Conversations about automation often start with a big dream: a company where every system is connected and AI runs the work on its own. While these plans grow in meetings, the team goes on copying the same spreadsheet from one system to another by hand every morning. We suggest starting from the other end: find the single task that is the most tedious and the most often repeated, and actually automate it within two weeks.
A one-week inventory
The first step is not choosing a tool but writing down, for one week, what gets done by hand. Instead of asking the team “what should we automate?”, ask “what did you copy from one place to another today?” For every recurring task, the following goes into a single spreadsheet that everyone can access:
- Who does it: the person or role doing the task.
- What is done: which information moves from which system to which.
- How often: how many times a day, a week or a month.
- How many minutes it takes: the rough time spent each time.
- What happens if it goes wrong: does it reach the customer, does it touch money, or is it easily spotted and fixed?
Keeping the spreadsheet takes five minutes from each person at the end of the day. The aim is not a perfect time study but a rough, honest picture; what really makes the difference is not whether a task takes 4 or 6 minutes but how many times a day it is repeated.
At the end of a week, the list is usually far more concrete than expected. Typical items: for an e-commerce business, moving marketplace orders into the shipping panel and syncing stock across channels; for a B2B manufacturer, entering quote requests from the web form into the CRM and assigning them to the right salesperson; for a clinic, sending appointment reminders and calling patients who didn’t turn up; for an education provider, transferring the enrolment form to the student list and to accounting.
The first task to automate is the one everyone complains about.
Scoring the tasks
For each task in the inventory, a simple score is calculated: time × frequency × error risk. Time multiplied by the number of repeats per week gives how many minutes a week the task eats up. A three-point scale is enough for error risk: 1, the error is easily spotted and fixed; 2, the error disrupts internal processes; 3, the error reaches the customer or touches money.
For example, a 3-minute copying task done 30 times a day comes to 450 minutes a week over five working days; with an error risk of 2, its score is 900. A 60-minute report prepared once a week, with an error risk of 1, stays at 60 points. Even if the report seems more tedious, the copying task costs far more time and produces far more errors.
A high score doesn’t make every task suitable for automation. Filter the list with these questions:
- Is the data digital, or is it read off paper?
- Can the rules be written down, or does someone make a judgement each time?
- Do the systems involved have an API, or at least a regular export?
- Will the process change in the coming months? If so, the process needs fixing first.
Choosing a single pilot
Of the tasks that pass the filter, the one with the highest score becomes the pilot. The pilot’s scope is kept narrow: one trigger, one outcome. “Automating the CRM” is not a pilot; “a quote request from the web form appears in the CRM within five minutes, assigned to the right salesperson” is a pilot.
Three things are written down before you start. First, an owner on the business side: the person who will accept the result and hear first when something goes wrong. Second, the measure of success: minutes spent per week, the number of errors and the time it takes to complete the task. Third, the baseline: the current state, as measured in the inventory week. Without these three, at the end of the pilot you can’t get beyond “I think it got better”.
A two-week pilot
In the first week, the automation is built in a test environment and run side by side with the manual work. The two outputs are compared every day, and each difference is noted separately as a rule error, a data error or an inconsistency that already existed in the manual work. In the second week, the automation goes into real use and the results are compared with the figures from the inventory week.
Three safeguards are always part of the pilot:
- Human approval. In steps where a mistake is expensive, the automation prepares the work and a person presses the final button. Messages to customers, price and stock changes, payment instructions and deleting records are examples of such steps. The approval step can be removed once trust has been built; it is not removed at the start.
- Error alerts. When a run fails, a named person is told by email or message. The most dangerous automation is the one that stops silently.
- Logging. Every run records when it ran, on which record and with what result. Once a week, someone checks that the number of records processed matches the expected number. A check that stops the same record being processed twice, and a short note explaining how to do the work by hand when the automation stops, are also part of this layer.
Measurement uses the three numbers written down at the start: minutes per week, the number of errors and completion time. Next to the time saved, the time spent maintaining the automation is written down too; if a workflow has to be fixed by hand every week, the gain exists only on paper.
At the end of the two weeks, the decision is written down: continue, fix and continue, or drop it. Like the others, this decision goes on record with an owner and a review date.
When you need AI
Automation and AI are often taken to be the same thing. When the input is structured and the rules are clear, as with moving form fields into the CRM, sending an order to the shipping panel or syncing stock, a rule-based workflow is cheaper, faster and more predictable. AI is useful where the input is free text or images: pulling the product, quantity and delivery date out of a quote request in an email, sorting customer messages by topic, summarising meeting notes or writing a first draft of a product description.
Our practical rule: build the workflow with rules first, and add AI only to the one step where rules are not enough. Where possible, check the AI’s output against a fixed list: does the extracted product code exist in the catalogue, is the delivery date in the past? Where the outcome is expensive, the output still goes through a person’s approval. Counting how many outputs are approved without correction during the pilot shows whether the AI step is really reducing the workload.
When you work with customer data, KVKK (Türkiye’s Personal Data Protection Law, Law No. 6698) needs separate thought. Before sending names, phone numbers, order history or health information to an AI tool, write down which data goes where and for what purpose; where possible, remove or mask personal data. Check in the contract whether the provider uses the data you send to train its models, and in which country the data is processed; transfers abroad are subject to separate conditions. In organisations that process health data, such as clinics, this data counts as a special category of personal data and needs stricter protection. Update your privacy notice and your data processing inventory together with your legal adviser.
After the pilot
When the first pilot succeeds, go back to the list and take on the second task in the same way. Every new automation gets an owner, a log and an error alert; so a few months later you have a set of small workflows, each one measured and each one maintainable. If you would like to choose and build the first pilot together with us, take a look at our automation and AI service.