How to Identify Repetitive Tasks to Automate in One Week

How to Identify Repetitive Tasks to Automate in One Week

To identify repetitive tasks worth automating, log everything you and your team do on a recurring basis for one week. Note how often each task appears, how long it takes and where the information comes from. At the end, cross frequency with effort, drop anything that requires delicate judgment, and pick a single candidate with clear rules. The goal is not to list everything that could be automated. It is to find one small, measured, safe task to test.

Key takeaways

• Memory exaggerates what annoys you and forgets what is small and constant. That is why a log kept during the week beats a brainstorm meeting.
• Frequency and effort per occurrence, taken together, show where time is really lost. A rare, long task is seldom the best first candidate.
• A good candidate has predictable input, an explicit rule, verifiable output and low risk if it goes wrong.
• Every estimated gain needs a discount: human review, exceptions and maintenance also cost time.
• Choose one task, define how you will know it worked, and agree in advance on what stopping means.

Why look at the real week instead of memory

When you ask someone what they do that is repetitive, the answer is usually the most irritating task, not the most expensive one. The monthly report that everyone dreads makes the list, while the dozens of small daily checks go unnoticed because they have become habit.

A short log, kept at the moment the task happens, corrects that bias. It also moves the conversation out of opinion: instead of arguing about whether something is very repetitive, the team looks at a list with counts. For a business owner or operations leader, that list is the basis for an investment decision, even a small one.

One caution from the start: the log is for understanding the work, not for watching people. Explain this to the team, record tasks rather than individuals, and let each person note their own work. If this point concerns you, read about measuring AI productivity without turning it into surveillance.

Day 1: set up a simple log

You do not need new software. A shared spreadsheet, a notebook or even a phone note will do, as long as everyone uses the same fields. The fewer the fields, the more people fill them in. We suggest these:

• A short task name, starting with a verb (for example: copy order to spreadsheet).
• Trigger: what started the task (a message, an email, a time of day, a customer request).
• Where the input information lives (WhatsApp, email, PDF, management system, spreadsheet).
• Where the result ends up (another system, a reply to the customer, a report).
• Time spent, in approximate minutes, with no stopwatch.
• Did you have to decide something, or follow a rule? Mark one of the two.
• Did something go wrong or fall outside the usual pattern? Note it in one line.

The decision field is the most important and the most ignored. It separates the mechanical task, which follows rules, from the task that looks mechanical but depends on context.

Days 2 to 4: log without changing how you work

The most common mistake is using the week to tidy up the work and, without noticing, changing what you are trying to observe. Work as you always do. The note should be made right after the task, in a few seconds. If logging requires a pause, people give up by Wednesday.

Some practical guidance for this period:

• Also note small interruptions, such as answering the same customer question that arrives in different forms.
• Include behind-the-scenes tasks: checking that something was done, chasing a colleague, looking for a file.
• Do not filter by gut feeling. If a task seems too small, log it anyway.
• If someone forgets to log, do not reconstruct from memory at the end of the day. A day with a gap is better than an invented day.

People who work heavily through WhatsApp often find that much of the volume sits in messages: confirmations, delivery-time questions, document sending. That is useful, but it comes with a privacy warning. When logging, record the type of data involved (name, phone number, document, financial data) without copying the content of conversations. Later this weighs on the decision about what can be automated and within which limits. A company policy for employee ChatGPT use is a good starting point for defining what can and cannot leave the company environment.

Day 5: add up, compare and correct the exaggeration

On the last day, consolidate. Group tasks that are the same thing under different names. For each group, calculate the weekly effort in two parts: how many times it happened and how long it takes each time. The calculation is frequency times time per occurrence, and it only serves as an order of magnitude.

Then apply discounts, the step almost everyone skips. The time automation saves is not the full time of the task. Subtract:

• The time for human review, which should exist at least at the start.
• Exceptions: cases that fall outside the rule and stay manual.
• Maintenance: rules change, systems change, document formats change.
• The time to set up and follow the pilot.

A task that looked large can shrink considerably after these discounts, and a small task that is very frequent and fully predictable can come out ahead. If the gain only holds up under optimistic estimates, treat that as a sign the candidate is not yet mature.

Illustrative scenario: three tasks, one choice

Imagine a building materials store with a small team. At the end of the week, the log shows three groups of tasks.

• Copying orders received on WhatsApp into the picking spreadsheet: it happens several times a day, takes little time each time, follows a clear rule and the output is easy to check.
• Answering delivery-time questions: also frequent, but it depends on checking stock and the day's route, and sometimes requires negotiating with the customer.
• Building the weekly report for management: it takes a lot of time, but happens once a week and mixes data from three sources with manual adjustments.

The report stands out because it is time-consuming, but it is rare, and automating it would mean handling judgment-based adjustments. Delivery-time questions are good for a later stage, with controlled access to stock. The safest candidate is the first: frequent, regular and verifiable. If the agent gets one order item wrong, the error shows up in the check before picking. This is an invented example to illustrate the reasoning, not a client result.

How to choose the candidate: a decision guide

After adding up and applying discounts, run each task through these questions. The more yes answers, the better the candidate.

• Does the input always arrive in the same format, or in a few known formats?
• Can you write the rule in a few sentences, including what to do when information is missing?
• Can someone check the result quickly, without redoing the work?
• Would a mistake be cheap and reversible?
• Does the task involve little personal data, or can the data be limited to the minimum necessary?
• Is there someone on the team who agrees to follow the pilot?

And these questions serve to eliminate:

• Does the decision directly affect a person's rights, such as credit, hiring or cancellation? Think about human review first and read about LGPD automated decision-making and the right to review.
• Does the task depend on relationship, tone or negotiation? It is probably better kept with people. See what not to delegate to AI in business.
• Is the process itself poorly defined? Automating a mess only speeds up the mess. Fix the process first.

Notice that the question is never whether AI can do it. It is whether the company can explain, limit and check it. Many good candidates do not even need generative AI: a deterministic flow (fixed rules, integration between systems) solves them at lower cost and with more predictability. AI comes in where there is free text, irregular documents or classification that fixed rules do not cover. Anthropic itself, in its Building effective agents guide, distinguishes predictable workflows from agents with autonomy, and recommends starting with the simplest solution that works.

Define the metric before you start

Once you have chosen the candidate, write down three things on one page before building anything:

• The baseline: how the task is done today, with the time and error rate the log showed, approximately.
• The success metric: for example, time per occurrence, number of corrections needed, or cases that required human intervention.
• The stop criterion: what would make the team adjust or abandon the pilot.

Without a baseline, there is no way to know whether things improved. Without a stop criterion, pilots become endless projects. And without an agreed metric, each person judges by how it felt, which almost always favors whoever wanted it to work.

How the map, limit, pilot and hand-off method fits in

The week's log is, in practice, the mapping step: understanding the current process and having a baseline. Next comes limiting: deciding which data the automation can see, which actions it can take and where a human approves. Then you pilot a real flow, with logs and the agreed metric. Finally, you hand off: deliver the documented flow to the team and decide, based on pilot data, whether to expand, adjust or stop. If the data includes personal data sent to providers abroad, also check international data transfer under LGPD when using AI providers and the guidance from Brazil's data protection authority, the ANPD.

Signs you are overestimating the gain

Before approving the pilot, look for these warnings:

• The calculation uses the full time of the task and ignores review and exceptions.
• The task is done by different people in different ways, and nobody has agreed on the rule.
• The gain depends on volume growing in the future.
• The savings are spread across small minutes from many people, without anyone actually gaining useful time.
• The only evidence is the feeling that the task is boring.

If two or more appear, go back to the log and refine it. It may be better to repeat the collection for another week, at a different point in the month, to check for seasonality.

Frequently asked questions

How long does it take to identify repetitive tasks? One week of logging is usually enough for small teams, plus one day to consolidate and decide. If the work varies a lot over the month, as with financial closings, repeat the log in another period before concluding.

Do I need a specific tool to log the tasks? No. A shared spreadsheet with a few fields is enough and has the advantage of being simple to fill in. Automatic monitoring tools can help later, but they start with a higher privacy cost and more resistance from the team.

Should every repetitive task become an AI automation? No. Many are solved with fixed rules, integration between systems or even by simplifying the process itself. AI makes sense when there is free-form language, irregular documents or classification that simple rules do not cover, and always with limits and checking.

What if the task involves customers' personal data? That does not rule out automation, but it calls for care: reduce the data to what is necessary, define who has access, record what the system did and confirm the legal basis with legal support. If in doubt, start with a task that uses little personal data and leave the more sensitive ones for later.

Sources

• Anthropic: Building effective agents
• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework