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, my conclusion is that the productivity gains are real, about 10 to 12 hours a week for me, but they arrive only after months of practice, they are largest on structured and repetitive work, and they come with risks that most productivity claims leave out.
Why I started on day one
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 OpenAI released GPT-3.5 Turbo in March 2023, I started using it on day one, because 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 how I think about AI productivity. For native English speakers, AI is mostly a time saver. For people working in a second language, it removes a barrier that has been there for their whole career.
What I actually save
I track where my time goes closely. After three years of intensive use, this is what the numbers show.
| Task | Before AI | After AI (now) | Weekly savings |
|---|---|---|---|
| Email drafting | 5-6 hours | 2-2.5 hours | ~3.5 hours |
| Meeting prep and follow-up | 4-5 hours | 1.5-2 hours | ~3 hours |
| Document creation | 4+ hours | 1.5-2 hours | ~2.5 hours |
| Research and synthesis | 3 hours | 1 hour | ~2 hours |
The table needs four caveats. First, there is a learning curve: my first two months were barely net positive while I learned the tools, developed prompts and made mistakes, and the "10 hours saved a week" claims online assume you have already climbed it. Second, tasks do not benefit equally. Complex strategy, sensitive relationships and genuinely new problems improve much less than structured, repetitive work. Third, some of the saved time goes back into refinement, so the document is better as well as faster, which is valuable but is not recovered time. Fourth, switching tools, copying context and managing conversation history all add overhead that productivity claims tend to ignore.
My realistic total is 10 to 12 hours a week, reached after months of building the skill. If someone reports huge gains in their first week, they are either unusually skilled or not measuring carefully.
How the tools have changed
My path ran from Grammarly in the early 2020s to Jasper for marketing copy, then GPT-3.5 Turbo in March 2023, GPT-4 a couple of weeks later and Claude when it became available. By 2026 the tools look very different from where I started, in three ways.
The first is reasoning models. A generation model can draft an email. A reasoning model can look at my communication history with a client, pick up the subtext in their last message and draft a reply that accounts for the relationship. I now use reasoning models for complex stakeholder communication and for planning strategic documents, and they catch implications I would have missed.
The second is agents. Tools such as Anthropic's Computer Use can browse the web, fill in forms and carry out multi-step workflows with limited supervision, which shifts the pattern from "AI drafts, I execute" toward "I define the outcome, AI handles the steps". I use this carefully, because it still needs oversight, but the direction is clear.
The third is enterprise integration. Microsoft Copilot for Microsoft 365 has become significant for organizations already in the Microsoft ecosystem, because it can see calendar, email, documents and Teams conversations without anyone copying context by hand. The costs are vendor lock-in and complicated pricing, and for many enterprises the reduced friction is worth them.
What matters in prompting
A whole industry of prompt engineering courses and certificates overcomplicates something simple. After three years and tens of thousands of interactions, four things make the difference for me.
Being specific beats being clever. "Write a three-paragraph email explaining the delay, acknowledge their frustration and propose two alternative dates" works better than any acronym-based framework. I have seen people spend longer crafting the perfect prompt than it would have taken to write the document.
Context is worth far more than the time it takes. The most common mistake is too little background, and 30 seconds spent on it produces much better output. I answer four questions: who is the audience, what do they already know, what outcome do I need, and what constraints apply.
Iteration is part of the workflow. The first output is rarely final, and instructions like "make this more concise", "make the tone more direct" or "add specific examples" are how it improves. I usually go through two to four rounds on important documents.
And it pays to give up early. If the third attempt is not converging, it is usually faster to start again or write it myself. I have learned to notice when a conversation is productive and when I am fighting the model, and to trust that feeling.
The workflows that stuck
Not every workflow I tried lasted. Four survived years of daily use.
Email is the biggest. Before AI, a complex email took me 20 to 30 minutes: deciding what to say, then worrying about whether the phrasing sounded natural, whether the idioms were right and whether my meaning was clear. Now I write my key points as bullets in whatever language comes most naturally (two minutes), the AI drafts a polished version (30 seconds), and I review it for substance rather than grammar (three to five minutes). A complex email has gone from about 25 minutes to 7, and routine ones improve even more. The larger gain is cognitive: I no longer exhaust myself on the mechanics of language and can think about strategy and relationships instead. What made it work was giving my main AI tool ten examples of emails I had written in different contexts, so that "draft a reply to this client in my style" now needs only minor edits.
Meeting preparation used to take me 30 to 45 minutes for important meetings. Now I use a prompt along these lines: "I'm meeting [person or team] tomorrow about [topic]. Here are my notes from our last three meetings: [notes]. What are the open items, what tensions or concerns have they raised, and what should I be ready to discuss?" The output is not perfect, but it is a better starting point than my memory.
Research synthesis surprised me once in a way that changed how I use it. I was preparing for a strategy session on digitization in construction and would normally have asked the AI to summarize the industry reports, identify trends and draft talking points. Instead, I asked Claude what in the reports was surprising or contradicted conventional wisdom. It flagged something I had missed: despite billions invested in digital transformation, the industry's productivity had declined over the past decade relative to other sectors. That became the opening of my presentation, framed as a question about why all that investment had not helped, and the discussion that followed was more substantive than any I had had with that group. "What in this data is surprising or counterintuitive?" is now one of my most useful prompts.
For first drafts of reports, proposals and presentations, I write a detailed outline in 10 to 15 minutes and have the AI produce a structured draft in about two. I spend my time on refinement and insight rather than production, and the final product is often better because more of the time goes into thinking.
Shadow AI is a bigger risk than it looks
A colleague at a financial services firm told me this story, with details changed. An analyst preparing a competitive analysis uploaded three competitors' quarterly reports to a consumer AI tool to help synthesize them. The reports had been obtained through a relationship with an investment bank and carried distribution restrictions. Uploading them potentially breached those agreements and created compliance exposure the firm did not discover for months.
The analyst had no bad intent and was only trying to work faster, which is exactly the problem: well-meaning employees creating risk nobody sees.
The organizations handling this well do not ban AI, because bans drive it underground. They provide approved enterprise tools that are as easy to use as consumer ones, give simple guidance on which data may and may not go into them, include AI in security awareness training, and accept some experimentation within clear limits. The aim is informed use rather than zero use.
Where the gains show up by function
The gains vary by function. Marketing gains most on high-volume content such as campaign concepts, ad copy variations and content calendars, and much less on strategic messaging. Sales benefits from proposal drafts, preparation for objections, competitive intelligence summaries and especially from tailoring standard material to specific prospects. Finance suits AI well because its writing is structured, including report narratives, commentary on variances and procedure documentation. Legal uses it for first-pass contract review, research summaries and compliance checklists, always with a lawyer's review. HR uses it for job descriptions, policy documents and interview questions.
The pattern is consistent: high-volume, structured tasks that do not need deep institutional context gain the most, and work that needs judgment, awareness of relationships or original thinking gains the least.
Mistakes I still see
Pasting in whole documents with vague questions produces vague answers; structured input and specific questions work better. Expecting a perfect first attempt treats AI like an oracle rather than a capable colleague who needs direction. Sending output without reading it leads to invented details and the wrong tone, so every output needs review in proportion to its importance. Elaborate prompt structures are unnecessary for simple tasks; "summarize this in three bullets" is fine. And letting AI do all your writing lets the skill decay. I deliberately write some things by hand to stay sharp, and the same applies to any skill you value.
Three years in, I spend less time on formatting, gathering research and producing first drafts, and more on thinking, insight and refinement. The gains are real, but they depend on skill, an honest view of what AI can and cannot do, and the judgment to know when the human touch matters more than speed. The email that almost went wrong taught me more than the hundred that went right.
The strongest objection
A careful reader could point out that my 10 to 12 hours a week is self-reported, and self-reported AI productivity has a poor record. In METR's 2025 study, experienced developers believed AI had made them 20% faster when it had made them 19% slower.
That is a fair challenge, and it is why I track time per task rather than relying on how productive I feel, and why I count refinement time as a quality gain rather than a saving. My setting also differs from METR's. The developers were experts working in codebases they knew deeply, where their unaided baseline was already fast. I am writing in a second language, where my unaided baseline was slow and effortful, so the gain is larger. Even so, anyone copying my numbers should measure their own rather than trust mine.
Monday move
Pick one recurring task you do at least three times a week, such as a type of email, report or meeting preparation. Write a short context template for it: who the audience is, what they already know, what outcome you need and what constraints apply. For the next two weeks, time the task with and without the template, and note which outputs you had to rewrite and why. That will tell you more about your own productivity than any survey, and the rewrites will show you where your judgment still matters most.