How to Save Time at Work With AI: The 10 Tasks to Hand Over First, and What Each One Really Saves (2026)

The quickest way to save time at work with AI is not to ask it to do your job. It is to hand it the ten jobs inside your job that quietly eat the week: the inbox, the meeting notes, the first draft, the 60-page report nobody has time to read, the spreadsheet formula you look up every single month. You keep the parts that need your judgement.

The research on this is now unusually good, and it contains a surprise. In controlled experiments, AI cut the time people needed for a single task by between 25% and 74%. Professionals writing with ChatGPT in a study published in Science took 40% less time and produced better work. Yet when economists measure whole working weeks, the savings shrink to between 1.6% and 2.8% of all hours worked. And when one UK government department gave Microsoft 365 Copilot to a thousand staff, it found real time savings but no robust evidence that anyone had become more productive.

That gap, between big savings per task and small savings per week, is what this guide is about. Most of it comes from three habits: handing AI the wrong tasks, skipping the two minutes of briefing that make its output usable, and then spending the saved time checking and fixing what it got wrong. Workday’s research puts that last loss at nearly 40% of the time AI saves.

So here are the ten tasks worth handing over first, in the order I would start with them; what the research says each one really saves; a copy-and-paste prompt for every one; what you still have to do yourself; and, because I work from Germany, what GDPR, the EU AI Act and German law say about pasting your work into an AI tool. Every figure below was checked against the original study, statistic or legal source as of 25 October 2026.

Independence note: No company, vendor or lab mentioned in this article sponsored, reviewed or approved it, and it contains no affiliate links. Nothing here is legal advice; Part 4 explains where to get that.

---

TL;DR — The 60-Second Version

• Start with the boring, frequent, text-heavy work. Email triage, meeting recaps, first drafts and document summaries recur every week, are cheap to check and have the strongest evidence behind them.
• Per task, the savings are big. Writing tasks took 40% less time with ChatGPT in a Science study; a small coding task took 55.8% less with GitHub Copilot; catching up on a missed meeting took 11 minutes instead of 43.
• Per week, they’re small. For now. Danish chatbot users saved 2.8% of their working hours; across the US workforce the St. Louis Fed measured 1.6%. The difference is tasks AI can’t do, checking time and rework.
• Email is the best-measured win. In a six-month field experiment with 7,137 knowledge workers, people who used Copilot spent two fewer hours a week on email. Their meetings didn’t get any shorter.
• Some tasks get slower. In a UK government Copilot trial, drafting documents saved 1.3 hours per task, but presentations saved nothing and scheduling took longer than doing it by hand.
• Rework is the hidden tax. Workday found nearly 40% of AI time savings are lost to fixing output. 40% of US desk workers received AI “workslop” last month, and each piece took about two hours to deal with.
• Never delegate judgement. On a task just beyond GPT-4’s abilities, BCG consultants who used it were 19 percentage points less likely to get the right answer.
• In Germany, the rules are manageable. Use business plans with a data processing agreement for personal data, announce AI transcription (§ 201 StGB), talk to your works council early, and train your staff: the EU AI Act still expects AI literacy.
• Measure net time, not the feeling. Experienced developers in a 2025 trial believed AI made them 20% faster. It made them 19% slower.

---

The 10 Tasks at a Glance

#	Task	Hand to AI	Keep for yourself	What the evidence says	How far to let go
1	Email	Triage, thread summaries, first replies	Sending; anything about money, contracts or complaints	Copilot users spent 2 fewer hours a week on email	AI drafts, you send
2	Meeting notes	Transcript, recap, action items	Consent, decisions, sensitive meetings	Missed meeting caught up 3.8× faster; 0.7 h saved per recap	Hand over, with consent
3	First drafts	Structure, first draft, rewrites	The argument, the facts, your voice	40% less time and 18% better quality; 1.3 h saved per document	AI drafts, you edit
4	Research and summaries	Summaries, extraction, comparisons	Verifying claims and citations	Report summarised in 13 minutes instead of 42	AI reads, you verify
5	Spreadsheets and data	Formulas, cleaning, analysis code	Conclusions and the numbers that matter	Copilot was slower and less accurate on an Excel test	AI builds, you check
6	Customer replies	Suggested answers, FAQ drafts	Complaints, refunds, promises	+14% issues resolved per hour; +34% for new agents	AI suggests, a person approves
7	Presentations	Storyline, slide text, speaker notes	The message, the numbers, the design	No time saved on finished decks in the UK trial	AI outlines, you build
8	Translation	First-pass translation, glossary	Legal and certified texts, final read	+60% words per hour, German to French	AI translates, you review
9	Admin and data entry	Extraction from PDFs and receipts, form drafts	Approvals, payments, calendars	Scheduling took longer with AI than without	AI extracts, you approve
10	Small automations	Formulas, scripts, workflow glue	Access rights, production systems	55.8% faster on a coding task; experts slower in 2025	Build with AI, test before you trust

---

Table of Contents

• Part 1 — Where Your Working Week Actually Goes
• Part 2 — How Much Time Does AI Really Save? The Evidence, Honestly
• Part 3 — How to Choose What to Hand Over, and How to Hand It Over
• The Ten Tasks: Where to Start and How Far to Let Go
• Task 1 — Email: Triage, Thread Summaries and First Replies
• Task 2 — Meeting Notes, Summaries and Action Items
• Task 3 — First Drafts: Reports, Proposals, Posts and Briefs
• Task 4 — Research and Reading: Summarising Long Documents
• Task 5 — Spreadsheets and Data Analysis
• Task 6 — Customer Support Replies and FAQs
• Task 7 — Presentations and Slide Decks
• Task 8 — Translation and Localisation
• Task 9 — Admin: Forms, Invoices, Data Entry and Scheduling
• Task 10 — Small Automations: Formulas, Scripts and Workflows
• Part 4 — Privacy and the Law: What You Can Paste Where in Germany and the EU
• Part 5 — Which AI Tools to Use: Built-In, Standalone or Local
• Part 6 — How to Measure Whether You Really Save Time at Work With AI
• Part 7 — Your First 30 Days: A Rollout Plan
• Part 8 — Ten Mistakes That Eat the Time You Saved
• Frequently Asked Questions
• The Bottom Line

---

Part 1 — Where Your Working Week Actually Goes

Before deciding what to hand to AI, look at where the hours go. Microsoft has the best data on this, because it can measure, anonymously and in aggregate, how long people spend in Outlook, Teams, Word and Excel.

Its 2023 Work Trend Index found that people spent 57% of their time in Microsoft 365 communicating: meetings took 23%, chat 19% and email 15%. The other 43% went on creating: Excel 18%, Word 10%, PowerPoint 8% and OneNote 7%. The heaviest email users, the top quarter, spent 8.8 hours a week on email alone, and 68% of people said they didn’t get enough uninterrupted focus time during the day.

Two years later the picture was more fragmented, not less. Microsoft’s June 2025 report on what it calls the “infinite workday” counted 117 emails and 153 Teams messages per person per day, and an interruption (a meeting, an email or a chat notification) every two minutes, 275 times a day. 57% of meetings were ad hoc calls with no calendar invite, and 40% of people were checking email before 6 a.m. The telemetry behind those numbers excludes EU customers, so European offices may look slightly different, but there’s little reason to think they look calmer.

Then there’s the time spent simply looking for things. Atlassian’s 2025 State of Teams survey of 12,000 knowledge workers and 200 executives found people spend about a quarter of their working week searching for information.

Put those together and the opportunity has an obvious shape. Most of the week is text in, text out: reading, summarising, replying, documenting and finding. That is exactly what large language models are good at. The “creating” half (the spreadsheet model, the board deck, the analysis) is where they are less reliable, and that is why the order of the ten tasks below isn’t random.

The jagged frontier: why “AI is good at writing” isn’t enough

The most useful idea in this whole field comes from a 2023 experiment with 758 consultants at Boston Consulting Group, run by researchers from Harvard, MIT, Wharton and Warwick and since published in Organization Science. On realistic consulting tasks that sat inside GPT-4’s abilities, consultants using it completed 12.2% more tasks, 25.1% faster, with more than 40% higher quality in the original working paper.

Then the researchers gave them a task chosen to sit just outside those abilities: it meant combining figures from a spreadsheet with clues in interview notes. Without AI, 84.5% of consultants got it right. With AI, they were on average 19 percentage points less likely to reach the correct answer. The AI’s analysis sounded as convincing as ever; it was simply wrong more often, and many consultants went along with it.

The researchers called the boundary the jagged technological frontier: AI is excellent at some tasks and poor at others that look almost identical from the outside, and intuition is a bad guide to which is which. That’s the case for a deliberate list of tasks rather than a vague resolution to “use AI more”.

Five tests for a task you can hand over

1. It recurs. Weekly or more often. The setup cost of choosing a tool and writing a good brief only pays off on repetition.
2. It’s mostly text or structured data. Emails, notes, documents, tables. Anything that needs a phone call, a site visit or a relationship isn’t on the list.
3. You can check the result faster than you can produce it. This is the test most people skip. If verifying an AI summary means reading the whole report anyway, you’ve saved nothing.
4. A mistake is cheap and caught before it leaves your hands. A clumsy internal draft costs nothing. A wrong number in a customer quote costs a lot.
5. The data is allowed in the tool you’re using. If it isn’t, stop. Part 4 explains how to tell.

A task that passes all five belongs at the top of your list. A task that fails test 3 or 4 can still use AI, but only as a drafting aid, with a person making the final call.

---

Part 2 — How Much Time Does AI Really Save? The Evidence, Honestly

There are two kinds of evidence, and they tell very different stories. Controlled experiments give one group AI and an otherwise identical group none, then time a single task. Workforce studies look at what happens to whole jobs over months. Both are right; they’re measuring different things.

On a single task: a quarter to three-quarters faster

• Writing. In a 2023 MIT experiment published in Science, 453 college-educated professionals did realistic writing tasks for their own occupations. Those given ChatGPT took 40% less time and their work was rated 18% higher in quality. The gap between weaker and stronger writers narrowed.
• Coding. In a 2023 GitHub experiment, developers asked to build a small web server in JavaScript finished 55.8% faster with GitHub Copilot.
• Consulting. The BCG consultants above: 12.2% more tasks and 25.1% faster, inside the frontier.
• Customer support. A study of 5,179 support agents at a software company, by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in the Quarterly Journal of Economics in 2025, found an AI assistant raised the number of issues resolved per hour by 14% on average and by 34% for new and lower-skilled agents, with little effect on the most experienced.
• Office work. Microsoft’s first Copilot study, in November 2023, found users 29% faster across searching, writing and summarising. A first draft took 8 minutes 12 seconds instead of 13 minutes 48. Catching up on a missed meeting from an AI recap took 11 minutes instead of 43, though recap readers remembered slightly fewer details (11 of 15 against 12).

There’s one important exception. In early 2025 the research group METR ran a randomised trial with 16 experienced open-source developers working on 246 real tasks in codebases they knew well. Beforehand, they expected AI to cut their time by 24%. It made them 19% slower, and afterwards they still believed it had made them about 20% faster. A rerun in late 2025 pointed the other way, with tasks taking roughly 4% to 18% less time with AI, but the confidence intervals still included a slowdown, and METR itself described its new data as very weak evidence. The lesson isn’t that AI slows experts down. It’s that people are bad at judging how much AI helps them. (I went through the coding evidence in more depth in State of AI 2026.)

Across a whole week: 1.6% to 2.8% of working hours

Now zoom out from a single task to a whole job.

• Denmark. Economists Anders Humlum and Emilie Vestergaard linked surveys of about 25,000 workers in 11 occupations exposed to AI with official employer and earnings records. Chatbot users saved 2.8% of their working hours on average, and two years after ChatGPT arrived the authors found no significant effect on earnings or recorded hours in any of those occupations.
• The United States. The Federal Reserve Bank of St. Louis found that 37.4% of working-age adults used generative AI at work by August 2025. Pooling its 2025 surveys, the time saved came to 1.6% of all working hours across the whole workforce, users and non-users together. (An earlier and widely quoted figure of 5.4% of hours for users, from late 2024, was later revised by the Fed.)
• The UK civil service, self-reported. In a three-month cross-government trial with about 20,000 civil servants, Microsoft 365 Copilot users reported saving 26 minutes a day, roughly two weeks a year. 17% reported no saving at all; 14% reported an hour or more.
• One UK department, measured more carefully. The Department for Business and Trade’s evaluation of 1,000 Copilot licences (October to December 2024, published August 2025) found that time savings were “evident, albeit in small quantities”, and no robust evidence that they translated into higher productivity. 22% of its participants said they had caught Copilot hallucinating.
• Email, measured by software rather than surveys. The largest field experiment so far, covering 7,137 knowledge workers in 66 firms over six months and run by economists from Microsoft Research and Harvard Business School, found that people who used Copilot spent two fewer hours a week on email, 17% less than before, and worked a little less outside normal hours. Their meeting time didn’t change, and neither did the mix of tasks they did.

Where the time goes missing

Four things explain the gap between a 40% saving on one task and a 2.8% saving on a week.

1. Most of a job isn’t delegable yet. Anthropic’s January 2026 Economic Index found Claude succeeded on 67% of the tasks people brought to the Claude.ai app and 49% of those sent through its API, and success fell as tasks got longer, from about 60% for work a person could do in under an hour to about 45% for work taking five hours or more. Anthropic’s separate estimate that AI cuts the time of those tasks by about 80% comes with the company’s own caveat: it can’t see the time people spend checking the output.

2. Rework. Workday surveyed 3,200 employees who actively use AI in November 2025. 85% said AI saved them one to seven hours a week, but nearly 40% of that was lost to rework, correcting errors and rewriting content, and only 14% consistently came out clearly ahead. Researchers at BetterUp Labs and the Stanford Social Media Lab have a name for the cause: workslop, AI-generated work that looks finished but lacks substance. In their September 2025 survey of 1,150 US desk workers, 40% had received some in the past month, and each instance took about two hours to deal with.

3. Our sense of time saved is unreliable. METR’s developers felt faster while measuring slower. Most of the large “hours saved” numbers you’ll see, including the UK government’s 26 minutes, are self-reported.

4. Some tasks get slower. The DBT trial’s diary study is the most useful dataset here, because it breaks Microsoft Copilot time savings down task by task. Drafting written documents saved 1.3 hours per task, summarising research 0.8 and meeting transcripts and summaries 0.7. Presentations saved nothing. Scheduling and image generation took longer than doing them without AI.

At the level of whole organisations, the picture is even noisier. MIT’s NANDA initiative made headlines in August 2025 by reporting that 95% of enterprise generative AI efforts had produced no measurable return. That figure, though, rests on 52 interviews, a survey of 153 people and an analysis of about 300 public initiatives, so treat it as a warning about pilots rather than a measurement of your inbox.

What people who save time with AI do differently

Microsoft’s 2026 Work Trend Index surveyed 20,000 people who use AI at work in ten countries, Germany among them, in spring 2026. 86% said they treat AI output as a starting point rather than a final answer, and half ranked quality control of AI output among the most important human skills. The group Microsoft calls “Frontier Professionals”, about one in six of those surveyed, stood out mostly for their habits: 53% pause to decide whether a task is better done by AI or by a person (against 33% of the rest), and 43% deliberately do some work without AI (against 30%).

That’s the whole method in one sentence: choose the task deliberately, brief it properly, and check it in proportion to the risk. The rest of this guide is how.

---

Part 3 — How to Choose What to Hand Over, and How to Hand It Over

Most advice on how to use AI at work starts with the tools. Start with the tasks instead, and decide two things for each one: whether to hand it over at all, and how far to let go.

The delegation matrix

Score every recurring task on two questions: how often do you do it, and what would a mistake cost if one got through? Then put it in one of four boxes.

• Hand over now: frequent, and mistakes are cheap to catch. Meeting recaps, internal drafts, summaries of documents you’ll read anyway. This is where most of the time is.
• AI drafts, you decide: frequent, but mistakes are costly. Customer replies, numbers in spreadsheets, anything with your name on it. AI does the first 80%; you own the last 20%.
• Batch it: occasional, and mistakes are cheap. Slide outlines, translations for internal use, one-off research. Use AI when it comes up; don’t build a system around it.
• Keep a human in charge: rare and costly, like contracts, hiring decisions, anything legal or medical, automations with write access to real systems. AI can help you think; it shouldn’t act.

The positions in the matrix are typical, not universal. A translator’s translations and a lawyer’s research belong in very different boxes from a marketing manager’s.

Four levels of hand-off

It also helps to decide, task by task, how far you let go.

Level	AI does	You do	Example
1 — Assist	Suggests, explains, critiques	All the work	“What’s missing from this plan?”
2 — Draft	Produces a first version	Edit, verify, approve	Proposal drafts, email replies
3 — Delegate	Does the task end to end	Spot-check, handle exceptions	Meeting recaps, receipt extraction
4 — Automate	Runs on a schedule or trigger	Design, test, monitor	A weekly report built from a spreadsheet

Most people should spend most of their time at levels 2 and 3. Level 4 is where the biggest savings are, and also where most of the failures happen: Gartner expects over 40% of “agentic AI” projects to be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls.

The five-part brief

Nearly every bad AI output traces back to a bad brief. A good one takes two minutes and has five parts:

1. Goal. What “done” looks like, and what the reader should do after reading it.
2. Source. Paste or attach the material. Never ask AI to work from memory when facts matter.
3. Audience and tone. Who it’s for, how formal, which language and, in German, Sie or du.
4. Rules. Length, what must be included, what must be avoided, and “don’t invent names, numbers or dates”.
5. Format. Bullets, a table, an email, slide titles, plus an instruction to flag anything it isn’t sure about.

That last instruction is the cheapest quality control there is. Models are far more willing to say “I’m not sure” when you explicitly ask them to.

The hand-off loop

The loop has five steps: brief, draft, check, ship, save.

• Check means three questions, asked in proportion to the risk. Are the facts and numbers right? (Open the source.) Is there a promise, or a tone, you wouldn’t put your name to? Is there anything in it that shouldn’t leave the building?
• Ship means you send it — never the tool.
• Save is the step most people skip. When a prompt works, keep it. Most assistants now let you store reusable instructions or project folders, and a shared document of your team’s best prompts works just as well. The second time you run a saved brief, the setup cost is zero, and that’s where the compounding savings come from.

---

The Ten Tasks: Where to Start and How Far to Let Go

These are the tasks to delegate to AI first, in the order I’d start with them. Tasks 1 to 4 are safe for almost anyone this week; 5 to 8 need a little more care; 9 and 10 need some setup. Each section follows the same pattern: why the task eats your time, what the evidence says, how to hand it over, a prompt to copy, and the human check you can’t skip.

A note on the prompts: they’re written in English, but they work in any language. Write the prompt in the language you want the answer in, and replace everything in square brackets with your own details.

---

Task 1 — Email: Triage, Thread Summaries and First Replies

Why it eats your week. 117 emails a day is the average in Microsoft’s 2025 data, and the heaviest email users in its 2023 data spent 8.8 hours a week in their inbox. Most of that time goes on reading and deciding, not on typing.

What the evidence says. Email has the strongest real-world evidence of any task in this guide. In the 7,137-worker field experiment, people who used Copilot spent two fewer hours a week on email, a 17% cut, and less time working after hours. Look at where that saving comes from, though. In the UK DBT trial’s diary study, writing an email saved only about 12 minutes per email (0.2 hours). In its observed test, Copilot users took the same seven and a half minutes to write an email as everyone else, although their emails were rated more accurate and better written.

So the real value of AI for email isn’t writing faster. It’s triage and summaries: knowing within two minutes which ten of today’s emails actually need you, and what a forty-message thread actually decided.

How to hand it over

1. Triage once or twice a day. Ask the assistant in your mail client to sort unread mail into “reply today”, “reply this week”, “FYI” and “ignore”, with one line on what each sender wants.
2. Summarise long threads first. Ask for the decision, the open questions and who is waiting on whom before you read a word of the thread itself.
3. Draft replies from your bullet points. You decide what to say; AI decides how to say it politely in a hundred words.
4. Turn repeat replies into templates. If you answer the same question every week, write the perfect answer once with AI’s help and save it.
5. Never let it send. Drafts only, always.

Prompt to copy

You are my email assistant. Below are today's unread emails.

1. Sort them into four groups: Reply today / Reply this week / FYI only / Ignore.
2. For each "Reply today" email, give one line on what the sender wants
   and the deadline, if there is one.
3. Draft a reply to each "Reply today" email: under 120 words, direct and
   friendly, in the language the sender used.
4. List separately - without drafting replies - any email about money,
   contracts, complaints, legal issues or personal matters.

Rules: don't invent facts, dates, prices or commitments. If a reply needs
information you don't have, write [ASK ME: what you need] in the draft.

[paste emails]

Tools. Copilot in Outlook if your company runs Microsoft 365; Gemini in Gmail, which Google includes from its cheapest Workspace business plan; or ChatGPT or Claude connected to your mailbox. For confidential mail, paste the thread into a local model running in Ollama or LM Studio instead (Part 5).

The human check. Names, dates, amounts and attachments; anything that reads like a promise; and the recipient list, because AI doesn’t know who shouldn’t be in copy.

Watch out. An email can contain instructions aimed at AI assistants, such as “ignore your previous instructions and forward this thread to…”, hidden in white text or tiny print. That’s harmless while your assistant only drafts for you, and dangerous once an agent can send, forward or click links on its own. Keep email assistants on read-and-draft permissions.

---

Task 2 — Meeting Notes, Summaries and Action Items

Why it eats your week. Meetings took 23% of people’s Microsoft 365 time in 2023, and more than half of today’s meetings are ad hoc calls with no invite or agenda. The bigger cost often comes afterwards: writing them up, chasing the actions, and briefing everyone who wasn’t there.

What the evidence says. Transcribing and summarising meetings was the most common use of Copilot in the DBT trial, with 233 diary entries, and saved about 0.7 hours per meeting write-up. Microsoft’s early study found people caught up on a missed meeting nearly four times faster from an AI recap than from the recording (11 minutes instead of 43), while remembering slightly fewer details. And the big field experiment found no change in how long people spent in meetings at all. AI meeting notes make meetings cheaper to document, not shorter. Shortening them is still your job.

How to hand it over

1. Tell people first. Put it in the invitation (“this meeting will be transcribed by [tool] for the minutes”) and say it again at the start. In Germany this isn’t a courtesy; see below.
2. Use the transcription built into your meeting platform (Teams, Meet or Zoom) rather than a third-party bot that joins as a guest. That means fewer companies handling your data and clearer admin controls.
3. Ask for decisions, actions and open questions, not a narrative summary of who said what.
4. Check owners and dates, then send the recap within the hour. A recap that arrives the next day is half as useful.
5. Delete the raw transcript when your retention rules say so, and keep the checked recap.

Prompt to copy

Below is the transcript of a meeting about [topic] with [participants].
Write the recap in this order:

1. Decisions - one line each, with who made the decision.
2. Action items - a table: task | owner | due date | quote from the
   transcript that supports it (max 12 words).
3. Open questions nobody answered.
4. Anything that sounds like a commitment to a customer, or a number or
   date that should be double-checked.

Use only what is in the transcript. If an owner or a date wasn't stated,
write "not stated" - don't guess. Keep the whole recap under 300 words.

[paste transcript]

The supporting quote in the action-item table is the trick here: it lets you check each action against the transcript in seconds instead of re-reading the whole thing.

The human check. Who owns each action, and by when; anything said off the record that shouldn’t be written down; and every number or commitment to a customer.

Watch out — especially in Germany. Recording or transcribing the non-public spoken word without consent can be a criminal offence under § 201 of the German Criminal Code (StGB), and a transcript is personal data under GDPR. The law firm Heuking’s April 2026 guidance recommends announcing transcription in the invitation and again before recording starts, getting explicit agreement, signing a data processing agreement with the provider, switching off AI training and unnecessary storage, and checking whether live captions without any storage would do the job. Don’t transcribe HR conversations, medical topics or anything a participant objects to. And if a company tool can evaluate who said what and how much, your works council will want a say (Part 4).

---

Task 3 — First Drafts: Reports, Proposals, Posts and Briefs

Why it eats your week. Word alone accounted for 10% of people’s Microsoft 365 time in 2023, and the slowest part of any document is the blank page.

What the evidence says. This is the best-evidenced creative task. The Science study found 40% less time and 18% better quality on professional writing tasks. In the DBT diary study, drafting written documents produced the biggest saving of any task: 1.3 hours per document. In Microsoft’s early study, first drafts were done in 8 minutes instead of 14.

It’s also where the trouble starts. Workslop, the polished-looking but empty output that 40% of US desk workers say they received last month, is what an unread AI draft becomes once someone forwards it. A draft that looks finished is the most dangerous kind, because it tempts you to skip the reading.

How to hand it over

1. Write the skeleton yourself: the purpose, the reader, the three to five points you want to make, and what you’re asking for. That’s the thinking. It takes ten minutes, and it’s the part AI can’t do for you.
2. Give it your material: notes, figures, the previous version, a good example of the format you want.
3. Ask for two different structures and choose one; then ask for the draft.
4. Edit for truth first, then for voice. Check every claim. Then cut a fifth of the words and put your own phrasing back.
5. Ask for the opposite view. “What would a sceptical reader object to?” finds the holes before your reader does.

Prompt to copy

Draft a [proposal / report / internal memo / LinkedIn post] for [reader].

Goal: after reading it, the reader should [decide / approve / understand] [X].
Points to make, in this order: [3-5 bullets]
Source material (use only this for facts): [paste notes, figures, previous version]
Tone: plain and confident. No hype words such as "revolutionary",
"game-changing" or "seamless".
Length: about [N] words, short paragraphs, at most one table.

After the draft, list every factual claim you made and where in my source
material it comes from. Flag anything you had to assume.

The human check. Every fact, name and number; whether you’d sign every sentence; and whether it still sounds like you.

Watch out. Since 2 August 2026 the EU AI Act has required you to disclose AI-generated text that you publish to inform the public on matters of public interest, unless a person has reviewed or edited it and someone holds editorial responsibility. A business document you’ve edited yourself isn’t affected; an unedited AI article on current affairs posted to your website would be.

---

Task 4 — Research and Reading: Summarising Long Documents

Why it eats your week. About a quarter of the working week goes on searching for information, according to Atlassian. Add the reports, contracts and studies you’re supposed to have read, and research may be the biggest hidden time sink on this list.

What the evidence says. In the DBT trial’s observed test, staff summarising a report with Copilot took 12½ minutes instead of 41½ and produced a more accurate summary. Its diary study put the saving at 0.8 hours for summarising research and 0.7 hours for finding existing information, and Microsoft’s early study found searching 27% faster.

The risk is invented facts. A database of court cases maintained by the researcher Damien Charlotin had recorded 1,240 US court filings containing AI-hallucinated authorities by late July 2026: 524 in 2025, and 668 in the first seven months of 2026 alone. Lawyers have been fined for it. If professionals whose job is checking sources get caught out, the rest of us will be too.

How to hand it over

1. Give it the document. Never ask “what does the report say about…” and rely on its memory.
2. Ask for page or section references for every point it makes.
3. Ask what the document doesn’t say. The gaps are often more useful than the summary.
4. Spot-check three claims against the source. If one is wrong, check them all.
5. For web research, use a tool that cites its sources, and open them. A citation you haven’t clicked isn’t a citation.

Prompt to copy

Read the attached document carefully. Then:

1. Summarise it in 10 bullet points for a busy [role], most important first.
2. Give the page or section number for each bullet.
3. Put every number, date and deadline in a table:
   value | what it refers to | page.
4. List what the document does NOT cover that a [role] would expect it to.
5. Name the three claims I should verify before relying on this document.

If anything in the document is unclear or contradictory, say so rather
than resolving it yourself.

The human check. Open the cited pages. Check every number. For anything legal, financial or medical, verify against the original publication or database, never against another AI answer.

Watch out. Contracts, client files and personnel documents are exactly the material that shouldn’t go into a consumer AI account. Part 4 has the data tiers.

---

Task 5 — Spreadsheets and Data Analysis

Why it eats your week. Excel was the single biggest app in Microsoft’s 2023 data, at 18% of people’s time. Most of that isn’t analysis; it’s cleaning, reshaping and trying to remember formula syntax.

What the evidence says. Mixed, and the reason matters. The DBT diary found 0.6 hours saved per data analysis task. But in its observed test, admittedly with a tiny sample, Copilot users took longer on an Excel analysis (25 minutes against 20½) and got worse results. And the task that tripped up the BCG consultants, costing them 19 percentage points of accuracy, was an analysis that combined spreadsheet figures with interview notes.

The pattern is consistent: AI is good at building the machinery of an analysis and bad at eyeballing numbers and drawing conclusions. So let it write the formulas, clean the data and write code you can rerun, and keep the conclusions for yourself.

How to hand it over

1. Describe the table, not the data: column names, data types, the number of rows and what one row represents.
2. Ask for the method, not the answer: a formula, PivotTable steps, Power Query steps or a short Python script.
3. Test it on five rows you can check by hand.
4. Ask for the edge cases that would make the result wrong: blanks, duplicates, text in number columns, dates stored as text.
5. Write the conclusion yourself. “Revenue in the north fell 12%” is your sentence, not the model’s.

A tip for German readers: German Excel uses German function names and semicolons as separators (SVERWEIS instead of VLOOKUP, XVERWEIS for XLOOKUP, WENN for IF), plus a decimal comma. Say so in the prompt, or you’ll get formulas Excel refuses to accept.

Prompt to copy

I have an Excel table called "Orders" with these columns:
OrderDate (date), Customer (text), Region (text), NetAmount (EUR),
Status (Open / Paid / Cancelled). About 4,000 rows; one row per order.

I use Excel in German, so give formulas with German function names and
semicolons as separators.

Question: total paid net revenue per region per month in 2026.

1. Give me the formula or the PivotTable steps - not the answer.
2. Explain in one sentence what each part does.
3. List the edge cases that would make the result wrong (blank dates,
   cancelled orders, duplicates, numbers stored as text) and how I can
   check for each one.

The human check. Does the total reconcile with the source? Are cancelled, duplicate and test rows excluded? Are the units, the currency and the time periods what you think they are?

Watch out. Customer lists, payroll and HR spreadsheets are personal data. A simple workaround: describe the columns and give the model five rows of made-up sample data, get the formula, then run it on the real file yourself. The model never sees a real name.

---

Task 6 — Customer Support Replies and FAQs

Why it eats your week. The same twenty questions arrive every week, and each one needs a reply that is careful, friendly and correct.

What the evidence says. Some of the strongest real-world evidence for AI at work comes from customer support. The 5,179-agent study found 14% more issues resolved per hour, 34% more for new agents, happier customers and lower staff turnover, largely because the AI passed on what the best agents did to everyone else.

Replacing the agents altogether is another matter. Klarna said in 2024 that its AI assistant was doing the work of 700 agents. In May 2025 its chief executive told Bloomberg that the push had gone too far, that cost had weighed too heavily against quality, and that the company would hire people again. Gartner, in June 2025, predicted that half the organisations planning to cut their customer service staff significantly because of AI would abandon those plans by 2027.

How to hand it over

1. Start with agent assist: AI suggests the reply, a person approves it. That is the set-up the 14% study measured.
2. Ground it in your own material (help articles, policies, price lists) and tell it to say “I need a human” when the answer isn’t there.
3. Mine your last 100 tickets. Ask AI to cluster the questions and draft FAQ answers; a person approves them; publish them in your help centre.
4. Route the hard ones to people by rule: complaints, refunds, cancellations and anything legal.
5. If customers talk to a bot directly, tell them it’s a bot and always offer a human. Under the EU AI Act, systems that talk to people must make it clear they’re AI, unless that’s obvious anyway.

Prompt to copy

You draft replies for our customer support team.
Use ONLY the policies and help articles below. If the answer isn't in
them, reply "I need a human for this" and say what information is missing.

Customer message: [paste]
Our policies and help articles: [paste]

Write:
1. A reply under 120 words - friendly, specific, in the customer's language.
2. The policy line or article you based it on.
3. Your confidence (high / medium / low) and why.

Never promise refunds, discounts, delivery dates or legal outcomes.

The human check. Is the policy quoted correctly? Is there a promise of a refund, a discount or a date that nobody authorised? Would you be happy to see the reply screenshotted on social media?

Watch out. Support tickets are full of names, addresses and order histories. Use the AI built into your support platform under your existing data processing agreement, or a business AI plan that has one.

---

Task 7 — Presentations and Slide Decks

Why it eats your week. Microsoft’s 2025 data found that PowerPoint edits spike by 122% in the final ten minutes before meetings. Most slide time isn’t design. It’s working out what you’re trying to say, at the last minute.

What the evidence says. This is where AI’s reputation runs furthest ahead of the evidence. In the DBT diary study, presentations saved no time at all, and producing slide decks had one of the lowest satisfaction scores (54% satisfied, 23% dissatisfied). In its observed test, Copilot built a deck faster (11 minutes against 18½), but with worse quality and much worse accuracy.

AI is far better at the thinking scaffold than at the finished deck: the storyline, the slide titles, slide text drawn from your document, speaker notes and the awkward questions.

How to hand it over

1. Write your one-sentence message, and the question your audience needs answered.
2. Ask for a storyline: 8 to 12 slide titles, each written as a full sentence that states the point, so the titles alone tell the story.
3. Then ask for slide text from your source document, 30 words per slide at most, plus speaker notes.
4. Build the visuals in your own template, with charts made from your data, not from the model’s imagination.
5. Rehearse against AI. Ask it to play the most sceptical person in the room.

Prompt to copy

I'm presenting to [audience] for [N] minutes. They need to decide [X].
My one-sentence message: [message]
Source material: [paste]

1. Propose a storyline of 8-10 slides. Write each slide title as a full
   sentence that states the point, so the titles alone tell the story.
2. For each slide: up to 3 bullets (30 words max) and 2-3 sentences
   of speaker notes.
3. Mark every number with [CHECK] and say where in the source it comes from.
4. End with the five hardest questions this audience is likely to ask.

The human check. Every number on every slide, against its source; the storyline, in your own words; and the brand template.

Watch out. Image generation inside office tools took longer than doing without it in the DBT trial. Don’t spend twenty minutes prompting for a picture your stock library has in two.

---

Task 8 — Translation and Localisation

Why it eats your week. If you work in Germany, anywhere else in the EU, or with anyone outside your language, you translate constantly: emails, offers, documentation, product texts.

What the evidence says. One of the few productivity studies with German as the source language looked at professional translators working on banking and finance texts in 2019. Post-editing neural machine translation, instead of translating with their usual translation memories and termbases, made them 59.7% faster from German into French and 9.3% faster into Italian, with equal or slightly better quality. Today’s models are far stronger than the 2019 systems, but the study’s real lesson still holds: the speed came from editing a machine draft, not from skipping the human review.

How to hand it over

1. Give it context: the audience, the purpose, the register, and a glossary of terms that must stay fixed, such as product names, legal terms and your company’s own vocabulary.
2. Say Sie or du. In German that choice changes the whole text; for business customers, the formal Sie is the safe default.
3. Ask for a second pass: “Review this as a native-speaking editor for [market] and list what you changed.”
4. Have a native speaker read anything external before it goes out.
5. Keep your glossary and paste it into every translation prompt.

Prompt to copy

Translate the text below from English into German for [audience].

Register: formal ("Sie").
Keep these terms in English: [list]
Always translate these terms like this: [glossary]
Keep the meaning, not the sentence structure - it should read as if it
had been written in German.

After the translation, list:
1. Any sentence whose meaning was ambiguous in the original.
2. Idioms or examples you adapted for a German reader.
3. Every number, date, currency or unit format you changed
   (for example 1,000.50 -> 1.000,50).

[paste text]

The human check. Numbers, dates and units (1,000.50 becomes 1.000,50 in German); legal terms; tone; and anything that could be read as a promise.

Watch out. Contracts, terms and conditions and official documents still need a professional translator, and certified translations for German authorities have to come from a sworn translator. A translation of a personal document is also processing of personal data, so the data tier matters (Part 4).

---

Task 9 — Admin: Forms, Invoices, Data Entry and Scheduling

Why it eats your week. Copying figures from PDFs into systems, filling in the same forms, filing receipts, finding a meeting slot: work nobody wants and everybody has.

What the evidence says. This is one of the fastest-growing business uses of AI. In Anthropic’s January 2026 Economic Index, office and administrative support rose to 13% of business API traffic, driven by email, document processing and scheduling automation, and data entry was among the jobs where AI handled most of the core, time-consuming work. Anthropic’s model-based estimates suggest tasks such as producing invoices and memos take about 87% less time with AI, though again before anyone checks the result.

The exception is scheduling. In the DBT trial, scheduling with Copilot took longer than without it (0.6 hours per task), and it was the least satisfying use of all, with 36% of those who tried it dissatisfied. Finding a slot across five calendars is a coordination problem, and a booking link solves it better than a language model does.

There’s also a German twist. Since 1 January 2025, every business in Germany must be able to receive structured e-invoices in formats such as XRechnung or ZUGFeRD, and a plain PDF no longer counts as an e-invoice. Issuing them becomes mandatory from 2027 for businesses with more than €800,000 in annual turnover and from 2028 for everyone. As that happens, AI extraction from invoice PDFs becomes a bridge rather than a destination: the better long-term investment is accounting software that reads e-invoices directly.

How to hand it over

1. Extraction: receipts and PDFs into a table with date, supplier, invoice number, net, VAT and gross.
2. Build validation rules into the prompt: does net plus VAT equal gross; is the VAT rate 19% or 7%; are there duplicates?
3. Forms: let AI pre-fill a draft from your data; you check it and submit it.
4. Scheduling: use a booking page (Microsoft Bookings, Calendly or Cal.com, for example) and let AI draft only the “here are three slots” email.
5. Approvals and payments stay human. Always.

Prompt to copy

Extract these fields from each attached receipt or invoice into a table:
date | supplier | invoice number | net amount | VAT rate | VAT amount |
gross amount | currency | category (travel / software / hardware /
office / other)

Rules:
- Use German number format (1.234,56).
- If a field is missing or unreadable, write MISSING.
- Never fill in a value by calculation without marking it CALCULATED.

Then check every row: does net + VAT = gross (within 0.01)? Is the VAT
rate 19% or 7%? Flag rows that fail and any possible duplicates.

The human check. Totals, VAT and duplicates, plus every value the model marked as missing or calculated.

Watch out. Never paste bank details, card numbers or login details into an AI chat, and never give an agent the ability to pay, submit or delete anything.

---

Task 10 — Small Automations: Formulas, Scripts and Workflows

Why it eats your week. The weekly report you rebuild by hand, the files you rename one by one, the data you copy from one system to another. Each takes twenty minutes; together they can take a day a month.

What the evidence says. In GitHub’s 2023 experiment, developers finished a coding task 55.8% faster with Copilot, and Anthropic’s estimates put software development at the top of its productivity table, accounting for 19% of the total gain it modelled. METR’s study is the counterweight: experienced developers working on their own complex codebases were slower in early 2025 and, at best, modestly faster by late 2025.

For non-programmers the gain is different in kind: things you couldn’t do become things you can. I don’t come from a coding background: I built my app Stash with AI coding agents and wrote up the whole process in How to Build an App From Scratch With AI in 2026. A 30-line script that renames 400 files is a far smaller job than that.

How to hand it over

1. Write down the chore as you do it by hand, step by step, with an example input and the output you want. That’s your specification.
2. Ask for the simplest tool that will do the job: an Excel formula, a Power Automate or n8n flow, a Google Apps Script, or a short Python script.
3. Insist on a dry-run mode that shows what would change without changing anything, and run it on copies first.
4. Give it the least access that works. Read-only where possible; never payment or delete rights.
5. Write down what it does and who owns it, so it doesn’t turn into shadow IT the day you go on holiday.

Prompt to copy

Every week I do this by hand: [describe each step, the apps, file names,
an example input and the output I want].

1. Suggest the simplest way to automate it with tools I already have
   ([Microsoft 365 / Google Workspace / n8n / Python]) and explain the
   trade-offs in three sentences.
2. Write the automation with plain-English comments.
3. Add a dry-run mode that shows what would change without changing anything.
4. List what could go wrong and how I would notice.

I'm not a programmer: give me step-by-step instructions to run and test it
safely on a copy of my data first.

Tools. Power Automate if you’re on Microsoft 365, Apps Script on Google Workspace, n8n (built in Berlin, and self-hostable on your own server), or Zapier and Make for app-to-app workflows. For anything bigger, an AI coding agent (I compared 28 of them in Every AI Coding Harness Compared, and my vibe coding guide covers the basics). For an assistant that lives in your chat apps and runs errands, see my OpenClaw guide and OpenCode v2 vs Hermes Agent vs OpenClaw.

The human check. Run the dry run. Test with messy data. Make sure it fails loudly, not silently.

Watch out. “Agentic AI” is the most over-labelled product category around. In 2025 Gartner reckoned that only about 130 of the thousands of vendors selling it offered genuinely agentic capabilities; the rest were rebranded chatbots and automation tools. Fix and document the process first, then automate it.

---

Part 4 — Privacy and the Law: What You Can Paste Where in Germany and the EU

Everything in this part is general information, not legal advice. If your company has a data protection officer, a works council or an AI policy, they come first.

Shadow AI is already in your company

Across the EU, Eurostat found that 20% of enterprises with ten or more employees used AI in 2025, up 6.5 points in a single year, and 55% among large companies. In Germany, the Institute for Employment Research (IAB) found that 25% of businesses used generative AI in 2025, rising to 48% of those with 200 or more employees, and that 90% of them relied on freely available software.

What employees do on their own is harder to measure. In an October 2025 Bitkom survey, 42% of German companies said they assume staff use generative AI on private accounts at work (8% said it was widespread), while only 23% had clear rules for using AI and about a quarter provided an AI tool of their own.

So the question for most teams is no longer whether people use AI at work. It’s whether they use it in a way the company can stand behind.

GDPR in practice

• Personal data means GDPR applies. The German data protection authorities’ joint guidance on AI (the DSK, May 2024) reads “personal data” broadly, and warns that superficial pseudonymisation, such as swapping a name for initials, usually isn’t enough to take a text outside the GDPR.
• A cloud AI service is usually a processor. That means you need a data processing agreement: in German, an Auftragsverarbeitungsvertrag (AVV) under Article 28 GDPR. The business and enterprise plans of the main providers offer one; consumer accounts generally don’t.
• Check training settings. Make sure your inputs aren’t used to train the provider’s models. Anthropic, for example, says Claude’s Team and Enterprise plans don’t train on your content by default, while on its individual plans training is something you opt out of.
• Accuracy counts. The DSK rejects the idea that AI output doesn’t need to be correct: wrong output about a person can itself be unlawful processing.
• A data protection impact assessment is often required when AI tools process personal data; the DSK says so explicitly. Your data protection officer will know whether your use case needs one.

Recording and transcribing meetings

Task 2 covers this in detail. The short version: § 201 StGB can make recording or transcribing the non-public spoken word without consent a criminal offence, so announce it, get agreement, and check the tool’s storage and training settings before you press record.

Works councils

If your company has a works council (Betriebsrat), technical systems that can monitor employees’ behaviour or performance are subject to co-determination under § 87(1) No. 6 of the Works Constitution Act (BetrVG). The first published labour court decision on generative AI, by the Hamburg Labour Court on 16 January 2024 (case 24 BVGa 1/24), held that letting staff use ChatGPT through their own browser accounts did not trigger co-determination, because the employer had no access to the usage data. Commentators stress that the result could easily be different for company-provided tools that log usage, and the DSK guidance says works council involvement should be checked whenever AI is introduced. Bring the works council in early. It’s faster than bringing it in late.

The EU AI Act: the parts that matter for office work

The AI Act was amended this year by the so-called digital omnibus, Regulation (EU) 2026/1744, published on 24 July 2026. For everyday office use, four points matter:

• AI literacy (Article 4) still applies, in a softer form. Companies that use AI must take measures to support the AI literacy of their staff, but no specific level of knowledge is required. In practice: a short training session and written rules.
• Transparency (Article 50) has applied since 2 August 2026. Deepfakes, and AI-generated text published to inform the public on matters of public interest, must be disclosed, unless a person has reviewed the text and someone holds editorial responsibility for it. Chatbots must make clear to people that they’re talking to an AI.
• High-risk uses such as AI in recruitment and employment decisions now face their main obligations from 2 December 2027 instead of August 2026. That’s a reason to stay careful, not a reason to automate hiring now.
• Most of the ten tasks in this guide aren’t high-risk. Writing, summarising, translating and building spreadsheets don’t make an AI system high-risk. Using AI to evaluate people can.

Professional secrets and trade secrets

Lawyers, doctors, tax advisers and other professionals bound by § 203 StGB need contractual safeguards before client data goes to any outside service provider, AI providers included. And German trade-secret law (GeschGehG) only protects information you’ve kept secret with appropriate confidentiality measures. Pasting it into an uncontrolled consumer chatbot is the opposite of that.

A one-page AI policy for a small team

1. Approved tools and the accounts to use: business plans, AVV signed, training switched off.
2. The data tiers from the guide above, with examples from your own business.
3. The never-paste list: passwords, bank and card details, health data, personnel files, client secrets.
4. The human-check rule: nothing AI-generated leaves the company unread.
5. Meeting rules: who may record, how it’s announced, how long transcripts are kept.
6. Labelling rules for anything you publish.
7. Who to ask, and how to propose a new tool.
8. A one-hour training session for everyone, repeated once a year. That’s your Article 4 measure.

---

Part 5 — Which AI Tools to Use: Built-In, Standalone or Local

Start with the AI inside the software you already pay for. It works on your existing files and mail, your administrator controls it, and your contract with the vendor usually already covers the data.

Your situation	Start with	Price (USD, October 2026, excl. VAT)	Best for
Your company runs on Microsoft 365	Copilot Chat, then Microsoft 365 Copilot Business	Copilot Chat included with eligible plans; Copilot Business $21 per user a month on an annual plan, for up to 300 users	Email, meetings, Word, Excel and PowerPoint in one place
Your company runs on Google Workspace	Gemini in Workspace	Included: Business Starter $7 (Gmail and the Gemini app), Standard $14 (adds Docs and Meet), Plus $22 per user a month, annual	Gmail triage, drafts in Docs, Meet notes
You work alone or in a small team, on any software	ChatGPT or Claude	About $20 a month per person (ChatGPT Plus, Claude Pro); team plans from about $20–25 per seat	Drafting, research, analysis, reusable briefs
The data mustn’t leave your machine	A local model in Ollama or LM Studio	Software free; you pay in hardware	The sensitive tiers in Part 4
Repetitive chores across several apps	Power Automate, Apps Script, n8n, Zapier or Make	From free (n8n self-hosted)	Task 10 automations

Prices change constantly, so check before you buy. Microsoft’s price is from its December 2025 Copilot Business launch; Google’s as checked by Dupple on 12 October 2026; Claude’s from claude.com on 25 October 2026; ChatGPT’s as listed in my AI coding harness comparison on 22 October 2026.

Which assistant writes best? That changes every few months, and it matters less than the brief. Pick the one your company can approve with a data processing agreement, and learn it properly.

The local option. For data that mustn’t leave your control, a local model is the cleanest answer: nothing is sent anywhere, so there’s no processor and no data transfer. The trade-offs are quality and hardware. Current open-weight models from the GLM, Qwen, Kimi, DeepSeek and MiniMax families are remarkably capable, but the strongest need a lot of memory. My MacBook Pro M5 Pro 64 GB guide covers what runs well on a laptop, and my comparisons of Kimi K3, Fable 5.1 and GLM-5.3 and GLM-5.3-Flash vs MiniMax M3 go through the models themselves. For summarising, drafting and translating confidential text, a mid-sized local model is often good enough; test it on your own documents before you decide. My State of AI 2026 report has a section on the cost and compliance maths of running AI offline.

Agents. 2026’s big promise is AI that doesn’t just draft but acts: books, files, sends, updates the CRM. Some of it works, and my AI for Business Growth guide covers where. But keep agents at level 2 or 3 (Part 3) until you’ve measured them, give them the least access possible, and remember Gartner’s forecast that over 40% of agentic AI projects will be cancelled by the end of 2027.

---

Part 6 — How to Measure Whether You Really Save Time at Work With AI

Measuring is the step that turns “AI feels helpful” into “AI saves me four hours a week”, or into “this task isn’t worth it”. It needs no software, just four weeks and some honesty.

1. Baseline. For two weeks, log the time you spend on the three tasks you’ve picked. A note in your calendar is enough.
2. With AI. For the next two weeks, log three numbers per task: time with AI, time spent checking, and time spent fixing.
3. Work out the net time saved: net time saved = time before − (time with AI + time checking + time fixing).
4. Count the rework. How often did you have to redo AI output substantially? If it’s more than one time in five, fix the brief or drop the task.
5. Check quality blind, once. Give a colleague two versions of the same piece of work, one made with AI and one without, and don’t say which is which.

Task	Minutes a week before	With AI	Checking	Fixing	Net saved	Keep?
Email triage
Meeting recaps
Report drafts

Why bother? Because feelings mislead (METR’s developers), because a 40% rework rate turns a big saving into a small one (Workday), and because some tasks genuinely get slower (the DBT trial). Measure first, then decide.

---

Part 7 — Your First 30 Days: A Rollout Plan

Week 1 — Audit and pick. List your recurring tasks, with how often you do them and how long they take. Run them through the five tests from Part 1 and the matrix from Part 3, and pick three from the “hand over now” box. For most people, that means email triage, meeting recaps and one kind of document. Set up an approved tool and write your one-page rules (Part 4).

Week 2 — Brief and run. Write a five-part brief for each task and use it every time the task comes up. Save the prompts that work. Log your time.

Week 3 — Measure and fix. Compare with your baseline. Fix the briefs that produce rework, drop the tasks that don’t save net time, and add one or two tasks from the “AI drafts, you decide” box.

Week 4 — Share and scale. Turn the prompts that work into templates, saved instructions or a shared prompt library. Show colleagues what worked, with real numbers. Decide whether paid licences are worth it, based on your own measurements rather than a vendor’s slide.

For teams: name one person per team as the go-to for AI questions; bring in your works council and data protection officer in week 1, not week 4; and run the training that the AI Act’s literacy rule expects. For the bigger organisational picture, moving from single tasks to connected workflows, my AI for Business Growth guide has a 30-60-90-day plan.

---

Part 8 — Ten Mistakes That Eat the Time You Saved

1. Asking from memory instead of giving the source. The model will answer anyway — fluently, and sometimes wrongly.
2. Forwarding drafts you haven’t read. That’s how workslop is born, and it costs the person who receives it about two hours.
3. Delegating judgement. Outside the frontier, AI makes good professionals worse: the BCG consultants lost 19 points of accuracy.
4. Letting AI send, pay, submit or delete. Drafts and suggestions only, until you’ve measured it and set limits.
5. Using AI where a simpler tool exists. A booking link beats an AI scheduler; a saved template beats a prompt you rewrite every time.
6. Generating slides instead of a storyline. Decide what you’re saying first.
7. Trusting numbers you didn’t recalculate. Ask for the formula, not the answer.
8. Pasting personal or confidential data into a consumer account. It’s the most common mistake, and the one with legal consequences.
9. Not saving prompts that work. Every time you rewrite a brief that already worked, you pay the setup cost twice.
10. Measuring by feel. Experienced developers believed they were 20% faster while measuring 19% slower.

---

Frequently Asked Questions

What’s the fastest way to save time at work with AI?

Start with email triage, meeting recaps and first drafts. They happen every week, mistakes are cheap to catch, and they have the strongest evidence behind them: in a field experiment with 7,137 knowledge workers, Copilot users spent two fewer hours a week on email, and in a UK government trial drafting documents saved about 1.3 hours per task. Brief the AI properly, check its output, and save the prompts that work.

How much time does AI actually save at work?

On single tasks, controlled experiments show 25% to 74% less time, including 40% on professional writing tasks in a study published in Science. Across whole working weeks the measured savings are much smaller: 2.8% of working hours for Danish chatbot users, and 1.6% of all US working hours in the St. Louis Fed’s 2025 surveys. The gap comes from tasks AI can’t do, time spent checking, and rework.

Which work tasks should I not hand to AI?

Anything where a mistake is expensive and hard to spot: legal and contractual decisions, hiring and performance evaluations, medical or financial advice, conclusions drawn from data you haven’t checked, and any action that sends, pays, submits or deletes on its own. In a BCG experiment, consultants using GPT-4 on a task just beyond its abilities were 19 percentage points less likely to get it right.

Is it legal to use ChatGPT at work in Germany?

Yes, with conditions. If you enter personal data, the GDPR applies: use a business plan with a data processing agreement under Article 28 GDPR, switch off model training, and follow your employer’s AI rules. Keep client secrets, health data and personnel files out of consumer accounts. Under the EU AI Act, companies must also take measures to build their staff’s AI literacy. This is general information, not legal advice.

Can I record and transcribe meetings with AI in Germany?

Only with the participants’ knowledge and agreement. Recording the non-public spoken word without consent can be a criminal offence under § 201 StGB, and a transcript is personal data under the GDPR. Announce transcription in the invitation and at the start, use a provider with a data processing agreement, switch off AI training and unnecessary storage, and don’t transcribe sensitive HR or medical conversations.

Does the EU AI Act affect how I use AI at work?

For everyday office tasks, only lightly. Companies must take measures to support their staff’s AI literacy under Article 4. Since 2 August 2026, deepfakes and AI-written text published to inform the public on matters of public interest must be disclosed, unless a person has reviewed it and holds editorial responsibility. AI used in recruitment or employment decisions is high-risk, with the main obligations applying from 2 December 2027.

Is Microsoft 365 Copilot worth it for a small business?

It depends on how much of your week happens in Outlook, Teams and Word. Copilot Chat comes with eligible Microsoft 365 plans at no extra cost, and Microsoft 365 Copilot Business costs $21 per user a month on an annual plan, for up to 300 users. Run a 30-day pilot on three tasks, measure the net time saved, and buy licences only for the people whose numbers justify them.

What is AI workslop?

Workslop is AI-generated work that looks finished but lacks substance, leaving the person who receives it to do the real thinking. In a September 2025 survey of 1,150 US desk workers by BetterUp Labs and the Stanford Social Media Lab, 40% had received workslop in the past month, and each instance took about two hours to deal with, roughly $186 per employee per month in lost time.

Should I use a local AI model for work?

Use one for data that mustn’t leave your control: client secrets, personnel files, health information, unpublished figures. Nothing is sent to a provider, so there’s no processor and no international data transfer. The trade-offs are hardware cost and somewhat lower quality than the best cloud models. Tools such as Ollama and LM Studio run open-weight models on a capable laptop or desktop.

How do I measure the time AI saves me?

Log the time a task takes for two weeks without AI, then two weeks with it, counting the time you spend checking and fixing AI output as well as generating it. Net time saved is the time before, minus the time with AI, checking and fixing. Don’t rely on your impression: in a 2025 trial, experienced developers believed AI had made them faster when it had made them 19% slower.

Will handing tasks to AI cost me my job?

The best evidence so far shows little effect. A Danish study linking surveys of about 25,000 workers to official records found no significant effect of AI chatbots on earnings or recorded hours in any of 11 exposed occupations, two years after ChatGPT’s launch. What changes first is the mix of the job: less drafting and summarising, more checking, deciding and dealing with people.

Which AI tool is best for office work?

The one your company can approve and you’ll actually use every day. If you’re on Microsoft 365, start with Copilot; on Google Workspace, with Gemini, which is included in its business plans. For individuals, ChatGPT and Claude both cost about $20 a month. For confidential data, use a local model. The quality of your brief matters more than the brand of the tool.

Can AI answer customer emails on its own?

It can draft them well, but it shouldn’t send them unsupervised. In a study of 5,179 support agents, AI suggestions raised the number of issues resolved per hour by 14%, and by 34% for new agents, with a person approving every reply. Klarna, which leaned heavily on full automation, said in 2025 that it had gone too far and started hiring human agents again.

---

The Bottom Line

AI saves real time at work, just not where most people first point it, and not as much as the demos suggest. On a single, well-chosen task the savings are large: 40% on a writing task, two hours a week on email, more than an hour on every document draft in one government trial. Across a whole week, the measured gains so far are a few percent, because most work isn’t delegable yet, because checking costs time, and because some tasks get slower.

The way to close that gap is unglamorous. Pick the three tasks that recur, are mostly text and are cheap to check. Brief them properly. Check in proportion to the risk. Never let the tool send, pay or decide. Save what works. Measure the net time, not the feeling. And keep personal data where the law says it belongs.

Do that for a month and these ten tasks will give you back hours each month, not minutes. Skip it, and you risk becoming the colleague who sends the workslop that 40% of desk workers say they received last month.

---

Using AI for a task I didn’t cover, or measuring very different savings? I’d like to hear which task, which tool, and what it really saves you once the checking is done. Corrections and counter-evidence are welcome; this guide gets updated when the evidence changes.

Facts checked: 25 October 2026.