Skip to content
AIupick

Guide for office workers · 6 jobs

AI at work: meetings, numbers, reports and whole tasks

Most people use AI at work to write emails a little faster. The bigger gains are elsewhere: decisions that do not get lost after a meeting, numbers you can check, research you can defend, and whole tasks handed off with a clear brief.

  1. 01

    Turn every meeting into decisions and owners

    The summary is not what you need from a meeting; the decisions and who owns what are. Save one format that runs on every meeting, so nothing depends on remembering to ask, and later ask across meetings instead of searching notes.

    Example

    A 45-minute weekly project meeting. The template returns four decisions with timestamps, six tasks (two marked "deadline: not said") and one open question about the budget. Fixing two owners takes a minute; three weeks later, "what did we decide about the launch date?" finds it.

    How to do it

    1. 1Say at the start that the meeting is recorded and why; recording rules differ by country.
    2. 2Save the prompt on this page as a template in your notes tool: Recipes in Granola, or AI Skills in Fireflies, which can run on every meeting.
    3. 3Before you send the result, correct the owners and remove anything that was only suggested.
    4. 4Weeks later, ask across meetings in the tool's chat, for example "what did we decide about pricing in September?"

    Built in

    Prompt to copy

    From this meeting transcript, list:
1. Decisions: what was decided, who decided, and the timestamp.
2. Tasks: owner, deadline as said in the meeting ("not said" if missing), and the timestamp where the owner agreed.
3. Open questions nobody answered.
Leave out anything that was only suggested. If you are not sure something was decided, put it under open questions.

    Where it goes wrong

    • Transcripts get speakers wrong when people talk over each other. The timestamp lets you check who really agreed.
    • AI turns "we could" into "we will". Every decision should be something someone actually agreed to.
  2. 02

    Walk into every meeting prepared

    Most people prepare by scrolling old emails five minutes before the call. A brief built from your past meetings tells you who is coming, what was said last time and what is still open, before you join.

    Example

    A renewal call with a customer you last spoke to in June. The brief lists the three people joining, the discount you mentioned in June and the open question about where their data is stored. You add "can we agree on the term?" to the note and go in prepared.

    How to do it

    1. 1Connect Google Calendar to Granola. For external meetings it prepares a brief overnight (on its Business and Enterprise plans).
    2. 2Open the note for the meeting: the brief is at the top of the chat panel. Connect Gmail if you want recent emails in it too.
    3. 3Write the one or two things you need from this meeting into the note before you join.
    4. 4Without briefs, ask your notes tool's chat the prompt on this page across your past meetings with the same people.

    Built in

    Prompt to copy

    From my past meetings and notes with [names or company]: what did we agree, what did I promise, what is still open, and what did they care about most?
Give the meeting and date for each point. Say "not found" rather than guess.

    Where it goes wrong

    • A brief only knows what was recorded. If the last conversation was a phone call nobody noted, it is missing.
    • It repeats what was said, not whether it is still true. Check prices and dates before you quote them.
  3. 03

    Analyse a spreadsheet so every number can be checked

    A number the AI writes in a sentence may have been calculated or guessed, and you cannot tell which. Make it describe the data first, calculate only with formulas or code you can see, and check one number by hand.

    Example

    A sales export of 4,800 rows. The description finds 37 duplicate orders and one region spelled three ways. After fixing those, revenue per region is a formula in the sheet, and you check the largest region's total with a filter before it goes in the report.

    How to do it

    1. 1Ask it to describe the data before any analysis: rows, columns, empty cells, duplicates and odd values.
    2. 2Fix what it found, then ask your question. In Excel, Copilot works in the sheet itself; elsewhere, ask for the formula or code behind each number.
    3. 3Recalculate one key number yourself, for example a total with a filter.
    4. 4Ask for any chart together with the table of numbers it shows.

    Built in

    Prompt to copy

    Before any analysis, describe this data: number of rows, each column and its type, empty cells, duplicate rows and values that look wrong.
Then answer: [question].
For every number you report, show the formula or code that produced it. If a number needs an assumption, state it.

    Where it goes wrong

    • Duplicates and blank rows change totals without anyone noticing. That is why the description comes first.
    • Averages hide outliers. Ask for the median and the largest and smallest values too.
  4. 04

    Write a research brief you can defend

    Deep research tools read hundreds of pages, but they can give a vendor blog the same weight as a regulator. Approve the research plan before it runs, say which sources count, and check the three claims your decision rests on.

    Example

    Deciding whether to sell in Switzerland next quarter. The plan you approve names the federal statistics office and the regulator as sources and treats consultancy reports as claims. The three claims that matter are market size, the licensing rule and import duty, and you open each source.

    How to do it

    1. 1Start Deep Research in Gemini, or Researcher in Microsoft Copilot, and give it the decision the brief is for, not only the topic.
    2. 2Read the research plan Gemini proposes, or answer Researcher's follow-up questions, and name the sources that count, such as regulators, official statistics and company filings.
    3. 3When the report is ready, list the three claims your decision depends on and open their sources yourself.
    4. 4Cut it to one page: the recommendation, the evidence, and what would change your mind.

    Built in

    Prompt to copy

    I need a brief to decide [decision] by [date], for [audience].
Prefer official and primary sources: regulators, official statistics, company filings. Treat vendor marketing as a claim, not as evidence.
Structure: recommendation; the 3 facts it depends on, each with source and date; the main counterargument; what would change the recommendation.

    Where it goes wrong

    • Sources that look independent often copy one original. Trace your key claims back to where they started.
    • A market figure from three years ago is not this year's. Check the date of every number.
  5. 05

    Hand off a whole task, not a paragraph

    The gain has moved from drafting text to finished work: a deck, a spreadsheet or a report built from your files. Brief the agent like a new colleague, with the outcome, the audience, the files and what it must not do, and read its plan before it starts.

    Example

    "A 10-slide deck for Thursday's sales review from these three files: last quarter's numbers, the pipeline export and my notes. Do not email anyone." Its plan uses the pipeline export for the forecast slide; you point it to the numbers file instead, and then it starts.

    How to do it

    1. 1Use the agent mode of a tool you already pay for: Work in ChatGPT, Cowork in Claude, or Cowork in Microsoft Copilot.
    2. 2Write the brief with the prompt on this page and attach the files it should use.
    3. 3Follow its progress. ChatGPT Work and Copilot Cowork ask you to approve important actions; read what it is about to do before you say yes.
    4. 4Check the result like a colleague's first draft: facts, numbers, and anything it assumed.

    Built in

    Prompt to copy

    Task: [what to deliver], for [audience], by [when].
Use these files: [list]. Use nothing else unless you tell me where it comes from.
Format: [deck of n slides / spreadsheet with these columns / 2-page report].
Do not send, delete or share anything.
Before you start, tell me your plan in 5 steps and what you will assume.

    Where it goes wrong

    • An agent that misreads the brief does a lot of wrong work quickly. Its plan is where you catch that.
    • It acts with your access. Start with tasks that read and create files, not ones that send or delete.
  6. 06

    Stop explaining your job to AI every time

    If you paste the same background into every chat, the AI starts from zero every time. Put the standing material, your role, templates, style guide and good past examples, into one workspace once, and every chat starts from it.

    Example

    A consultant writing weekly client updates. The workspace holds the client's brief, the update template and three past updates the client liked. "Draft this week's update from these notes" now comes back in the right format and tone, and you only check the facts.

    How to do it

    1. 1Create a Project in Claude, or a notebook in Microsoft Copilot, for one area of work, such as one client or one report.
    2. 2Add what it should always know: templates, the style guide, and two or three examples of finished work you liked.
    3. 3Write the standing instructions with the prompt on this page.
    4. 4Replace files when they change; an old template in the workspace produces old work.

    Built in

    Prompt to copy

    Standing instructions for this workspace:
I am [role] at [kind of company]. The audience is usually [audience].
Write like the examples in this workspace: [length, tone, format].
Always use our template for [document type].
If a request conflicts with these files, say so instead of guessing.
If a request is unclear, ask me one question rather than filling the gap.

    Where it goes wrong

    • Company files in a personal account can break your company's rules. Use the account your company approves.
    • Examples get copied closely. Only add examples you would be happy to see repeated.

Rules for office workers

  • Use the tools and plans your company approves; a personal account is not covered by a company contract.
  • Keep customer data, contracts and passwords out of personal AI accounts.
  • Check whether your plan uses your chats for training: each tool page shows it under "Your data".
  • Recalculate any number you will act on, and read everything AI drafts in your name before it is sent.

Go deeper

Esc
↑ ↓ to moveEnter to open