How to Identify Repetitive Work That Is Worth Automating

Where do you find the candidates?

Repetitive work hides in plain sight, because the people doing it stopped noticing years ago. It rarely appears in a process document, and almost never in a job description. It shows up in habits: the file opened every morning, the group message sent every evening, the spreadsheet rebuilt on the last working day of the month.

• Ask each team what they would stop doing tomorrow if they could, and why they cannot.
• Look at the recurring entries in people’s calendars, especially the ones marked as blocked time.
• Open the shared inbox and count the messages that get the same reply.
• Find the spreadsheets whose figures are typed in from another system.
• Watch month-end and quarter-end: peaks reveal work that is manual all year.
• Ask why overtime happens, and on which days of the week it lands.

Write each candidate as a unit of work, not a department: one supplier invoice, one dealer order, one admission enquiry, one weekly outlet report. If you cannot describe the trigger and the finish in a sentence, you are looking at a function rather than a process, and it is too large to judge.

Which signals tell you the work is worth automating?

Eight signals separate work that repays automation from work that merely feels tedious. The first four say how much the work costs you. The last four say how safely it can be handed to software.

1. Frequency: how often the work arrives, and whether it arrives evenly or in peaks.
2. Volume: how many cases in a typical month, counted rather than estimated.
3. Time per case: how long one case takes from arrival to finished record, including waiting.
4. Rework: how often a case comes back for correction, and what causes it.
5. Handoffs: how many people or systems the case passes through on the way.
6. Exceptions: what share of cases break the pattern, and whether they are predictable.
7. Data availability: whether you can reach the inputs and produce real examples today.
8. Risk: what happens if the work is done wrong, and how quickly that would be noticed.

How do you score candidates without pretending to be precise?

Score each signal from one to three, where three always means more attractive to automate. Keep it rough. The purpose is to make a conversation comparable, not to produce a number that decides for you.

Read the shape, not only the total. A candidate scoring three everywhere except data availability is not ready; it is a data access task first. A candidate scoring three on volume and one on exceptions is a candidate for automating part of the work, usually the reading and preparing, while people keep the deciding.

What are the red flags?

Some work should be left alone, however irritating it is. These signs are worth more weight than a high score elsewhere.

• It happens rarely. Building and maintaining a system costs more than the work it saves.
• The decision is a judgement call, made differently by two experienced people for good reasons.
• The process is broken. Automating it makes the faults arrive faster and less visibly.
• Nobody owns it. Without a person who answers for the outcome, quality has no home.
• The inputs cannot be reached, or exist only on paper in another office.
• The rules are disputed between teams, or change with every negotiation.

How do you measure the work as it stands today?

Scores rank candidates. A baseline tells you later whether anything actually improved. Take a fortnight of real cases, not a reconstruction from memory, and record time from arrival to finished record, waiting time, how many needed rework and how many turned out to be exceptions. Keep the awkward cases in the sample; they are the ones that decide whether a system survives contact with a normal week.

The NIST AI Risk Management Framework makes the same point for systems meant to augment or replace human activity: managing them requires some form of baseline for comparison, and it notes this is difficult to systematise because software carries out tasks differently from people. That is a reason to measure the whole workflow rather than one step. If the reading gets faster and the approval queue grows, nothing has been gained.

While you measure, price the work in rough terms: hours multiplied by a loaded hourly cost, plus the cost of the errors you found. The real cost of repetitive work goes into that arithmetic; a rough figure agreed with the finance lead is enough to decide whether a candidate is worth a project at all.

What does this look like on a real shortlist?

Notice that the loudest complaint, the enquiries, is not the best first candidate, and the quietest task, the monthly report, may be the easiest win. Notice too that two candidates split into an automated part and a human part. That split is usually where the safe value sits.

How do you turn a shortlist into a decision?

Before committing to anything, test the cheaper answer. A clearer form, a corrected master list, a shared inbox rule or one setting in software you already pay for will sometimes remove most of the work. Automation should be the option that wins on merit, not the option nobody questioned.

Then decide the pace deliberately. The Economic Survey 2025–26 chapter on India’s AI ecosystem suggests sequencing over speed, and classifying uses as deploy now, pilot or defer according to readiness across data, skills and legal frameworks. That is a sound habit for a single business too: one candidate to build, one to pilot in a limited way, and the rest deferred with a written reason and a date to revisit.

For the final comparison between the two or three that survive, choosing the first workflow with a clear head weighs impact, repetition, readiness and risk together, and ends with a one-page case. The solutions overview shows the shapes these systems usually take, and BYBO’s Blueprint is the paid version of this exercise, with the workflow mapped, the baseline costed and a recommendation on what to automate or leave alone.