Start with a task that happens often, works the same way every time, needs little judgement, and eats a real number of minutes each month. That combination is what makes automation pay back. It is almost never the task you would have named off the top of your head, because the tasks you remember are the ones that irritated you, and irritation is a poor guide to cost.
This article is about the step before choosing a tool: working out which of the forty or so things you repeat every month is even worth examining. It ends with a shortlist and a decision for each item on it.
- You remember loud tasks, not expensive ones. The two overlap less than people expect.
- Seven days of rough logging beats any amount of recall, as long as the log survives a bad day.
- Rank by minutes per month, a number you can actually measure, rather than a composite score you invent.
- Stability decides whether an automation pays back. A task whose rules move every month will cost more than it saves.
- Automate is one of four possible answers, and it is not the most common one.
Why the task you want to automate is usually the wrong one
Ask someone what they would automate and they answer immediately. That speed is the problem. The answer arrives from memory, and memory of your own work is skewed in a consistent direction.
What sticks is interruption, unpredictability and indignity. Re-entering the same figures into a second system sticks because it feels beneath you. Chasing someone for a file sticks because it went wrong. Meanwhile the forty minute task you do every Friday afternoon in a calm and orderly way leaves almost no trace, and it may be the single largest recurring cost in your month.
Annoyance measures how a task feels while you do it. Cost measures how much of your month it takes. Nothing forces those two to agree.
There is a second distortion. Tasks that are spread thin hide well. Five minutes, six times a day, is two and a half hours a week, and you will never notice it, because no single instance is worth noticing. Tasks that arrive in one block are visible whether they matter or not.
So the audit is not really a measurement exercise. It is a correction for a known bias, and that is why it is worth seven days of mild inconvenience.
Design the audit so a bad week cannot kill it
Most time audits fail for exactly the reason productivity systems fail, which is that they ask for continuous discipline and then get abandoned on the first day that goes badly. An audit that requires discipline ends up measuring your discipline rather than your work.
So build it to survive being done badly.
- Log three times a day, not continuously. Late morning, mid afternoon, end of day. Write down what you repeated since the last check. Recall over two or three hours is good enough and costs about ninety seconds each time.
- Cap it at six fields. Task, how long, how many times, what triggered it, whether it needed judgement, and whether it went wrong. More fields means better data in theory and no data in practice.
- Estimate in fives. Five, ten, fifteen, thirty, sixty minutes. Precision beyond that is false anyway, since you are recalling rather than timing, and rounding removes the excuse to skip an entry because you are not sure.
- Treat a missed day as normal. Do not backfill and do not restart. Six days of honest logging is a usable sample. A restarted audit is almost always an abandoned one.
- Add the invisible items at the end. On day seven, spend ten minutes listing the monthly and quarterly things a seven day window could not catch. Reporting cycles, invoicing, renewals, reviews.
Keep it in whatever you already have open. A notes file or the back of a notebook is fine. Setting up a tracking system for the audit is procrastination with a spreadsheet.
Record what triggered each task. A surprising number of recurring tasks are triggered by "someone remembered", and those are worth fixing before anything else, usually with a recurring reminder rather than software. A task that only happens when you think of it is not a process, it is a liability.
What to record, and why each field earns its place
| Field | Example | What it decides later |
|---|---|---|
| Task | Copy pricing pages into the comparison doc | Whether two log entries are the same task |
| Minutes | 15 | Half of the cost calculation |
| Times | 3 this week | The other half, and payback speed |
| Trigger | Monday reminder | Whether the task is reliable or depends on memory |
| Judgement | Low | Whether a machine can do it at all |
| Went wrong | Twice, wrong figures pasted | Error cost, which changes the answer entirely |
Six fields, one line per entry. If a line takes longer than fifteen seconds to write, the audit is too heavy and you will stop.
Download the audit template as a CSV if you would rather start from a file. It is six columns and a few example rows, nothing more.
Turn the log into a shortlist without inventing a score
There is a temptation here to build a formula: multiply frequency by time by friction, divide by difficulty, sort the column. It looks rigorous and it mostly is not, because you are multiplying numbers you invented on scales you invented, and the ranking will swing wildly depending on choices you made arbitrarily.
Use a real number instead.
Monthly cost equals times per month multiplied by minutes per run. That is measurable, it is comparable across tasks, and when you tell someone a task costs four hours a month, it means something. Sort the list by that and stop.
Now apply three filters to the top of the list. These are not scores, they are questions with consequences.
Does it need judgement? If a competent new colleague could not do it correctly from a written instruction, a machine cannot either. High judgement tasks are not automation candidates, though the collection and formatting around them usually are.
Is it stable? Has the way you do this changed in the last three months? If yes, anything you build now becomes maintenance. This is the filter people skip, and it is the most common reason an automation is quietly abandoned six weeks later.
What does a mistake cost? An error that produces a malformed internal file is cheap. An error that sends a client a wrong figure is not, and it means the first version needs a checking step, which roughly doubles the work of building it.
The task at the top of your sorted list that also passes all three filters is where to start. Frequently there is nothing at the very top that passes, and the right first target sits third or fourth. That is a normal result and a useful one.
Four decisions, and automate is only one of them
Every task on the shortlist gets exactly one of these. Deciding between them takes a minute and saves weeks.
Eliminate. Name the person who reads the output. If you cannot, stop producing it and wait to see who objects. Recurring work accumulates in every job because starting something is a decision and stopping it never is. This option is chosen far less often than it should be, and it is the only one with an instant payback and no maintenance.
Simplify. Reduce the number of steps, tools or approvals before touching any software. Half the tasks that look like automation candidates are really a two system problem, where the same information is entered twice because nobody has connected the two systems or, more often, because nobody has asked whether the second system needs it at all.
Automate. Frequent, stable, low judgement, meaningful monthly cost, tolerable error cost. Build the smallest version that handles the common case and stops when it meets something unfamiliar.
Keep human. Some tasks are the work. Deciding, judging, deciding not to, and any conversation where the point is that a person had it.
Unread output that arrives reliably looks like a working system. Once it is automated, the last signal that anyone would have noticed its absence is gone, and it will still be running in three years. Ask the eliminate question before the automate question, always in that order.
Tasks that look automatable and are not
Four patterns come up often enough to be worth naming.
The task that is about to disappear. A system migration, a process review or a departing colleague is going to change the whole shape of this work in two months. Building now means building twice. Ask what is scheduled to change before you start.
The task with quiet failure. Some errors announce themselves and some do not. A script that silently skips two rows because the input format shifted will keep producing plausible output indefinitely. If you cannot describe how you would notice a wrong result, you are not ready to automate it.
The task that is really a relationship. The monthly check in that is nominally about a report is often the only regular contact you have with someone whose cooperation you need later. Automating the report can quietly remove the reason the conversation happens.
The task that varies more than it appears. It looks identical every time until you write down the exceptions, and then it turns out you make a judgement call in one run out of five without noticing. Our guide on mapping a workflow before you automate it covers how to surface those exceptions, and it is the natural next step once the shortlist exists.
A worked example
Here is the shape the process usually takes, using a common case: someone in a small marketing team who is certain that expense reports are the thing to fix, because they hate doing them.
The audit says otherwise. Expenses turn out to be about twenty five minutes, once a month. Real, irritating, and twenty five minutes.
Meanwhile the log shows the same entry appearing on most days: gathering content from competitor and partner pages into a working document, ten to fifteen minutes at a time, several times a week. Nobody had ever thought of it as a task. It was just the thing you do before the real work starts. Multiplied out, it is a few hours a month, and it fails all the memorability tests because each instance is short and unremarkable.
Run the filters. Judgement is low for the collection part, since deciding what matters happens afterwards. It is stable, because the pages change but the act of pulling text off a page does not. Error cost is low, and a mistake is obvious immediately.
So the first target is a step inside a task nobody had named, rather than the task that prompted the whole exercise. That is the usual outcome, and it is why the audit is worth doing rather than skipping straight to a tool.
Note what does not follow. The judgement half of that work, reading the material and deciding what it means, stays exactly where it was. What gets automated is the collection, which is the part that was never the point.
The collection step is usually the first win
Pulling clean text out of pages, PDFs, presentations and metadata is the step that shows up at the top of most shortlists, because it is frequent, stable and needs no judgement. Our tools do that part and nothing else. Free, and they run in your browser.
See the toolsWhat to do with the rest of the list
Resist starting three things. The value of this exercise is the ranking, and the ranking is only useful if you act on it one item at a time. Finish the first automation, use it for a month, then come back to the list with a better sense of what building actually costs you.
Keep the log. In three months it becomes a comparison rather than a memory, and the second audit takes two days instead of seven because you already know what the entries look like.
Frequently asked questions
Which tasks should I automate first?
Frequent, stable, low judgement tasks with a real monthly cost. Rank candidates by times per month multiplied by minutes per run, then remove anything that needs judgement, has changed recently, or fails expensively without warning. Whatever is left at the top is your answer.
How long should the audit run?
Seven working days. That catches daily and weekly patterns, which is where the hidden cost usually sits. Add monthly and quarterly items from memory at the end, since a week cannot see them. Running it longer rarely improves the shortlist because most people stop logging honestly after the first week.
Should I automate the task that annoys me most?
Not first. Annoyance tracks interruption and unpredictability rather than time. Once you have ranked by measured cost, let annoyance break ties between similar candidates, since a task you dread has a real cost in avoidance and procrastination that the minutes do not capture.
What if my work is different every day?
Then the audit is more valuable, not less. Genuinely varied work still contains repeated mechanical steps, usually around finding, formatting and moving information. Those steps are the candidates. If seven days of logging turns up no repeated entries at all, your work is judgement work and the honest answer is that there is nothing here to automate yet.
Is it worth automating something that only takes five minutes?
It depends entirely on how often. Five minutes twice a day is over three hours a month and is usually worth it. Five minutes twice a month is twenty minutes a year of savings against however long you spend building, and it is not. Frequency, not duration, decides this one.
How do I know when to stop auditing and start building?
When you have one task that survives all three filters. One is enough. The point of the audit was never a complete inventory, it was to stop you building the wrong thing first.
Where to start this week
Open a blank note. For the next seven working days, three times a day, write one line per repeated task using the six fields above. On day seven, add the monthly items, sort by minutes per month, and apply the three filters to the top five.
Then take the one task that survives and map it properly before you build anything, which is the subject of the next article in this series. The audit tells you where to look. The map tells you what to build.
How this was put together: the framework here comes from our own experience building small tools for recurring problems, and from the patterns we see in the workflows those tools get used for. It is not a controlled study, and we have deliberately not quoted percentages or average time savings, because we have not measured them at a scale that would make such figures honest. The monthly cost calculation is arithmetic on your own numbers, so it holds regardless. Treat the three filters as judgement aids rather than rules.