In this guide 6 sections
Find automatable manual work by watching tasks as they really happen, then log how often each runs, how long it takes and how often it hits an exception. Score candidates on time cost and risk before choosing one pilot, because automating a messy process only produces the same mess at higher speed.
Observing how the work happens
Written procedures describe the intended task. Watched in person, the real task usually differs, with shortcuts, a check added after one old mistake, or a second look because someone distrusts the first system. Automate the real version.
Spend time beside the person doing each task that comes up in complaints or visibly fills a large part of their day. Typical candidates are timecard entry, reconciliation, status updates and report compilation, so ask each person what they would do with an extra hour. Then ask what uses that hour today. Look for paper that gets retyped, from supplier delivery notes to a shoebox of expense receipts, and any figure typed twice by different people.
Frequency, handling time and exceptions for each candidate
Measure three things for every candidate task, each taken over a representative period of normal work rather than a single convenient afternoon.
Multiply frequency by time for a weekly cost in hours, and keep the exception count beside it, because together they show how safe the task is to automate. Exception rate matters as much as time. A ten-minute task that runs cleanly every time is often a safer candidate than a five-minute task that needs a judgment call on one attempt in five.
- How often the task happens, counted across a whole month-end cycle so that the busiest days and the quietest days both appear in the count.
- How long it takes each time, timed with a stopwatch or a screen recording instead of an estimate from memory offered in a meeting.
- How often something unusual interrupts the normal flow, such as a missing field, an odd customer request or an approval that has to be chased by phone.
Clean data and stable rules
Automation amplifies whatever it is given. If the data feeding a task holds duplicates, mixed date formats or empty fields, automating it multiplies errors faster than a person could catch them by eye. Data need not be perfect. It must be consistent enough for a rule to work reliably, so deduplicate the records and fill blank fields before anything is switched on.
Check that the rules are genuinely rules. A task can look routine yet depend on unwritten judgment, such as waiving the fee for one long-standing client or approving a late order from a trusted supplier. Write those exceptions down first. A rule hardcoded into a macro by someone who left is the classic trap: nobody knows what it does, so nobody dares touch it. Until then it is not ready.
Scoring against effort and risk, not only time saved
Rank candidates on two axes. The first is how much time or cost the task consumes today, and the second is how much damage it would do if the automation got something wrong. Compiling a report from three sources is a safer first candidate than a task that touches customer payments or fiscal filings, however much that task seems to save.
Suppose an accounts team retypes the same purchase order details into three separate systems several times a day, and a slip affects only one order at a time. High frequency with contained damage goes first. A lower-volume payday run waits.
One measurable pilot before expanding
Automate the highest-scoring candidate first. Define the measure of success before starting, whether hours saved per week, error rate before and after, or turnaround time from request to completion. Do not start three at once. A single pilot is easier to debug, easier to trust and easier to prove worthwhile to the people who must live with the result.
Once the pilot is stable and the team trusts it, move to the next candidate using the same scoring sheet, and report the measured result first. Staged rollout builds confidence. That matters when a change reaches staff who have done a task by hand for years, because they judge the next proposal by how the first one went.
Manual-work discovery worksheet
A sheet holding one row for each candidate task, with columns for observed frequency, average handling time, exception rate, a data-quality check, and an effort-versus-risk score. Fill it in from direct observation over one to two weeks, not from memory, so the scores reflect what actually happens on the floor and in the inbox.
Open a blank worksheet to printQuestions owners ask
How do we find hidden manual work?
Watch tasks being done instead of reading process documents, and ask staff what they would skip if they had an extra hour in the day. Hidden work clusters wherever two systems do not connect and someone bridges the gap by hand, such as copying figures from an order form into a stock sheet.
Which tasks need human judgement?
Any task whose rule is an unwritten exception applied case by case, such as a discretionary discount or a decision that depends on context missing from the data. Those need their exceptions written down before any automation, and some are better left with a person who can weigh the case.
How long should a pilot run?
Long enough to cover one full cycle of the task, such as a month-end close, with the success measure recorded before and after. Weekly tasks need several runs.