Measuring time savings: cases times minutes, before and after

Whether automation saves time is shown not by a promise but by a calculation with your own figures: how many cases a month, how many minutes each takes today, how many after the change, checking included? Two often-quoted studies offer no shortcut. One measures, in the customer support of a software company, about 14 percent more issues resolved per hour; the other arrives in Denmark at about 3 percent self-reported time savings. The gap arises because the two studies measure different things.

The calculation: cases times minutes

A case is a unit of work that repeats: an incoming enquiry, an incoming invoice, a quotation to be followed up. Which of these suits a fixed workflow, meaning a sequence of steps a program carries out, is explained in Process automation in a small business: the basics. The saving comes from two figures that you collect separately:

cases per month × (minutes today minus minutes afterwards) = minutes saved per month

Count the cases in the mailbox, the filing system or the quotation list, ideally over three months, because single months fluctuate. Time the minutes on real cases, not from memory. As a rule of thumb I take 15 to 20 cases. That is the author’s practical assumption, not a requirement from any source. After the change you measure in the same way, for at least four weeks (also a practical assumption), and count everything the workflow newly creates: checking results, correcting errors, following up on wrong answers.

An example with invented data, an office with three workflows:

Task	per month	Minutes today	Minutes after	Total today	Total after
Sorting enquiries	40	4	3	160	120
Assigning incoming invoices	50	4	3	200	150
Following up quotations	20	15	10	300	200
Total				660	470

todayafter change

• Sorting enquiries40 × 4 min against 40 × 3 min

  today 160 min/month
  after change 120 min/month

• Assigning incoming invoices50 × 4 min against 50 × 3 min

  today 200 min/month
  after change 150 min/month

• Following up quotations20 × 15 min against 20 × 10 min

  today 300 min/month
  after change 200 min/month

Source: Worked example with invented data, neither a measurement nor a forecast

Text version of the graphic

• Sorting enquiries: today 160 min/month, after change 120 min/month (40 × 4 min against 40 × 3 min)
• Assigning incoming invoices: today 200 min/month, after change 150 min/month (50 × 4 min against 50 × 3 min)
• Following up quotations: today 300 min/month, after change 200 min/month (20 × 15 min against 20 × 10 min)
• Source: Worked example with invented data, neither a measurement nor a forecast

Saved are 660 minus 470, so 190 minutes or 3 hours 10 minutes a month. A full-time position has about 173 working hours a month (40 hours × 52 weeks ÷ 12 months). The just over 3 hours are just under 1.8 percent of that. This is a worked example, not a forecast.

Two counter-calculations belong here. First, set-up costs time. Suppose it took 20 hours, then it is paid off only after 20 ÷ 3.17, so a little over six months, and upkeep is not yet included. Second, the saving depends on the number of cases: if you do a task ten times a year, you save hardly anything even if the time is halved. That is why When you should not automate a process says that rare tasks usually do not pay off.

What the customer support field study measures

Brynjolfsson, Li and Raymond analysed data on 5,179 employees in the chat support of a software company that offers business software for small and medium-sized firms in the United States. Most worked in the Philippines, and a chat lasted 40 minutes on average. An assistant based on a language model gave them suggested replies in real time, which they could accept or ignore. The assistant was introduced in stages, mostly between November 2020 and February 2021, after a seven-week randomised pilot with 50 employees.

The measurement came from operating data, not from self-reports: chats resolved per hour. The working paper reports 14 percent more (from an average of 2.12 resolved chats per hour, plus 0.30). Chats became 3.8 minutes or 9 percent shorter, and the resolution rate rose by 1.3 percentage points from a baseline of 82 percent. The effect was unevenly distributed: in the fifth with the lowest prior performance, resolved chats per hour rose by 34 percent, while the best performers saw no increase. The abstract of the journal version (Quarterly Journal of Economics 2025) gives 15 percent for 5,172 employees. I could not retrieve the full text, so whether the study reports 14 or 15 percent depending on the version is documented only from the abstract.

The authors write themselves that their data allow no statement on wages or employment, and that the results need not hold for other firms and processes. Their case had a stable product and always similar questions. The assistant supports people in conversation and does not automate an office process.

What the Danish survey measures

Humlum and Vestergaard invited 115,000 employees in Denmark to a survey in November and December of 2023 and of 2024, and received about 25,000 valid responses per round from about 7,000 workplaces. It covers 11 occupations considered particularly suited to chatbots, among them office clerks, accountants, marketing and customer service. Trades occupations are not included. The responses were linked to register data on income and working hours.

The time saving here is a self-report in rough bands: less than 15 minutes a day, 15 to 60 minutes, or more than an hour. From that the authors calculate the share of the users’ total working time. In the version of May 2025 the mean is 2.8 percent, between 0.6 and 6.8 percent depending on occupation and employer policy. The typical user uses the chatbot only every third or fourth working day, and on days of use respondents report 25 minutes saved on average. Where employers encourage use, the value is higher (3.6 against 2.2 percent). The revised version of March 2026 gives about 3 percent.

The register data show no effect: the authors compare employees in workplaces that encourage chatbot use with those elsewhere, before and after the launch of ChatGPT. They find no significant effect on earnings and recorded working hours. The May 2025 version rules out effects above 1 percent, the March 2026 version those above 2 percent. The authors themselves point out that self-reports can overstate the saving.

Why the figures cannot be mixed

The 14 and the 3 percent differ in three respects. The activity is different: chat support against eleven office and knowledge occupations. The way of measuring is different: operating data on resolved cases against a self-assessment in three bands. And the reference is different: performance per hour in one task against a share of total working time. The Danish authors themselves attribute the gap to such experiments to different occupations, infrequent use and a lack of accompanying measures by employers. Anyone who puts both values in one sentence is therefore not comparing like with like. The same principle applies to red tape figures, covered in Red tape and working time: what 3, 7 and 23 percent measure.

For your office this means: neither study shows whether you will reach 3 or 14 percent. Neither has examined trades businesses or workflow automation.

What the display in n8n means

n8n is software for such workflows, which you can run yourself or rent as a service. On request it shows a “Time saved” figure. According to the vendor’s documentation (as of November 2026), for this you enter a fixed number of minutes per run for each workflow, or you set time values at points within the workflow. n8n extrapolates that over the number of production runs. So the display multiplies your estimate by a count. The number of runs is a useful guide for “cases per month”, even if a run does not always match exactly one case. As evidence for the minutes, the display is no use.

What you can do now

1. Choose a recurring task and count how often it occurred over the last three months.
2. Time 15 to 20 real cases one by one, not in a block, and note the minutes.
3. After the change, measure in the same way, including checking and correcting, and set the hours saved against the hours for set-up and operation.
4. In n8n or another tool, enter only time values you have measured yourself, and label them as such.