TL;DR. The xkcd chart “Is It Worth the Time?” shows how long you can spend speeding up a routine task before you lose time over five years. Its cells are 24-hour days: the “6 days” allowed for shaving 5 minutes off a daily task is about 152 hours, or 19 eight-hour working days, and the chart has no term for maintenance, failures or tool fees. In a CEOtudent recomputation that adds 7 hours a year of upkeep and a 2% failure rate, a 1-minute daily saving goes from 30 hours of allowed build time to minus 11. For the value of an hour, BLS data show employers paid $49.46 per hour worked for civilian workers in June 2026 ($33.85 of it wages), and Eurostat puts the 2025 EU average hourly labour cost at €34.9. The CEO move: run every idea through one formula and a four-gate rule before building. The student move: log the manual task for four weeks first, because guessed inputs are what sink the estimate.
This piece is the calculator. Why upkeep quietly turns small wins into losses, with an Automate Now / Wait / Never scorecard, is covered in Automation Debt: Why Automating Too Fast Creates More Work Than It Saves. Which tasks deserve a calculation at all is covered by the priority score in The 80/20 of Knowledge Work Automation. If you have not yet measured where your hours go, start with The Workflow Audit.
This article is general information about estimating time and cost. It is not personalised financial advice.
What does the famous “is it worth the time” chart actually measure?
xkcd comic 1205 (April 2013) asks: “How long can you work on making a routine task more efficient before you’re spending more time than you save? (Across five years)”. Rows are the time shaved off (1 second to 1 day), columns are frequency (50 times a day to yearly), and each cell is the total time saved over five years.
The grid is a break-even ceiling with three built-in assumptions:
- Every calendar day counts. A daily task runs 365 times a year; most work tasks happen on working days only.
- Cells are 24-hour days. The 5-minutes-daily cell is 152.1 hours: 19.0 eight-hour days, or 18.8 days at the 8.1 hours that full-time US workers worked on days they worked in 2025, per the BLS American Time Use Survey.
- Zero upkeep. The automation never breaks, needs no checking and costs nothing to run.
xkcd comic 1319 (“Automation”, January 2014) draws the gap: in its “reality” panel, writing code turns into “debugging”, “rethinking” and “ongoing development”, and the original task never shrinks. Google’s Site Reliability Workbook reproduces that comic and states the test: time saved should be at least proportional to “the time invested in first developing and then maintaining an automated solution”. The 2013 grid leaves out the second half.
What is an hour of your time actually worth?
The formula needs a value per hour (V). Three lenses give different answers. Employer cost: for paid work, an hour costs wage plus benefits and employer taxes, which the BLS Employer Costs for Employee Compensation (ECEC) series measures. Wage benchmark: the median hourly wage for your occupation group (BLS OEWS) or Eurostat’s hourly labour cost. What the freed hour becomes: a saved hour that dissolves into extra email is worth close to nothing, and only you can judge this.
| Benchmark | Figure as published | Period | Source |
|---|---|---|---|
| Employer cost per hour worked, civilian workers | $49.46 | June 2026 | BLS ECEC |
| of which wages and salaries / benefits | $33.85 / $15.61 | June 2026 | BLS ECEC |
| Employer cost per hour worked, private industry | $46.89 (benefits 30.0%) | June 2026 | BLS ECEC |
| Median hourly wage, all occupations | $24.51 | May 2025 | BLS OEWS |
| Median hourly wage, management | $60.83 | May 2025 | BLS OEWS |
| Median hourly wage, computer and mathematical | $52.54 | May 2025 | BLS OEWS |
| Median hourly wage, business and financial operations | $39.74 | May 2025 | BLS OEWS |
| Median hourly wage, office and administrative support | $22.81 | May 2025 | BLS OEWS |
| Average hourly labour cost, EU / euro area | €34.9 / €38.2 | 2025 | Eurostat |
| Lowest and highest EU country | €12.0 (Bulgaria), €56.8 (Luxembourg) | 2025 | Eurostat |
Sources: BLS ECEC June 2026 news release; BLS OEWS May 2025 Table 1; Eurostat, Hourly labour costs, Statistics Explained (data extracted 31 March 2026). Eurostat labour costs include non-wage costs, so they are closer to the ECEC concept than to a wage.
OEWS medians are wages only. One simple conversion scales them by the ECEC civilian ratio of total compensation to wages: $49.46 / $33.85 = 1.461, so benefits are 31.6% of total cost.
| Occupation group | Median wage per hour | Loaded value per hour (x 1.461) |
|---|---|---|
| All occupations | $24.51 | $35.81 |
| Office and administrative support | $22.81 | $33.33 |
| Business and financial operations | $39.74 | $58.07 |
| Computer and mathematical | $52.54 | $76.77 |
| Management | $60.83 | $88.88 |
CEOtudent analysis of BLS OEWS May 2025 and BLS ECEC June 2026. The civilian average ratio is applied to every group; real benefit shares differ (30.0% in private industry, 38.8% in state and local government), so the right-hand column is an approximation.
For personal tasks such as household admin, the employer-cost lens does not apply; use your after-tax hourly pay as an upper reference and the third lens as the real test. Scale matters more than precision: at $58.07, one hour a week for a year is worth about $3,020.
What goes into the full net-value formula?
The calculator adds what the classic chart omits: failures, upkeep, tool fees and incidents, plus a capped learning credit.
Net value = V * [ H*n*(1-f)*t/60 - H*n*f*r/60 - H*(M + W) - B + O ] - H*(S + q*L)
Payback (weeks) = B / weekly net hours
weekly net hours = [ n*((1-f)*t - f*r)/60 - M - W - (S + q*L)/V ] / 52
t = minutes saved per successful run; n = runs per year; f = share of runs that fail; r = extra minutes per failure; M and W = maintenance and monitoring hours per year; B = build hours including learning and testing; O = learning credit in hours; H = horizon in years (the shorter of task life and tool life); V = value of an hour; S = tool cost per year; q and L = expected incidents per year and the cost of each. CEOtudent editorial framework.
Failures cost more than the lost saving. A failed run loses its minutes and adds time to notice and redo, so f and r enter separately. Incidents are separate from routine failures. A report that did not generate is a failure; an email to the wrong client or a duplicated payment is an incident, rare but expensive, so it is modelled as q x L. Learning credit is capped. Google’s workbook lists indirect benefits a pure hours comparison misses, including “Enhanced technical skills and career growth” and “Increased process clarity and standardization”. They are real and easy to inflate. CEOtudent editorial rule: credit learning only when you can name the next automation that reuses the skill, and cap O at half the build hours.
How does the xkcd grid change once maintenance and failure are added?
The table recomputes the core of the 2013 grid in hours. Each cell shows the original xkcd ceiling (five years, no upkeep), the adjusted five-year ceiling and the adjusted two-year ceiling. The adjustment uses one set of illustrative assumptions: 2% of runs fail, each failure costs 10 extra minutes, maintenance takes 6 hours a year and monitoring 1 hour a year. Frequencies follow the comic (365 days a year).
| Time saved per run | 5 times a day | Daily | Weekly | Monthly |
|---|---|---|---|---|
| 30 seconds | 76 / 9 / 4 | 15 / -26 / -10 | 2 / -34 / -13 | 0 / -35 / -14 |
| 1 minute | 152 / 84 / 33 | 30 / -11 / -5 | 4 / -32 / -13 | 1 / -34 / -14 |
| 5 minutes | 760 / 680 / 272 | 152 / 108 / 43 | 22 / -15 / -6 | 5 / -30 / -12 |
| 30 minutes | 4,562 / 4,406 / 1,762 | 912 / 853 / 341 | 130 / 92 / 37 | 30 / -6 / -2 |
| 1 hour | 9,125 / 8,877 / 3,551 | 1,825 / 1,747 / 699 | 260 / 219 / 88 | 60 / 24 / 9 |
Cells: xkcd ceiling / adjusted 5-year / adjusted 2-year, in hours of build time affordable before the automation loses time. CEOtudent analysis recomputing xkcd comic 1205 with illustrative maintenance, monitoring and failure assumptions; arithmetic in calc.py.
A negative cell means the automation loses time even with zero build time. Under these assumptions the minimum saving per run just to cover upkeep and failures is 0.44 minutes for a 5-times-a-day task, 1.38 minutes for a daily task, 8.45 minutes weekly and 35.92 minutes monthly.
- Small savings on infrequent tasks almost never pay. The weekly and monthly columns below 30 minutes all go negative; the original chart still shows small positive allowances.
- Frequency pays for upkeep. At 5 times a day, even a 1-minute saving keeps 84 hours of five-year headroom.
- Horizon is the biggest lever. Moving from five years to two cuts every positive adjusted cell by 60%, because the allowance is the yearly net saving multiplied by the years. If the task may not exist unchanged in two years, use the two-year number.
Why is maintenance the term most people leave out?
Personal automations depend on things you do not control: a page layout, an export format, an API, a colleague’s template. How fast those change is the best guide to M and H.
| Source | What was measured | Figure as published |
|---|---|---|
| Pew Research Center (2024) | Webpages existing at some point 2013-2023 that were inaccessible in October 2023 | A quarter |
| Pew Research Center (2024) | Pages from 2013 not available; pages from 2023 not available | 38%; 8% |
| Ochoa et al. (2021, abstract) | Maven Central library upgrades complying with semantic versioning; clients affected by breaking changes | 83.4%; 7.9% |
| Panko (2008) | Accuracy per formula or line of code; chance of an error in a sheet with 10 root formulas | 95%-98%; 18% to 40% |
| Google SRE Workbook (2018) | Availability of a service with nine dependencies, each at 99.99% | 99.9% |
| Google SRE book (company-reported) | Average share of time Google SREs report on toil in quarterly surveys; target cap | About 33%; below 50% |
| Dehaghani and Hajrahimi (2013) | Share of software life-cycle cost attributed to maintenance, citing earlier estimates | About 90% |
Full citations in the Sources list. Ochoa figures are from the abstract only. The 90% figure is an estimate the authors cite from earlier literature, not their own measurement; published maintenance shares vary widely by definition and era.
- Page-dependent flows decay fastest. Pew found 8% of pages that existed in 2023 were gone by October that year. If an automation reads a specific page, assume a short horizon and a high M.
- Versioned interfaces break less than feared. Analysing 119,879 library upgrades, Ochoa and colleagues found most breaking changes hit code no client used. A stable, versioned interface earns a longer H than a scraped screen.
- The external world drives the cost. Sculley and co-authors at Google, writing about machine learning systems: “Experience has shown that the external world is rarely stable. This background rate of change creates ongoing maintenance cost.”
With no history, log every minute spent fixing or adjusting the automation for eight weeks and annualise it. Writing the manual procedure down first also cuts maintenance, as the SOP Renaissance piece argues: much breakage starts where the process was never specified.
How do failures compound in a multi-step automation?
Most personal automations are chains of steps. If each step works independently with probability p, the chain works with p to the power of the step count. The SRE Workbook uses this product rule for availability (0.9999 to the tenth power is 99.9%), and Panko uses it for spreadsheets: at 95% to 98% accuracy per formula, a sheet with 10 root formulas has an 18% to 40% chance of an error (recomputed here as 18.3% and 40.1%).
| Reliability of each step | 3 steps | 6 steps | 10 steps |
|---|---|---|---|
| 99.9% | 99.7% | 99.4% | 99.0% |
| 99.5% | 98.5% | 97.0% | 95.1% |
| 99.0% | 97.0% | 94.1% | 90.4% |
| 98.0% | 94.1% | 88.6% | 81.7% |
| 95.0% | 85.7% | 73.5% | 59.9% |
CEOtudent analysis applying the product rule described in the Google SRE Workbook and Panko (2008); per-step reliabilities are illustrative assumptions and steps are assumed independent.
Six steps at 99% each succeed on 94.1% of runs: about 14.6 failed runs a year for a task run 250 times. Estimate f per chain, not per tool; cutting steps often beats adding error handling.
Lisanne Bainbridge’s 1983 paper “Ironies of Automation” adds the human side: “it is the most successful automated systems, with rare need for manual intervention, which may need the greatest investment in human operator training.” For a personal automation, r grows over time because the manual fallback is a skill you stop practising.
How much repetitive time is there to win back?
The BLS American Time Use Survey for 2025 reports that the civilian population spent an average of 0.17 hours a day on household management, a category covering financial management, household organisation and planning, and household mail and messages. That is about 10.2 minutes a day, or roughly 62 hours a year. On an average day 18.6% of people did some, and they spent 0.92 hours (about 55 minutes). Purchasing goods and services took 0.69 hours a day, about 252 hours a year.
For paid work, none of the official sources opened for this piece measures the share of knowledge workers’ time spent on repetitive digital tasks; vendor surveys on this exist and are not used. The most transparent measured example is company-reported: Google says quarterly surveys of its site reliability engineers show average toil of about 33%, with answers from 0% to 80%. Its definition of toil, work that is “manual, repetitive, automatable, tactical, devoid of enduring value, and that scales linearly as a service grows”, is a good checklist for your own week.
Recoverable admin time for most individuals is tens of hours a year, not hundreds: enough for a few well-chosen automations, not a dozen fragile ones.
What do worked examples look like with every term included?
Four typical personal automations; every input is an illustrative assumption, not a measurement. V is $58.07, the loaded business-and-finance value above.
| Input (illustrative assumption) | A: weekly status report | B: daily inbox-to-task rule | C: monthly invoice reconciliation | D: daily web price check |
|---|---|---|---|---|
| Minutes saved per run (t) | 25 | 3 | 90 | 4 |
| Runs per year (n) | 48 | 230 | 12 | 365 |
| Build hours (B) | 8 | 3 | 20 | 10 |
| Maintenance + monitoring hours/yr (M + W) | 4 + 0.5 | 2 + 0.5 | 6 + 1 | 12 + 1 |
| Failure share (f) / extra minutes per failure (r) | 5% / 20 | 2% / 5 | 8% / 45 | 10% / 8 |
| Horizon (H) | 2 years | 2 years | 3 years | 1 year |
| Tool cost per year (S) | $0 | $0 | $120 | $60 |
| Incidents per year x cost (q x L) | 0 | 0 | 0.10 x $500 | 0 |
| Learning credit (O) | 0 | 0 | 5 hours | 0 |
| Result | A | B | C | D |
|---|---|---|---|---|
| Gross value of time saved | $2,207 | $1,309 | $2,885 | $1,272 |
| Build | -$465 | -$174 | -$1,161 | -$581 |
| Maintenance and monitoring | -$523 | -$290 | -$1,220 | -$755 |
| Failures and incidents | -$93 | -$45 | -$275 | -$283 |
| Tools | $0 | $0 | -$360 | -$60 |
| Learning credit | $0 | $0 | +$290 | $0 |
| Net value over horizon | $1,127 | $800 | $159 | -$407 |
| Payback | 30 weeks | 19 weeks | 176 weeks | 173 weeks |
| Net under stress test | $134 | $263 | -$1,486 | -$1,491 |
| Verdict (four-gate rule) | Build | Build | Wait | Do not build |
CEOtudent analysis; all inputs are illustrative assumptions; arithmetic reproduced in calc.py. Stress test: saving 25% lower, maintenance and monitoring 50% higher, failure share doubled.
- A passes, narrowly under stress. Build, upkeep and failures consume 49% of its gross value.
- B is the quiet winner. Tiny build, short payback, positive stress result: the typical profile of a few-step automation of a frequent task.
- C is positive only because of the learning credit. Without it the net is -$131, and its payback exceeds half its horizon. Simplify the process or keep a manual approval step, then recalculate.
- D loses money despite running daily. Upkeep and failures absorb 82% of its gross value, and a page-dependent flow justifies only a short horizon. The original chart would have approved it.
The fill-in worksheet
Fill the “Your value” column from a four-week log of the manual task, then compute the bottom lines.
| Input | Symbol | Your value | How to estimate it honestly |
|---|---|---|---|
| Minutes saved per successful run | t | Time 4 or more manual runs, take the median, subtract review time you keep | |
| Runs per year | n | Count real occurrences in your log, not intentions | |
| Build hours | B | First estimate x 1.5 if the tool is new to you | |
| Maintenance hours per year | M | Log fixes for 8 weeks after launch; before launch, 30 minutes a month | |
| Monitoring hours per year | W | Weekly checking minutes x 52 / 60 | |
| Failure share | f | Per-step reliability to the power of the step count (chain table) | |
| Extra minutes per failure | r | Notice + redo + clean up | |
| Horizon in years | H | Shorter of task life and tool stability | |
| Value of one hour | V | Loaded wage for paid work; judgement for personal time | |
| Tool cost per year | S | Subscription plus per-run fees | |
| Incidents per year, cost of each | q, L | Worst wrong output, and how often it could slip through | |
| Learning credit | O | Zero unless the next reuse is named; at most B / 2 | |
| Gross hours | H x n x (1 – f) x t / 60 | ||
| Cost hours | H x n x f x r / 60 + H x (M + W) + B – O | ||
| Net value | V x (gross hours – cost hours) – H x (S + q x L) | ||
| Payback weeks | B / weekly net hours | ||
| Stress-test net | Recompute with t x 0.75, (M + W) x 1.5, f x 2 |
CEOtudent editorial framework.
The four-gate decision rule
| Gate | Test | If it fails |
|---|---|---|
| 1. Positive base case | Net value over the horizon is above zero | Do not build. Eliminate, batch or simplify the task. |
| 2. Payback inside half the horizon | Payback weeks no more than H x 52 / 2 | Wait. Shrink the scope or cut steps, then recompute. |
| 3. Survives the stress test | Stress-test net is zero or above | Wait. The estimate rests on optimistic inputs; log four more weeks. |
| 4. Contained downside | Expected incident cost (H x q x L) under 10% of gross value, and no unreviewed irreversible step | Build only with a human approval step on the irreversible action, then recompute. |
CEOtudent editorial framework. All four pass = Build; gate 1 passes but 2 or 3 fails = Wait; gate 1 fails = Do not build.
Gate 1’s fallback mirrors the SRE Workbook: “rejecting a toil-intensive task should be the first option you consider.” The cheapest automation is the task you stop doing. Gate 4 follows the delegation logic in Augment, Don’t Automate: reversible, low-stakes steps can run unattended, irreversible ones keep a person in the loop, and that review time belongs in t.
Where do the CEO and the student meet in this calculation?
The CEO half is capital allocation: build hours are an investment, maintenance a recurring cost, an incident a liability. No CEO approves a project without a payback period and a downside case; the four gates are a one-person investment committee.
The student half is measurement. Guessed inputs tend to be wrong in the flattering direction: runs overstated, maintenance assumed zero, failures assumed not to compound. Four weeks of logging the manual task teaches more about the process than the build, and Bainbridge’s irony says the manual skill is worth keeping alive after launch. Treat the first eight weeks after launch as a test of the payback figure.
What this does not mean
- Small automations are not bad. Example B shows the opposite; the method penalises fragility and infrequency, not size.
- The numbers are not precise. The hourly benchmarks are official; every worked-example input is an illustrative assumption.
- Non-time benefits still count. Fewer errors or less stress can justify a small time loss; enter them as incident reduction or a capped learning credit so they stay visible.
- It is built for one person. When many people use one automation, n multiplies and the economics change.
Frequently asked questions
How do you calculate whether an automation is worth building?
Multiply minutes saved per successful run by runs per year and the horizon, convert to hours, subtract build, maintenance, monitoring and failure time, multiply by the value of an hour, then subtract tool fees and expected incident costs. Build if the result is positive, pays back within half the horizon and survives a stress test.
Is the xkcd “Is It Worth the Time?” chart still useful?
Yes, as a ceiling with zero upkeep. Its “6 days” for a 5-minute daily saving is about 19 eight-hour working days, and with maintenance and failures added many smaller cells turn negative.
What hourly value should you use for your own time?
For paid work, a loaded wage: your wage scaled by the employer cost ratio, about 1.461 for US civilian workers in June 2026. For personal time, your own judgement, with after-tax hourly pay as an upper reference. Eurostat’s 2025 EU average hourly labour cost was €34.9.
How much maintenance should you assume for a new automation?
Log it for eight weeks and annualise. Before launch, the grid here uses 6 hours a year plus 1 hour of monitoring as an illustrative start. Page-dependent flows deserve more: Pew found 38% of pages from 2013 were gone a decade later.
Why do multi-step automations fail more than single tools?
Reliability multiplies. Six steps that each work 99% of the time succeed together 94.1% of the time, about 14.6 failed runs a year for a task run 250 times.
Should learning value count in the ROI?
Only with limits. Google’s SRE Workbook lists skill growth and process clarity as real indirect benefits. CEOtudent’s rule: credit learning only when the next reuse is named, capped at half the build hours.
Sources
- Munroe, R. “Is It Worth the Time?” xkcd comic 1205, April 2013.
- Munroe, R. “Automation.” xkcd comic 1319, January 2014.
- U.S. Bureau of Labor Statistics. Employer Costs for Employee Compensation – June 2026. News release USDL-26-1494, 9 September 2026.
- U.S. Bureau of Labor Statistics. Occupational Employment and Wages – May 2025. News release and Table 1.
- U.S. Bureau of Labor Statistics. American Time Use Survey – 2025 Results. News release and Table 1.
- U.S. Bureau of Labor Statistics. American Time Use Survey activity coding lexicon (household management categories).
- Eurostat. Hourly labour costs. Statistics Explained, data extracted 31 March 2026 (dataset lc_lci_lev).
- Rau, V. “Eliminating Toil.” Chapter 5 in Site Reliability Engineering: How Google Runs Production Systems, ed. Beyer, B. et al. O’Reilly Media and Google, online edition. Company-authored.
- Google SRE team. “Eliminating Toil.” Chapter 6 in The Site Reliability Workbook. O’Reilly Media and Google, 2018. Company-authored.
- Sculley, D. et al. “Hidden Technical Debt in Machine Learning Systems.” Advances in Neural Information Processing Systems 28, 2015.
- Dehaghani, S. M. H., Hajrahimi, N. “Which Factors Affect Software Projects Maintenance Cost More?” Acta Informatica Medica 21(1), 2013, 63-66.
- Ochoa, L., Degueule, T., Falleri, J.-R., Vinju, J. “Breaking Bad? Semantic Versioning and Impact of Breaking Changes in Maven Central.” arXiv:2110.07889, 2021 (abstract only).
- Pew Research Center. “When Online Content Disappears.” 17 May 2024.
- Panko, R. R. “Thinking is Bad: Implications of Human Error Research for Spreadsheet Research and Practice.” arXiv:0801.3114, 2008.
- Bainbridge, L. “Ironies of Automation.” Automatica 19(6), 1983, 775-779.
CEOtudent analyses: the loaded hourly value table, recomputed xkcd grid, chained-step table, worked examples, formula, worksheet and four-gate rule. Grid and example inputs are illustrative assumptions; every derived figure is reproduced in a calculation script. Deliberately not used: vendor surveys on time spent on repetitive tasks, widely repeated textbook ranges for the maintenance share of software cost (no primary text opened), and any single figure for the value of personal non-work time.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:















