Why I Stopped Letting AI Write My Emails

For about four months, I let AI write most of my emails. It was faster, it was tidier, and it quietly cost me something I didn't notice until a friend replied to a message of mine with two words: "you okay?" I wasn't unwell. I just didn't sound like me anymore. This is what I learned pulling back — and the one half of my inbox I still happily hand over.

Let me be clear about where I'm coming from, because this is not an anti-AI piece. I write about AI tools most weeks, I pay for two of them, and I'll defend the good ones loudly. But somewhere last winter the "draft a reply" button crept into my email app, and I started tapping it the way you tap snooze — reflexively, before I'd really decided to. Within a fortnight, a genuinely large share of the words leaving my account had been written by a model and lightly waved through by me. It felt like progress. It took a while to see what it actually was.

The two-word reply that made me stop

That "you okay?" landed harder than it should have. I went back and reread a week of my own sent mail, and the pattern was embarrassing once I saw it. Every message had the same shape: a warm opener, three neat paragraphs, a friendly sign-off. Every one was grammatically immaculate. And every one was oddly weightless — the written equivalent of a handshake from someone looking over your shoulder. My actual voice, the one that runs sentences on too long and drops in a dry aside, had been sanded off. What was left read like a very polite stranger doing an impression of me.

The uncomfortable part is that the AI hadn't done anything wrong. The drafts were good. That's exactly the problem — they were good in the generic way a stock photo is good. And email, more than almost anything else I send, is where people expect to actually hear me.

Why "good enough" is the wrong bar for a personal message

Here's the distinction I'd completely missed. Not all email is the same job. A message to reschedule a delivery and a message to a friend going through a hard patch are both technically "emails," but they're doing utterly different work. For the first, tidy and fast is the entire spec. For the second, tidy and fast is almost an insult — the effort is the message. When you route both through the same "write this for me" button, you flatten that difference, and the personal ones are the ones that lose.

I'd written before about how AI doesn't automatically make you more productive, and this turned out to be the same trap in a different room. The tool made the task faster while quietly making the outcome worse, and because the task-speed is the bit you feel in the moment, you don't notice the trade until something breaks. A relationship email that lands flat doesn't throw an error. It just slowly makes you sound like nobody in particular.

"The drafts weren't bad. They were good in the generic way a stock photo is good — and email is exactly where people expect to hear an actual person."

The tell nobody warns you about

Once I started looking, I couldn't unsee it in other people's mail either. There's a texture to AI-written email. It leans on "I hope this message finds you well" and "I wanted to reach out." It resolves everything into balanced, reasonable-sounding paragraphs. It never rambles, never leaves a thought half-finished, never makes the tiny grammatical choices that mark writing as human. It reads like the model of an email rather than an email — and the more of it your contacts receive from everyone, the more a fully-generated one blends into the beige.

That's the strategic reason I pulled back, not just the sentimental one. If everyone's inbox is filling with the same smooth machine-prose, the thing that stands out is a message that plainly came from a person. In a sea of "good enough," sounding like yourself is becoming a genuine competitive advantage — and outsourcing it wholesale is throwing that away to save ninety seconds.

The half I still hand to AI, gladly

So I didn't quit AI in my inbox. I split my email into two piles and only gave it one of them. The line is simple: if the words carry a relationship, I write them; if they only carry information, I let the machine help.

What still goes to AI:

• Summarising before I reply. A forwarded thread twelve messages deep gets condensed to "here's what they're actually asking" before I answer a word of it. This is the job it's genuinely brilliant at.
• Skeletons for admin. Scheduling, form-style requests, "please find attached" logistics — mail where the wording truly doesn't matter and a template is the honest answer anyway.
• Editing my own draft. I write the thing, then ask AI to tighten it or flag where I've waffled. Crucially it's editing my words, not inventing new ones.
• Diplomacy passes. When I've written something too blunt at 11pm, "make this firmer but not rude" is a fair use — it's still my point, just better-mannered.

Notice the through-line: in every one of those, I'm the author and the AI is the assistant. It's the same principle I lean on when I use it to turn a messy meeting into clean, structured notes — the machine handles the shape and the drudgery, I keep the judgement and the voice. The failure mode is only ever the reverse: a blank prompt, a generated reply, a rubber stamp.

What actually changed when I switched back

Two things, and only one of them was what I expected. The first: my personal emails got shorter and rougher and, by every reaction I've had since, warmer. Turns out people don't want polished from a friend; they want present. The second, more surprising one: writing my own replies made me answer fewer emails — because when I had to actually compose a response, I noticed how many messages didn't really need one at all. The AI draft button had been so frictionless that I'd been replying to things reflexively. Removing it added just enough friction to make me think.

I still love these tools, and I still think most people underuse them for the boring 80% of their inbox. If you want the honest version of where AI earns its place in a real week — and where it quietly doesn't — the rest of what I've tested across AI tools and workflows lives in one place. But the reply you send to a person who'll remember it? Write that one yourself. It's the cheapest way left to sound like you mean it.

Questions people actually ask

Is it a bad idea to let AI write your emails?

Not always, but it depends entirely on what the email is doing. For anything that carries a relationship — a reply to a client, a note to a friend, an apology, a pitch — a fully AI-written message tends to read as smooth and hollow, and people can feel it even when they can't name it. For low-stakes admin where the content matters more than the voice, AI drafting is genuinely useful. The mistake is treating all email as one job. It isn't.

Can people tell when an email was written by AI?

Often, yes — not from any single word but from a texture. AI email tends toward tidy three-paragraph structures, a warm-but-generic register, and phrases like "I hope this message finds you well" or "I wanted to reach out." Once you've seen enough of it you start recognising the shape, the way you learn to spot a form letter. The tell isn't a mistake; it's the absence of the small imperfections that make writing sound like a person.

What is AI actually good for in your inbox?

The boring, structural half. I use it to summarise a long thread before I reply, to draft the skeleton of a form or a scheduling email where the wording doesn't matter, to tighten something I've already written in my own words, and to translate my own blunt notes into something more diplomatic. In every case I'm the author and the AI is the assistant — not the other way round. That distinction is the whole point.

How do you use AI for email without losing your own voice?

Write the message yourself first, then hand it to AI to edit rather than asking it to generate from scratch. Editing keeps your phrasing, your rhythm and your judgement in the driving seat while still catching waffle and softening anything too sharp. The moment you start from a blank prompt and paste whatever comes back, the voice you get is the model's, not yours — and over enough emails, people notice the difference.