't the phrasing, it was access to the source files.
Common questions
What is the best AI tool for email?
Two questions first, then the product: which facts have to go into your mail, and where do they live? Which license and which permissions does the feature you need for that require? If the context sits in the mailbox, your own provider's assistant is the short route: Copilot in Outlook on Microsoft 365, the Gemini assistant in Gmail on Google Workspace, both licensed by feature. If it sits in source files beside it, a separate tool pays off.
Is there an AI assistant for Outlook?
Yes, in two stages. Microsoft lists drafting a message, matching your own style, coaching and thread summaries for Copilot in Outlook. Without the add-on license, Microsoft lists Copilot Chat in Outlook for questions about your inbox, calendar and meetings plus actions directly in Outlook; with the add-on license, chats and enterprise data come along. Check before you buy which stage holds the feature you want.
Does an AI email tool send mail automatically?
The built-in features in Outlook, Gmail and Apple Mail are drafting aids per their makers, and you send yourself. Tools with mailbox write access can do more: Anthropic lists sending and forwarding for the Microsoft 365 connector, but only with separately enabled write tools and extra permissions.
Will a tool really write in my style?
Microsoft, Apple and Superhuman all advertise style matching without showing on their product pages what the style is derived from. It becomes reliable where you supply the material: three to five real emails of your own from different situations, plus a list of things you would never write.
What does an AI email tool cost?
Retrieved 20 September 2026: Microsoft lists the Microsoft 365 Copilot Business add-on from $21.00 per user and month with an annual commitment, and from $18.00 during the promotion. Google Workspace starts at €6.80 per user and month including the Gemini assistant in Gmail. Superhuman charges $25 or $33 per member and month billed annually.
Is a plain chat window enough?
For individual emails, yes, and it's the cheapest sensible start. It just knows nothing about your case. Once you're collecting the same facts by hand for the third time, the step to a role with access to your source files pays off.
How to take it further
Answer the five questions for your case on one page: where the facts for your mail live, whether you have emails of your own as a tone sample, who presses send, what the tool may see and what license that costs. Usually the expensive part isn't the tool, it's that the source facts sit in five different places.
Sonja exists with exactly this structure as a ready-made package in my community Claude Practitioners, alongside the other AI employees and the course that goes with her. If you get stuck sorting your sources, bring the question to one of the calls.
