The first process to automate should be a repetitive, frequent task with clear rules, with data already available in digital form and low risk if a mistake happens. Almost every company has a few activities that everyone says “should be automated”. Ideas are not in short supply; choosing is the hard part. A well-chosen first case finishes on time, and the team gains confidence for the next one.
Below is a simple method you can apply without technical knowledge. Its aim is not to find the “perfect” process, but a first case small enough to see through and clear enough to tell whether it was worth it.
In short, in seven steps
1. Take an inventory of repetitive tasks. 2. Score them on frequency, volume, clarity of rules, data availability and risk. 3. Narrow the winner down to a small case. 4. Decide where AI helps and where simple rules are enough. 5. Decide where a person steps in. 6. Measure before and after. 7. Avoid the common mistakes.
1. Take an inventory of repetitive tasks
Start with the people who do the work, not with the org chart. Ask each department to note down, for a week or two, the tasks they repeat that don't require a substantive decision. You don't need a complicated format: a shared spreadsheet is enough.
For each task, ask for a few basic details. Without them, the discussion stays at the level of impressions.
- What actually happens, step by step, and what triggers the task (an email, an order, a deadline).
- How often it occurs and roughly how long one run takes, as estimated by the person who does it.
- Which applications are involved: email, Excel, ERP, CRM, the invoicing platform, the online store.
- Where the data comes from and where it ends up.
- What happens when something goes wrong: who notices, what the correction costs, who is affected.
You will usually end up with a long, mixed list: entering orders, consolidating reports, sending reminders, sorting emails, checking documents received from suppliers. That's normal. The next step puts it in order.
2. Score each candidate on five prioritisation criteria
Give each task a score from 1 to 5 on the criteria below. Don't aim for mathematical precision; what matters is that the same people score all the tasks, using the same logic.
- Frequency: how often the task occurs. Something daily has more potential than something monthly.
- Volume: how many cases, or how much cumulative time, it takes up in a given period.
- Clarity of rules: can you explain to a new colleague, in writing, how it is done? The fewer “it depends” exceptions, the better.
- Data availability: does the information needed already exist in digital form and can a system access it? Poorly scanned paper or data that lives in one person's head lowers the score.
- Risk: what happens if the automation gets it wrong. Here the score is reversed: low risk means a high score.
Add up the scores and sort the list. The top three to five tasks are your real candidates. High-risk tasks, such as payments, decisions about customers or communications with legal effect, can be automated later, but they are rarely a suitable first step.
Practical tip
If two tasks score about the same, choose the one the team wants automated. A process people find annoying is more likely to be adopted than one chosen purely from the spreadsheet.
3. Choose a small, measurable first case
Once you have chosen the process, narrow it down. “We're automating invoicing” is too broad. “We automatically capture the data from invoices received from our main suppliers and prepare it for review in accounting” is a case with a clear start, end and result.
A good first case has a single main input, a result that is easy to check and an owner on the company's side. If describing it takes several paragraphs and lots of “but what if”s, it is still too big.
4. AI automation or simple, rule-based automation?
Not every automation needs artificial intelligence. Many processes can be handled with clear rules: when a new order comes in, a record is created in another system and a notification is sent. These are classic automations, predictable and easy to maintain.
AI has its place where information is unstructured: freely written emails, documents in different formats, requests that need to be classified by content. Here a language model can extract data or suggest a classification, while the rest of the flow stays rule-based.
- Without AI: copying data between applications, deadline reminders, generating reports from the same sources, approvals at fixed thresholds.
- With AI, as a step in the flow: reading documents in varied formats, classifying incoming messages, summarising long texts, suggesting a reply that a person checks.
- Combined: the AI extracts and suggests, the rules validate the format, a person confirms the unclear cases.
The rule of thumb: use the simplest solution that solves the problem. AI adds flexibility, but also uncertainty, so it is worth introducing only where it genuinely adds something.
5. Decide from the start where a person steps in
Good automation doesn't replace people; it takes the mechanical work off their hands and leaves them the decisions. That's why, before implementation, you should decide clearly what a person checks and when.
For example, the system can prepare an invoice's data, but a colleague in accounting approves it before it is recorded. Or an email can be classified automatically, but low-confidence cases go to a separate list for review. As you see how the system behaves, you can adjust the thresholds without giving up control.
Important limitation
Every automation will meet cases nobody foresaw. What matters is what it does then: a well-designed system stops and asks for help rather than guessing. Make sure there is a clear path for exceptions and a person who receives them.
6. Measure before and after automation
Without a baseline measurement, you won't be able to say whether the automation helped. You don't need a complicated system; a few weeks of consistently recorded data are enough for an honest comparison.
- Time spent on the task, per case or cumulatively per week.
- The number of cases processed in a given period.
- Errors found and the time needed to correct them.
- Time from trigger to completion, important where a customer or colleague is waiting.
- The share of cases that go to human review and the reasons they end up there.
- Feedback from the people who work with the process: what is easier, what has become harder.
Use the same measurements before and after, over a comparable period. Results can vary greatly from one company to another, so don't rely on general figures found online, but on your own data.
7. Common mistakes in a first automation project
- Starting with the most complicated process so the investment is “worth it”, and never reaching the end.
- Automating a process that doesn't work manually either. First clarify the steps, then automate them.
- Not involving the people who do the work. They know the exceptions no outsider can see.
- Choosing AI for a task that a simple rule would have handled.
- Not deciding who maintains the automation after launch, when an application or a form changes.
- Not measuring anything beforehand, so that after a few months nobody knows whether it helped.
How we work on the first automated process
At WEB-AZ we usually start with a short analysis of your processes, carried out together with the people who work with them every day. We choose a single case, agree what is measured and where human validation stays, then implement it and monitor it in operation. Only once the first case runs reliably do we discuss the next one.
Next steps