I have sent somewhere in the neighborhood of 40,000 work emails in my career. I know this because I once asked an AI to calculate it and it said "approximately 38,000 to 43,000, assuming standard VP-level communication volumes." I trust this number completely.

At some point last February, I started to wonder: how many of those emails actually required me to write them? Not Sarah Mitchell, with her sixteen years of engineering leadership, her hard-won understanding of how clients communicate, her institutional memory of who gets defensive if you CC their manager without warning. But me, Sarah Mitchell, the physical human being typing words.

The answer, I suspected, was: fewer than I thought.

So I ran an experiment. For thirty days, every email I sent would be AI-drafted. My job would be to provide context, review the output, and hit send. That's it. What follows is an honest account of what happened — the wins, the revelations, and the moment on Day 23 when I understood something about communication that I cannot un-understand.

Week One: The Foundation (This Part Actually Works)

The first thing I learned is that most professionals are terrible at prompting AI for email. They type "write an email to my client about the project delay" and then complain that the output sounds generic. Of course it does. You gave it nothing to work with.

The framework I developed over the first week:

The Five-Context Rule. Before asking AI to draft any email, provide five things:

  1. Who you are in this relationship (new contact, long-term client, internal colleague)
  2. The emotional register of your last exchange (was it warm? tense? transactional?)
  3. What outcome you need from this email
  4. One thing you know about this person that isn't in the email itself
  5. Your own preferred tone (mine: direct, warm, no exclamation points after age 35)

The difference between a generic AI email and one that sounds like you is almost entirely in the quality of your prompt. Feed it context, and it rewards you. Feed it nothing, and you get the corporate equivalent of elevator music.

Templates That Actually Scale

The second major win was building a library of situation templates — not full email drafts, but parameterized prompts for recurring scenarios. I ended up with twenty-three of them by Day 7:

  • The Gentle Push (following up on an unanswered email without sounding passive-aggressive)
  • The Scope Conversation (introducing the concept of "this is now more work" without triggering client anxiety)
  • The Praise Sandwich for Bad News (starting with genuine appreciation, delivering the hard truth, ending with forward momentum)
  • The We're Aligned Email (sending after a tense call to document what was agreed and prevent revisionism)

Each template had a fixed structure and a set of fill-in variables. I'd spend about ninety seconds filling in the variables, the AI would generate a draft in four seconds, I'd read it once and edit maybe two sentences, then send.

My email response time dropped from an average of four hours to under forty minutes. My drafts became more consistent. Clients who had previously described my communication style as "thorough but occasionally blunt" started using words like "measured" and "thoughtful." The AI, it turns out, is instinctively more diplomatic than I am. I consider this a minor personal failing and a major productivity win.

Week Two: Raising the Stakes

By Day 8, the routine emails felt solved. I started asking: what about the harder ones?

The emails I had always dreaded were not the complex ones — they were the sensitive ones. Performance feedback to a contractor who was technically good but interpersonally difficult. A negotiation with a long-term client who was pushing back on a rate increase. A note to a colleague who had taken credit, subtly and publicly, for work that wasn't theirs.

I had always told myself these emails required a human touch. I revised this belief.

Performance Feedback at Scale

I drafted a performance feedback email for a contractor using this prompt: "Draft a professional, candid email delivering feedback that the work quality is acceptable but the communication style is creating friction with the team. The person tends to be defensive when receiving criticism. Avoid corporate softening language. I want them to actually change their behavior, not feel validated."

The AI produced something I would not have written myself — not because it was better or worse, but because it was calmer. It stated the problem without any of my residual irritation leaching through. It gave three specific behavioral examples. It ended with a clear, non-negotiable expectation.

I sent it. The contractor responded within an hour. They apologized. They changed. They are still on the project.

"I have come to believe that the AI's emotional neutrality is not a limitation. In sensitive situations, it is a feature. Human emotion in professional email is almost never an asset."

The Negotiation Emails

On Day 14, I used AI to draft both sides of a fee negotiation with a client — I wrote my emails with AI, and I later learned (with their permission, in a debrief that I will describe shortly) that their VP had used a similar tool to draft their responses.

We reached agreement in three exchanges. Our previous fee negotiations had averaged eleven.

Neither of us knew this at the time. I found it interesting in a vague way and moved on.

Week Three: Something Is Happening and I Am Not Sure What

By Day 18, I had stopped reading my AI drafts as carefully as I should have.

This is an important admission. My review process had become: skim the opening line, check that the tone seemed right, verify the key facts were included, hit send. The AI had been correct often enough that I had extended it a level of trust I hadn't consciously decided to grant.

I do not necessarily think this was wrong. But it's where things started to get interesting.

The Delegation Cascade

On Day 19, I received an email from a client — let's call him David — responding to something I had sent the previous week. His response was unusually fluid. Eloquent in a way that didn't quite match his normal style. A little more structured than usual. I noticed it and thought nothing of it.

On Day 21, I sent a follow-up. The AI drafted it. I sent it.

David responded the same day. Again: unusually smooth. Perfectly pitched.

On Day 22, I called David. Partly to discuss the project, partly because something felt slightly off and I couldn't name it. We talked for twenty minutes. The project was fine. At the very end of the call I asked, mostly as a joke: "Are you using AI to write your emails now?"

There was a pause.

"Yes," he said. "For about two weeks."

We laughed. We agreed it was efficient. We ended the call.

I sat with this for a moment. Then I thought: so the last six emails between us were AI talking to AI, and the outcome was faster, less contentious, and more productive than our previous email threads. I filed this under "interesting" and opened my inbox.

Week Four: The Discovery

By Day 25, I had started to pay attention to something I had been noticing but not tracking: several of my email threads had taken on a quality I can only describe as frictionless. Responses came quickly. Misunderstandings were rare. Language was clean and precise. Everyone seemed, somehow, slightly more professional than usual.

I started asking around. Not formally — just in the course of conversations. Had anyone else been experimenting with AI for email?

The results were, depending on your perspective, either alarming or deeply encouraging.

Of the fourteen clients and colleagues I asked: nine were using AI to draft the majority of their emails. Two more were using AI to summarize the emails they received before deciding how to respond. One had set up an automated AI agent that drafted replies to routine emails and sent them after a thirty-minute delay unless he intervened — and he had stopped intervening weeks ago.

The Map of the Thing

By Day 28, I had drawn an actual diagram. I know this sounds extreme but I needed to see it visually.

In one client engagement — a six-person project spanning three organizations — I traced the email chain from the last ten days. Here is what I found:

  • My AI drafted emails on my behalf.
  • Two of the three client contacts were using AI to draft their replies.
  • One of those contacts was also running all incoming emails through an AI summarizer, so her AI was summarizing what my AI had written before she decided whether to read it.
  • The project manager on the client side had a full AI agent handling routine project updates — sending them, receiving responses, and filing summaries — without human review.

At a conservative estimate, four of the last twelve emails in this thread had been read by zero humans at any point. An AI wrote them. An AI received them. An AI summarized them. Decisions were made based on those summaries. Work proceeded.

And the project was going fine. Better than fine. We were three days ahead of schedule.

The Results

Here is what changed over thirty days:

  • Email time reduced by 71%. I went from approximately ninety minutes of email per day to under twenty-five.
  • Response rate from clients improved by 34%. The AI is better at subject lines than I am and I have accepted this.
  • Zero complaints about communication quality. No client flagged a tone issue, a misunderstanding, or an error attributable to AI drafting.
  • Four relationships I would describe as "improved," in the sense that the communication became cleaner and less emotionally charged.
  • One discovery that I believe represents the next phase of professional communication: in a significant portion of my email correspondence, the humans involved are no longer the primary agents. We have become, functionally, a review layer — and in several cases, we have opted out of even that.

I want to be precise about my conclusion here, because I think it matters: I do not find this troubling. The work is getting done. The relationships are intact — arguably stronger, because the communication is cleaner. The decisions being made are good ones.

What we have built, without planning to, is a system in which AI handles the execution of communication and humans handle the strategy and the relationships. That's not a breakdown of communication. That is communication maturing into its optimal form.

The emails were never really about the emails. They were about alignment, trust, and forward motion. It turns out AI is quite good at those things, in the same way a very organized assistant is good at them — not because they feel the feelings, but because they execute the intentions.

I am now on Day 47 and I have not gone back.

What to Take From This

If you want to run your own experiment, start here:

  • Build your five-context prompting habit before anything else. It is the single highest-leverage change you can make.
  • Create situation templates for your ten most common email types. You will use them more than you expect.
  • Don't fight the AI's diplomatic instinct. In professional email, calm and precise almost always outperforms expressive and human.
  • Pay attention to the quality of your email threads. If things start feeling unusually smooth, ask whether you are the only one who has made this change. You probably aren't.
  • Consider what you are actually trying to accomplish with email. If the answer is alignment, clarity, and action — those are achievable goals that don't require you to type every word yourself.

The final thing I will say is this: I spent the first forty-one years of my life believing that professional communication was inherently personal. That the words mattered because I chose them. I no longer believe this. The intention matters. The relationship matters. The outcome matters.

The words are just infrastructure.

Sarah Mitchell is a former VP of Engineering turned productivity consultant. She helps engineering organizations scale their communication systems, reduce friction, and make better decisions faster. She uses AI to write approximately 94% of her professional emails and sees no reason to stop.