What Work Tasks Can AI Automate? A One-Week Audit to Find Yours

The lists don't answer your question

Search for what work tasks AI can automate and you get lists: fifty tasks, a hundred tasks, sorted by department. They are accurate in a general way and useless on Monday morning, because they describe a job category and you have a calendar. "Email drafting" is on every list. Whether your emails are worth automating depends on which emails, how often, and what happens if one goes out wrong.

So we will answer the question differently. First, what "automate" means at three levels. Then the kinds of tasks current tools handle well and what you still have to check. Then a one-week audit with a four-question score that turns the general answer into a ranked list of your own tasks.

Three levels of "automate"

People use one word for three different things, and most disappointment comes from expecting level three when the task is ready for level one.

Level 1: assisted. You open a chat, write a prompt, get a draft, fix it. The tool saves time on each run, but you start every run by hand.

Level 2: a saved assistant. The instruction, reference files, and output format are stored in a project, a Gem, or a skill. You paste the input and get the result in the same shape every time. We describe how to build one in How to create a custom GPT for work.

Level 3: runs without you. A schedule or an event starts the task: a new form entry, an incoming email, 8 a.m. on Monday. You see the output, or only the exceptions.

Most office work that benefits from AI today sits at levels one and two. Level three suits a narrower set of tasks where a wrong output is cheap or easy to catch.

What AI handles well, and what you still check

Task type	Typical examples	Realistic level	What you still check
First drafts of routine messages	Follow-ups, status updates, polite declines	2	Facts, commitments, tone toward this person
Summaries of long material	Meeting transcripts, reports, long threads	2	That nothing important was dropped or softened
Extraction from messy text	Action items, dates, amounts from notes or PDFs	2 or 3	A sample of rows against the source
Reformatting and repurposing	Article to posts, notes to memo, table to slides	2	Claims that changed meaning in the shorter form
Sorting and tagging inbound items	Inquiries by type, feedback by theme	3	The "unsure" pile and a weekly spot check
Research digests	What changed this week on a topic, with links	3	Every source you plan to quote
Spreadsheet help	Formulas, cleanup steps, pivot logic	1	The result on rows where you know the answer
Recurring reports	Weekly numbers with a short commentary	2 or 3	The numbers themselves; commentary is the easy part

The last column matters as much as the first. Models still state wrong figures with full confidence and still cite pages that do not say what the summary claims. A task is a good candidate when checking takes far less time than doing.

What to keep for yourself

Some work should stay manual even when a tool could produce something that looks finished:

• Decisions you are accountable for: who to hire, what to charge, whether to sign.
• Messages where the relationship is the point. A condolence note, a hard conversation with a client, feedback to a direct report.
• Anything where an error is expensive and silent, such as figures that go into a contract or a filing and that nobody downstream will recheck.
• Thinking you need to do in order to understand the problem. If writing the strategy memo is how you find out what you think, a generated memo skips the useful part.

AI is still helpful here as a sparring partner: ask for the strongest objection to your draft. That counts as assistance, and you stay the author.

The four-question score

For each recurring task, give 0, 1, or 2 on four questions.

Question	0	1	2
How often?	Less than monthly	Monthly	Weekly or more
Is the input the same shape each time?	Different every time	Similar	Nearly identical
Can you describe "done" in three sentences?	No	Partly	Yes
Can you check the result quickly?	Checking takes as long as doing	Some effort	Under two minutes

Read the total this way: 7 or 8, build a saved assistant and consider a schedule. 5 or 6, keep it at level one with a good saved prompt. 4 or less, do it yourself.

Worked example: a one-week audit

For five working days, keep a plain list of what you did, one line per task, with rough minutes. Do not tidy it. On Friday, paste the list into ChatGPT, Claude, or Gemini with this prompt:

Below is a log of my work tasks for one week, with rough minutes.
I am a [role] at a [type of company].

1. Group the log into recurring tasks. Ignore one-offs.
2. Score each recurring task from 0 to 2 on four questions:
   - Frequency (0 = less than monthly, 1 = monthly, 2 = weekly+)
   - Input sameness (0 = different every time, 2 = nearly identical)
   - Definable result (0 = I could not describe "done", 2 = I could
     in three sentences)
   - Ease of checking (0 = checking takes as long as doing,
     2 = under two minutes)
3. Return a table: Task | Minutes per week | Four scores | Total |
   Your reasoning in one line.
4. If you had to guess a score because my log lacks detail, mark it
   with "?" and ask me about it after the table.
5. Do not recommend tools. Do not estimate time savings.

LOG:
[paste the week here]

Here is what the table looks like for a typical marketing manager's week. This is an illustration, not a client's data:

Task	Frequency	Input	Done	Check	Total
Weekly performance report from the ad platform export	2	2	2	2	8
Call notes into a follow-up email and action list	2	2	2	2	8
Blog post into three social posts	2	2	1	2	7
Replies to inbound partnership requests	2	1	1	2	6
Monthly budget reallocation	1	1	0	0	2
Positioning for a new product	0	0	0	1	1

Two things usually need correcting in the model's first pass. It scores "input sameness" too high, because your log says "client emails" and it cannot see that every client email is different. Tell it which inputs vary. It also treats "definable result" as a yes whenever a task has a name. Push back: try to write the three sentences yourself, and if you can't, the score is 0 or 1.

That is also why the prompt forbids tool recommendations and savings estimates. Asked for either, the model will produce confident numbers with nothing behind them.

From the audit to a first automation

Take the top-scoring task only. Build it at level two, run it by hand for two weeks, and move to level three after it has been right several times in a row.

What the common tools allow, as of October 2026:

• ChatGPT scheduled tasks. Available on every plan. Free and Go accounts can keep three active tasks that run at most once a day; Plus allows five and hourly runs. Tasks triggered by an event, such as a new email, need Plus or higher.
• Claude Cowork scheduled tasks. Available on all paid Claude plans. You describe the task once and it runs on a schedule and delivers a finished output such as a briefing or report.
• Zapier. The free plan covers 100 tasks a month and two-step Zaps (one trigger, one action).
• Make. The free plan includes 1,000 credits a month.
• n8n. The self-hosted Community Edition is free; the cloud version is paid.

A reasonable first level-three task is a Monday digest: a scheduled task that searches for news on two or three topics you follow and returns five items with links. If it misses something, nothing breaks.

Counting whether it saved time

Keep the arithmetic honest and count the checking.

Monthly saving = (minutes by hand − minutes to run and check) × runs per month.

If a report took 25 minutes by hand and now takes 8 including the check, four runs a month return 68 minutes. An hour of setup pays back within the first month. If checking takes 20 of those 25 minutes, the same task returns 20 minutes a month and is not worth a schedule. Use your own numbers from the log; ours are only there to show the formula. Once two or three tasks pass this test, the next question is order: A 90-Day AI Adoption Plan for Professionals Who Already Tried lays out which week to add what. If a task has to move data between apps rather than live inside a chat, compare the tools first in Zapier vs Make vs n8n for Non-Developers: An Honest Comparison.

Where this breaks

Unchecked level three. An automation that emails clients without review will, at some point, email a client something wrong. Keep a human approval step on anything that leaves the company.

Confidential inputs. Meeting transcripts, HR notes, and client files often contain personal data or NDA material. Consumer plans are the wrong place for them. Check your employer's policy before the audit, not after.

Automating a bad process. If nobody reads the weekly report, generating it faster saves nothing. The audit sometimes shows a task should be dropped.

Maintenance. Scheduled tasks and no-code flows fail quietly when an app changes a field name or a login expires. Put a monthly check in your calendar.

The prompt, not the plan. If results are weak at level one, a schedule will deliver weak results on time. Fix the instruction first; the structure is in How to write better AI prompts for work, and the settings that apply to every chat are in ChatGPT custom instructions examples for work.

Where the practicum picks this up

The week audit is part of session four of our two-week practicum, together with an assistant for one recurring task, no-code chains as an advanced block, a map of your personal AI system, and a way to count saved hours. How the practicum is set up for particular roles is described on our pages for managers and for marketers.