Free lesson

Where AI Really Saves Time — and Where It Steals It

Where AI Really Saves Time — and Where It Steals It — Save Time With AI: Email, Documents, Meetings and Planning. The promise and the actual numbers; The jagged frontier; The task audit; Your baseline and a fictional example; Decide whether the test worked. Practical assignment: Audit your week.

Demo edition. The materials are undergoing editorial review.

Topics covered

  • The promise and the actual numbers
  • The jagged frontier
  • The task audit
  • Your baseline and a fictional example
  • Decide whether the test worked

The promise and the actual numbers

The goal of this lesson is to find three of your own tasks where AI may save time and measure how long they take now.

Research results depend heavily on the task. In a 2023 experiment, 453 professionals completed short occupational writing assignments. With ChatGPT, average completion time fell by 40% and rated quality rose by 18%. These were roughly 20–30-minute tasks; the researchers explicitly noted that factual accuracy was not checked.

In another study, an assistant raised customer-support agents’ productivity by nearly 14% on average. Less-experienced workers gained most, while the effect for the most experienced workers was small.

Everyday work shows more modest results. Surveys of about 25,000 Danish workers in 2023–2024 found that users reported saving an average of 2.8% of working time—about one hour in a 40-hour week. In a UK government trial from September to December 2024, more than 20,000 civil servants reported an average saving of 26 minutes per day, but 17% noticed no clear time saving. These were self-reports, not independent time measurements.

The practical conclusion is limited: a strong result on a short, well-matched task does not mean an entire working week will become 40% shorter.

The jagged frontier

In a study of 758 consultants, AI helped on tasks that matched its capabilities: participants completed more tasks, worked about a quarter faster and received higher ratings. On a task outside the tool’s capabilities, however, AI users were 19 percentage points less likely to reach the correct answer.

The researchers called this the jagged technological frontier. Human-perceived difficulty is a poor guide to AI fit. A model may rewrite an awkward email well yet fail to apply every tax, discount and exception in a budget, even though the arithmetic looks simpler.

Feeling faster can also mislead. In METR’s 2025 study, 16 experienced developers completed 246 real tasks in projects they knew. With early-2025 tools, they took 19% longer, although afterwards they believed they had been about 20% faster. This is a narrow result for one group and tool generation, not a rule about all software development.

A Reddit report illustrates the same practical problem without proving it statistically: the writer received different income totals from several models and had to check everything manually. Verification time therefore belongs in the full cost of using AI.

The task audit

Assess individual recurring actions, rather than “AI” as one category. Record four features for each task:

  • Frequency per week. A small saving accumulates only on recurring work.
  • Minutes per occurrence. This shows the possible size of the gain.
  • Whether it is mostly text. Drafting, rewriting, extracting points and summarising are often easier to check than calculations or decisions with hidden conditions.
  • Cost of an error. Mark it low or high after considering money, legal consequences, health, confidentiality and reputation.

Good first candidates are frequent, mainly textual tasks whose errors are easy to notice and repair. When an error would be costly, AI may suggest questions or provide a second check, but an accountable person must verify the result against the primary material.

Inspect your calendar, sent mail and work chats for the last three working days; use a week if your days vary greatly. Name the action precisely: “reply to supplier questions,” rather than “do email.”

A common mistake is starting with the grandest job, such as a whole strategy or budget. Quality is harder to benchmark there, and a hidden error costs more. A dull recurring task is often the better first test.

Your baseline and a fictional example

A baseline is how long the task takes without AI. The next time you do it, record the start and finish. For short actions, time a batch: “10 replies took 45 minutes” is more reliable than recalling one email. Later, compare that with the full AI-assisted time: prompt preparation + waiting + reading + verification + correction.

Fictional example. Diana, a project coordinator, reviewed three working days and found these candidates:

  • client and supplier replies: about 20 per day; a baseline batch of 10 took 45 minutes;
  • follow-up emails after 12 weekly meetings: 15 minutes each, or 180 minutes per week;
  • reading venue and supplier contracts: about 2 hours per week.

She marked budgets “AI only assists with checking” because an error involving a taxable service charge could change the client total. Contracts are not automatically low-risk either: a summary can aid navigation, but it does not replace reading the clauses on which a decision depends.

If your employer restricts tools or data, those rules take priority. When you do not know the rules, practise with fictional or non-confidential material.

Decide whether the test worked

After the first attempt, do not record only whether you liked the output. Write down two numbers and one decision:

  1. Baseline time without AI.
  2. Full time with AI, including checking and correction.
  3. Decision: use it, change the method or reject it for this task.

Fictional example continued. Without AI, Diana’s 10 replies took 45 minutes. With AI, draft preparation took 12 minutes and checking plus correction took 18, for a total of 30 minutes. The batch saving was 45 − 30 = 15 minutes, or 15 ÷ 45 ≈ 33%. For a budget, generation took 8 minutes, but comparison and correction took 58; the total was 66 minutes against a 60-minute baseline. AI added 6 minutes there.

Do not transfer the email result to the budget: these are separate tasks on different sides of the jagged frontier. Do not omit checking to improve the figure either. That would measure draft generation, not completed work.

Practical assignment

Audit your week

Continue to learning

Course outline

  1. Where AI Really Saves Time — and Where It Steals It
  2. Email in 10 Minutes a Day
  3. Documents: Read Less, Understand More
  4. Write Faster Without Losing Your Voice
  5. Meetings: Prepare, Capture, Follow Up
  6. Planning: Your Week, Your Projects, Your Priorities
  7. Templates and a Workspace: Stop Starting From Zero
  8. Automation and Agents: What You Can Hand Over Completely
  9. Work Data, Honesty and Your Team
  10. Measure It, Then Build Your 30-Day System