Don't Replace Me, a book by Dmitry Kargaev, includes a three-day plan called Your First 72 Hours for starting to use AI at work. You finish with one boring task handed over and one workflow rebuilt. Day 1 and Day 3 take an hour each, Day 2 takes 30 minutes.

Kargaev wrote the book and owns dontreplace.me, so this is the author's own plan, not an outside review. It follows chapters 1, 6 and 11 to 15 of the second edition (2026).

Before you start: what the plan needs

The plan needs a chat assistant, a sheet of paper, your calendar, sent folder and chat history from last week, and a stopwatch. It follows the second edition, checked against the book on 11 November 2026. Tool names and prices are left out on purpose, because they go stale.

The book's premise is that AI takes over tasks much faster than it takes over jobs, and it does it unevenly, company by company, desk by desk. So the plan starts with tasks, and with what your employer allows. Do these five steps once, before Day 1.

  1. Run the distance test. Find out which AI tools your company pays for or allows, whether a written policy exists, and find one person who uses AI daily. The book prints a script for asking your manager or IT.
  2. On the same paper, add one "appendix" item: a task you fear a machine will take, such as status reports or meeting notes.
  3. Choose your daily assistant, the one your distance test says is allowed. The book's loadout has two slots, one chat assistant and one specialty tool, and it advises learning tool categories rather than brands. Leave slot two empty for now. The book fills it from day 15.
  4. Turn off training on your chats if your work is sensitive. Business and enterprise plans are generally excluded from training by contract, while consumer defaults change often.
  5. Write three sentences of custom instructions about you and your job, and make one project holding two documents you lean on.

Day 1: audit last week and time one task

The Day 1 audit in Don't Replace Me finds the one task worth handing to AI, and a stopwatch tells you whether the handover helped. Give it an hour.

  1. List 30 to 50 tasks from last week, with the minutes each took. Pull them from your calendar, your sent folder and your chat history.
  2. Mark each task "needs me" or "needs a checker". A task needs a checker when you would know a good result on sight and could confirm it in under two minutes.
  3. Add up the minutes in each group and see what share of your week sits under "needs a checker". The book calls the target the 40% rule. That figure is the author's own rule of thumb, with no study behind the exact number, so your share is the one that counts.
  4. Put a star beside one task: the dreaded one that comes round again and again. One star only. This is the task you hand over first.
  5. The next time the starred task comes up, do it with the assistant and run the stopwatch. Write the time next to what the task normally costs you. The book's example numbers: cutting a daily task from 45 minutes to 10 would save about 140 hours a year. They are illustrations, not a promise.

The stopwatch is there because of a METR experiment from early 2025, with 16 developers and 246 tasks. The developers expected to be about 24% faster. They were 19% slower, and afterwards still believed they had been about 20% faster. In February 2026 METR said today's tools likely help more, and that many developers refused to work without AI, so it redesigned the study. The book's lesson is to measure instead of trusting the feeling: "Chatting with an AI feels productive in the same way that rearranging your desk feels productive."

Day 2: brief the assistant like a smart intern

A good brief gives the assistant real material and context, then asks it to attack its own answer. Day 2 takes 30 minutes.

Two facts from chapter 12 explain the second half. Clever Hans, investigated in Berlin in 1904 by a commission of 13 people, was right 89% of the time when the questioner knew the answer and 6% when the questioner did not, by Pfungst's count. Chat assistants lean toward what you seem to want, a habit called sycophancy that comes from training on human ratings.

The book's brief template has these fields:

For a weekly status email, a filled-in brief might read: act as a project coordinator who writes short, plain updates; I want this week's status email; the readers are two directors who skim; the goal is to flag the one decision I need; here are my notes and last week's email; avoid jargon; here is an earlier email that worked; ask me up to three questions first; tell me where you are unsure. That example is invented for illustration.

  1. Open a new chat and fill in every field for your starred task. Paste in the real material, such as notes, an email thread or a report, as long as your distance test says you may.
  2. Send it and answer any questions the assistant asks. The book's comment on a brief like this: "Same machine. The second prompt has about ten things in it that only you knew."
  3. Send a second message asking the assistant to read as the toughest reader of its own answer and to list the three weakest points and every fact it is unsure of.
  4. Fix those points yourself, time the whole run, and save the brief in a note so Day 3 can reuse it.

This came from a book.

Don't Replace Me

200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff.

Get the Book →

Day 3: rebuild one workflow with belt, call and scar tags

Rebuilding a workflow means redrawing a whole repeating job around AI, not adding AI to one step. The book calls bolting AI onto old steps a motor in the basement. Factories were electrified from the 1880s, yet productivity barely moved until the early 1920s, when slightly more than half of drive power was electric. Day 3 takes an hour.

  1. Pick one repeating workflow and write down every step.
  2. Tag each step with one word. Belt means the step exists because of old machinery. Call means it needs your judgment. Scar means someone added it after something went wrong.
  3. Redraw the workflow. Hand the belts to the assistant, using your Day 2 brief saved as reusable instructions. Keep the calls yourself. Ask about the scars.
  4. Run the new workflow once with the stopwatch. Fix the step that broke.

Check the output before you send it

Checking AI output yourself matters because its mistakes now look finished. Chapter 14 says hallucinations have become rarer but subtler: a real link with the wrong claim behind it. Asking the same model "are you sure?" is not a check, because it is "the same guesswork running a second time". In Mata v. Avianca a lawyer, Steven Schwartz, cited fake cases, and Judge P. Kevin Castel fined $5,000 on 22 June 2023 (Schwartz, the signing lawyer and the firm).

Run the book's seven steps on anything that carries your name:

  1. Underline every specific claim.
  2. Open every source and find the sentence that supports the claim.
  3. Check a different way than the output was made.
  4. Redo the math in a spreadsheet.
  5. Match every quote word for word.
  6. Read it as the recipient would.
  7. Double-check the most embarrassing mistake it could contain.

Then spend 20 minutes running the list over anything AI-assisted you already sent last week.

Telling your boss you use AI

Chapter 15 argues that hiding AI use is now the riskier move, and that the new risks are sending unchecked output, which the book calls workslop, and leaking data. Say what the AI did, what you did, what changed and what you protect, and own every claim in what you send. The book prints a script for that conversation, two questions to ask about policy and sharing, and a one-line reply for the hallway. Before Friday, read the actual AI policy, check the training switch on your personal account, and book 15 minutes with your boss, bringing your before and after times from Day 1. In the book's words: "Being careless with data is a much faster way to lose your job than AI ever will be." The longer version is in should you tell your boss you use AI.

Where the plan comes from

The plan is Part IV, "Your First 72 Hours", of Don't Replace Me: A Survival Guide to the AI Apocalypse. It is not a prompt cookbook, and each chapter ends in one rule and one exercise. The book is on Amazon: Kindle ebook $2.99, paperback $14.99 (174 pages), hardcover $24.99. The introduction is a free download at dontreplace.me. There is nothing to buy after the book: no course, no community.