I almost sent an AI-generated email that would have damaged a relationship.
It was a Friday afternoon. A long-standing stakeholder had sent a terse message about a project delay, and I was tired, so I asked Claude to draft a reply. The draft was technically correct. It acknowledged the delay, explained the cause and proposed a revised timeline.
The tone was wrong. The AI did not know that this person had championed our project internally, had taken a personal risk for us and was now facing questions from their own leadership. They did not need a status update. They needed reassurance that their advocacy would pay off. I rewrote it from scratch in two paragraphs: a personal acknowledgement of their support, a commitment to making them look good, and the timeline last.
Since then I have not used AI for relationship-sensitive messages without a lot of my own judgment on top. After three years of daily use, I estimate that it saves me about ten to twelve hours in a typical week. That is a personal estimate, not a controlled comparison or a productivity target for anyone else. The useful part of the story is how I reached it, and what I still rewrite.
Why I started
English is not my first language. For years I relied on Grammarly and Jasper to keep my professional writing correct, and dictionaries were constant companions. Operating in a second language while trying to be persuasive in executive communication carries a real cognitive load.
When I began using generative AI in 2023, I could immediately see what it would do for someone like me: close the gap between how I think about strategy and how fluently I can express it. Instead of spending mental energy on grammar and phrasing, I could concentrate on what I wanted to say. The AI handled the linguistic polish, and I supplied the substance and judgment.
That shapes my experience. A fluent writer may have a much faster unaided baseline. For me, getting a sentence right had often taken attention away from getting the thought right. Removing some of that friction changed how much energy I had left for the conversation itself.
What I count as a saving
Email drafting, meeting preparation, document creation and research synthesis account for most of the difference. The estimate has limits. My work varies from week to week; I cannot replay the same week without the tools. Some of what feels faster is a task I would previously have done less thoroughly.
My first two months were barely net positive while I learned the tools, developed prompts and made mistakes. The gains grew as I found recurring tasks worth keeping. Complex strategy, sensitive relationships and unfamiliar problems benefit much less reliably than structured work.
Some saved drafting time goes straight back into refinement. A better document is valuable, but the time spent improving it has not been recovered for another task. Switching tools, copying approved context and managing conversation history also have a cost. I count those activities as part of the job, rather than timing only the model's response.
This matters for a team budget. My estimate does not establish that an employer could remove that many hours from a role, or turn them into revenue. The next question is what the person actually does with the capacity: follow up earlier, prepare better, take on another assignment, or simply make the working day more manageable.
What matters in prompting
Being specific beats being clever. “Write a three-paragraph email explaining the delay, acknowledge their frustration and propose two alternative dates” works better for me than an elaborate acronym. I have seen people spend longer crafting the perfect prompt than it would have taken to write the document.
Context is worth the effort. I answer four questions: who is the audience, what do they already know, what outcome do I need, and what constraints apply? In the Friday email, the missing constraint was the relationship. A technically accurate account of a delay could still hurt the person who had backed us.
Iteration is part of the workflow. Instructions such as “make this more concise,” “make the tone more direct” or “add the missing example” help me see what I actually want to say. Important documents usually need several rounds.
It also pays to give up early. If another revision is not improving the result, I start again or write it myself. I have learned to notice when a conversation is productive and when I am fighting the model.
The workflows that stuck
Email is the biggest. I write my key points as bullets in whichever language comes most naturally, have the AI draft a polished version, then review it for substance, tone and commitments. The larger gain is cognitive: I no longer exhaust myself on the mechanics of language and can think about strategy and relationships instead. Examples of my own writing help when they are appropriate to share in the approved tool; they do not remove the need to review a reply.
For meeting preparation, I use a question along these lines: “I'm meeting this team tomorrow about this topic. Here are the approved notes from our previous discussions. What are the open items, what concerns have they raised, and what should I be ready to discuss?” The result gives me a starting point. I check it against the notes, especially where a tentative comment has been turned into an agreement.
Research synthesis changed for me when I stopped asking only for summaries. Before a strategy session on digitization in construction, I asked what in the reports was surprising or contradicted conventional wisdom. That question prompted a more useful discussion than a list of trends. I still use it, but I go back to the report before putting the suggested claim in a presentation. A surprising interpretation deserves more checking, not less.
For reports and proposals, I write a detailed outline before asking for a draft. That is where I decide the argument, audience and evidence. If I hand over that thinking as well as the typing, the result becomes fluent without having much to say. The draft is useful when it helps express a position I can already explain.
A convenient tool still needs permission
A colleague in financial services described an analyst uploading material with distribution restrictions to a consumer AI tool. The intention was to work faster. The concern was whether the upload was permitted. I did not investigate that incident, and it is not evidence about every consumer tool or a legal conclusion about the firm's agreements.
The practical lesson is to check before sharing. An approved account, clear rules for documents and a usable route for questions make responsible use easier. People need to know what can go into the tool, where the output can be used and whom to ask when the answer is unclear. A general encouragement to experiment leaves too much for an employee to guess.
Measure one task, including the rewrites
Feeling productive is a poor substitute for measuring the work. In METR's study of experienced developers using early-2025 tools, participants felt faster after using AI, while the measured work took longer. That finding concerns a particular coding setting; it does not settle my experience of writing in a second language. It does make me more careful about treating a feeling as a result.
For a recurring task, compare the complete time with and without assistance. Count preparation, checking and corrections. Note whether the output was useful to its recipient. A meeting brief completed sooner has limited value if it misses the unresolved issue everyone came to discuss.
I deliberately write some things by hand to stay sharp. Other things benefit from a fast draft and a careful review. The distinction has become easier to make with practice, though it is never automatic.
The email that almost went wrong taught me more than the hundred that went right. I had saved a few minutes producing words and was about to spend a relationship. The useful habit now is to ask, before sending: what does this person need me to understand? That remains my work.