Lesson 2 of 8 · 6 min read

Pick the first target

Choose a first task that reads information and produces a draft, and know which tasks to leave for later.

Read and draft first

Your audit gave you a ranked list. This lesson narrows it to one task. The best first task reads information you already have and produces a draft you check. It does not send, pay, book or delete anything.

The strongest evidence for that rule comes from the UK Department for Business and Trade. From October to December 2024, it gave 1,000 staff a Microsoft 365 Copilot license and measured what happened. Written tasks saved the most time.

Other tasks went the other way. Excel analysis was slower and less accurate with Copilot. Scheduling and image generation took longer than doing them by hand. The department’s report also found no evidence that the time saved led to higher productivity.

Lesson 7 returns to that last finding. For now, it explains the shape of a good first target. It is writing-heavy, low-risk, and easy to check against what you would have written yourself.

Know where the frontier is

In 2023, researchers from Harvard Business School and Boston Consulting Group tested AI with 758 BCG consultants. On 18 tasks that suited the AI, consultants who used it finished 12.2% more tasks. They worked 25.1% faster, and their work was rated 40% higher in quality.

One task sat outside what the AI could do well. On that task, consultants working without AI reached the right answer 84% of the time. Consultants using AI were right 60 to 70% of the time. The AI output read well, and many consultants accepted it without checking.

Ethan Mollick, one of the authors, called this the jagged frontier. AI is strong on some tasks and weak on others that look similar. You cannot tell which side a task falls on by reading its description. You find out by testing it on a real example. That test is the AI-fit column in your audit.

The gains were also uneven across people. Consultants in the bottom half of the group improved by 43%, and the top performers improved less. Your own result depends on the task and on you, so the test on a real example decides.

Turn the sheet into a decision

Sort your tasks into three groups. High dread with good AI fit is your first target. High dread with poor AI fit means wait, or keep yourself in charge and use the assistant for parts of the task. Low dread means leave the task alone for now, even if it would be easy to automate.

Then apply three filters. Drop any task whose output leaves the company without review. Drop any task that moves money or changes a system other people rely on. Drop any task where you could not tell a good result from a bad one.

The third filter matters most. If you cannot judge the draft, you cannot check it. An automation you cannot check is a risk you have handed to the people downstream.

Unchecked output has a cost for other people. In a September 2025 survey by BetterUp and Stanford, 40% of US desk workers said they had received AI-generated “workslop” in the past month. That is work that looks finished but needs someone else to fix it. Your first automation should never add to that pile.

Good and poor first targets

Good first targets share a shape. They read from sources you already have and write something you would have written anyway. A weekly status report built from your notes and your team’s updates fits. So does a morning summary of your inbox and calendar, or follow-up notes drafted from a meeting transcript.

Other good examples depend on your role. A support lead might draft a weekly summary of ticket themes. A sales manager might draft notes before each client call from the account history. A recruiter might draft a weekly pipeline update for the hiring managers.

Poor first targets also share a shape. They need precise calculation, they touch other people’s calendars, or they send something outside the company. Spreadsheet analysis, meeting scheduling across several people, and replies sent straight to customers all fall here. Some of them may suit automation later, once you have checks you trust.

When you keep the task, split it

A task in the high-dread, poor-fit group still has parts an assistant can help with. Mollick describes two ways people work with AI. Centaurs divide a task, handing whole parts to the AI and keeping others. Cyborgs mix their work with the AI’s at a fine grain, sentence by sentence.

For a first project, the centaur pattern is safer. Give the assistant the part that reads and drafts, and keep the part that decides. For Dana’s reconciliation, that might mean an assistant drafts the list of mismatched lines, while Dana does the arithmetic and signs off.

Write the choice down

Write your choice and the reason in two sentences. Name the person who will check each draft. In most cases that person is you.

Keep the second and third tasks on the list. Once the first automation is stable, you will add the next one with half the effort, because the method stays the same.

Worked example · Larkfield Supply

Dana narrows the list

Dana’s audit sheet from Lesson 1 has six tasks. Dana applies the groups and the filters in order.

  1. Supplier payment approvals go first. A payment cannot be undone, so the task fails the money filter.That answers the question you guessed at: Supplier payment approvals take Dana 45 minutes a week and score 3 for dread. Guess: what rules them out as a first automation?
  2. Expense coding goes next. The codes feed the finance system, so a wrong code would reach another team without review.
  3. The spreadsheet reconciliation stays on the list but drops to the bottom.
    What would you do next?Why not automate the reconciliation first, given its high dread score?Think it through, then show the answer

    Its error cost is high, and it is spreadsheet analysis. The UK trial found AI slower and less accurate on Excel analysis. Dana’s own test run produced totals that did not match. It belongs in the high-dread, poor-fit group, where Dana stays in charge.

  4. That leaves the status report, the meeting follow-ups and inbox triage. All three read information Dana already has and produce drafts.
  5. Inbox triage scores lower on dread than the other two. Dana moves it to third place.
    What would you do next?The report and the follow-ups both pass every filter. Which should come first?Think it through, then show the answer

    The report. It has the highest dread score and a successful test run, and only the leadership team reads it. The follow-ups go to people outside Larkfield, so a weak draft has a wider audience.

Dana picks the Friday status report and writes two sentences: it is the task Dana dreads most, and every draft can be checked before the leadership team sees it. Dana checks each draft. Meeting follow-ups come next.

Starter prompts

  • Rank my audit

    Here is my audit table:

[paste your table]

Rank these tasks as first targets for an AI assistant that only produces drafts for me to check. Favor tasks with high dread, low error cost, and mistakes that can be undone. Exclude tasks that send anything outside my company, move money, or change a system other people rely on. Explain each ranking in one sentence.

  • Argue against my choice

    I plan to automate this task first: [describe the task]. Make the strongest case against starting here. List what could go wrong, what I might not notice, and which task from my list would be a safer start. Be specific to my task.

When not to use AI

  • Tasks where nobody would notice a mistake until it cost money.
  • Tasks whose correct output you cannot describe. Automating them hides the problem instead of solving it.
  • Tasks that help you learn a new role. Doing them yourself for a while builds the judgment you will later need to check the drafts.
Recall from memoryName the three filters that drop a task from your first-target list.Answer it without looking back, then show the answer

Drop any task whose output leaves the company without review. Drop any task that moves money or changes a system other people rely on. Drop any task where you could not tell a good result from a bad one.

Exercise · your project

Choose your first task and write down why.

  1. Group your audit tasks by dread and AI fit.
  2. Remove tasks that send outside the company, move money, or cannot be undone.
  3. Pick one task and write two sentences on why it comes first.
  4. Name the person who checks each draft, and list your second and third choices.

Sources