Lesson 1 of 8 · 6 min read

Audit your week

Time your recurring tasks and score them, so you automate the right one first.

Narrated with an AI voice. Captions are available in the player.

Transcript

Video transcript

This video uses an AI-generated voice.

Dana runs operations at Larkfield Supply, a forty-person office-supplies distributor.

Dana wants an AI assistant to take one recurring task off the list.

So Dana picked the task dreaded most: the monthly reconciliation.

The assistant’s totals did not match the ledger.

Finding out why took the rest of the afternoon.

Dana’s verdict: “I picked the job I hate, not the job it can do.”

Think of the first task you would hand over. How did you choose it?

Here is a quick guess before we start.

Inbox triage takes Dana twenty minutes, every morning.

The Friday status report takes forty-five minutes, once a week.

Which costs Dana more time each week? Make your guess.

Here is what Dana did instead: a one-week audit before choosing anything.

Three steps: list the tasks, time them with a clock, and score them.

On Monday, Dana lists every task that repeats.

The status report, inbox triage, supplier payments, the monthly reconciliation.

Then Dana times each run with a clock instead of guessing.

People misjudge this. In a 2025 METR study, developers believed AI tools made them faster.

The clock showed them slower. METR now marks that number out of date, but the gap is the lesson.

On Friday, Dana multiplies runs by minutes.

The status report: forty-five minutes, once a week. Forty-five.

Inbox triage: twenty minutes, five mornings a week. One hundred.

So the answer to the guess is inbox triage.

Small daily tasks hide, because no single run feels long.

Last, Dana scores each task.

Dread runs from one to five. So does error cost, for what a mistake would cost.

Can a mistake be undone before it does harm?

And one real test: give the assistant last week’s inputs, and judge the draft.

Pause the video and explain in one sentence.

Supplier payments score only three for dread. Why does Dana rule them out?

Because a payment cannot be undone once it is sent.

The reconciliation is out too, because it failed its test.

Inbox triage costs the most time, but its dread is middling.

The status report wins: dread five, low error cost, and a draft Dana checks first.

Before the audit, Dana chose by feel and lost an afternoon.

After it, the choice came from numbers Dana measured.

Now, from memory.

Before you choose a task to hand over, what do you measure, and how?

Each task’s weekly minutes: runs times minutes, timed with a clock.

Then you score dread, error cost, and whether a mistake can be undone.

Your turn. List the tasks you repeat this week, and time every run.

The lesson page has the audit sheet and two prompts that build it with you.

Next lesson, you turn your sheet into one choice.

Start with a list, not a tool

Most people pick their first automation by feel. They choose the task that annoyed them most recently, or the one a colleague automated last month. That choice often goes wrong in one of two ways. The task is too rare to repay the setup time, or it carries a risk that a draft cannot carry.

This lesson replaces the feeling with a short audit. You list the work you repeat, time each task, and score it in six columns. The audit takes about a week of light note-taking. At the end you have a ranked list, and the next lesson turns it into one choice.

The audit also gives you a baseline. In Lesson 7 you time the same task again with the automation running. Without the first measurement, you cannot tell whether the automation saved time or only moved it somewhere else.

Time tasks instead of estimating them

People misjudge how long their work takes, and they misjudge how much AI changes it. METR, a research group, ran a controlled study in early 2025 with 16 experienced software developers working on 246 real tasks.

Before the study, the developers expected AI tools to make them 24% faster. Afterward they believed the tools had made them about 20% faster. The clock showed the opposite. With AI tools, they took 19% longer.

METR updated its work in February 2026 and now marks that result as out of date. Its newer data is noisier, and METR says the new signal is unreliable. The finding that lasts is not the 19%. It is the gap between what people believed and what the clock recorded.

So time each task with a clock. Note when you start, when you finish, and any interruptions. One week of real numbers is worth more than a month of guesses, and it costs you a minute a day.

Score each task in six columns

Stanford researchers built WORKBank, a survey of 1,500 workers about 844 tasks across 104 occupations. They asked workers which tasks they wanted AI to take on. They asked AI experts which of those tasks current AI could do. Workers rated 46.1% of the tasks as good candidates for automation.

The reasons workers gave are useful for your audit. The reason given most often was freeing time for higher-value work, at 69.4%. Tedious, repetitive work came next, at 46.6%. Stress came third, at 25.5%. The main reason to keep a task was distrust of AI accuracy, at 45.0%.

WORKBank also asked how much involvement workers wanted to keep, on a five-level Human Agency Scale. In 47 of the 104 occupations, the most popular answer was an equal partnership between the worker and the AI. This course takes the same approach. The assistant drafts, and you decide.

Frequency is the first column: how many times a week the task runs. Minutes per run is the second, taken from your clock, not your memory. Dread is the third, scored from 1 to 5. A 5 is the task you put off until the last possible moment.

Error cost is the fourth column, also scored from 1 to 5. A 5 means a mistake would cost money, trust or a customer. Reversible is the fifth: can a mistake be caught and undone before it does harm? Mark it yes or no.

AI fit is the sixth column. Give your assistant one real instance of the task and judge the result. Mark it good, partial or poor. Do not guess this column. Tasks that look alike can produce different results.

Run the AI-fit test fairly. Use a real instance from last week, with the real inputs. Give the assistant the same information you had, and no more. Then compare its result with what you sent. A fair test takes ten minutes and tells you more than any article about what AI can do.

Add one note beside each task: who receives the output. A draft that only you read carries less risk than one that goes to a customer or a supplier. Lesson 2 uses this note to filter the list.

No published instrument measures dread. The dread column borrows from the WORKBank reasons. The frequency, error-cost and reversibility columns come from practitioner scoring guides. Treat the sheet as a working tool, not a validated test.

Add up the minutes

Multiply frequency by minutes to get the weekly cost of each task. A task that takes 45 minutes once a week costs 45 minutes. A 10-minute task you do twice a day costs 100. People often overlook the second kind, because no single run feels long.

Practitioner guides suggest starting with tasks that run more than three times a week and take more than 15 minutes each run. Those thresholds come from vendor blogs, not research. Use them as a first filter, and let your own numbers decide.

Worked example · Larkfield Supply

Dana audits a week at Larkfield Supply

Dana timed six recurring tasks over one week and scored each one. Each line below is one row of the audit sheet.

  1. The Friday status report runs once a week and takes 45 minutes. Dread 5, error cost 2, reversible as a draft. A test run produced a usable first draft from the week’s notes.
  2. Supplier payment approvals run three times a week at 15 minutes each. Dread 3, error cost 5. A payment cannot be undone once it is sent.
  3. Meeting follow-up emails run four times a week at 10 minutes each. Dread 4, error cost 2, reversible as drafts.
  4. The spreadsheet reconciliation runs once a month and takes two hours. Dread 5, error cost 4. A test run produced totals that did not match the ledger.
  5. Inbox triage runs every morning for 20 minutes. Dread 3, error cost 2, reversible.
  6. Expense coding runs twice a week at 15 minutes each. Dread 4, error cost 3, reversible until the codes are submitted.
    What would you do next?Which task should Dana automate first?Think it through, then show the answer

    The Friday status report. It has the highest dread, a low error cost, and a draft Dana can check before anyone sees it. Its test run worked. Payment approvals cannot be undone, and the reconciliation failed its test run.

The weekly minutes come to 45 for the report, 45 for approvals, 40 for follow-ups, about 30 for the reconciliation, 100 for triage and 30 for coding. Inbox triage costs the most time, but its dread score is middling. Dana starts with the report and puts follow-ups second.That answers the question you guessed at: Inbox triage costs Dana more minutes each week than any other task. Guess: is it the task Dana automates first?

Starter prompts

  • Audit interview

    I want to find recurring tasks in my work that an AI assistant could draft for me. Interview me about my last working week. Ask one question at a time. For each task I mention, ask how often it happens, how long it takes, how much I dislike it on a scale of 1 to 5, what a mistake would cost on a scale of 1 to 5, and whether a mistake can be undone. When I say I am done, give me the answers as a table.

  • Turn notes into the audit sheet

    Here are my notes on recurring tasks from this week:

[paste your notes]

Turn them into a table with these columns: task, runs per week, minutes per run, weekly minutes, dread (1-5), error cost (1-5), reversible (yes or no). Calculate weekly minutes as runs per week times minutes per run. Sort by weekly minutes, highest first. Do not invent numbers. If a value is missing, write "missing".

When not to use AI

  • Tasks whose value is your own judgment or relationship, such as feedback to a colleague. A draft from an assistant changes what the other person receives.
  • Tasks that move money, sign contracts or make legal commitments. Leave them out of a first project, whatever their score.
  • Tasks you do once a quarter or less. Setting up and checking the automation would cost more than doing the task.
Recall from memoryName the six columns of the audit sheet, and the one you must test rather than guess.Answer it without looking back, then show the answer

Frequency, minutes per run, dread, error cost, reversible, and AI fit. You fill in AI fit by giving your assistant one real instance of the task and judging the result. Minutes alone do not decide: inbox triage costs Dana the most time, but the report ranks first.

Exercise · your project

Time and score the recurring tasks in your own week.

  1. List every task you repeat at least once a week. Aim for eight to ten.
  2. For one week, note the start and end time of each run.
  3. Score dread and error cost from 1 to 5, and mark whether a mistake can be undone.
  4. Test AI fit once: ask your assistant to do one real instance, and judge the result.
  5. Multiply runs by minutes, and sort the sheet by weekly minutes.

Sources