When automating a process makes sense, and when it does not yet

The question of automation now comes up in almost every company. There are plenty of tools and their capabilities keep growing, so it is easy to assume that it is enough to switch one of them on. My experience of running operations shows that success is decided by the state of the process at the moment automation begins, not by the choice of technology. Below I describe the criteria I use to judge whether a given process is ready.

When automation makes sense

The process repeats and has volume

A task performed a few times a month will rarely repay the cost of automation. A task performed dozens of times a day, by several people, is a completely different situation. Count how many times a week the process takes place and how many minutes a single run takes. If the result is small, start by simplifying the process, not with a tool.

The rules are stable and can be described

If an experienced employee can explain how they make a decision, and they do it the same way every time, the rules can be transferred to a system. When every case requires individual judgement and the answer depends on context known only to one person, automation usually creates more problems than it solves.

The input data is available and reasonably clean

An automated solution is only as good as the data it receives. If information arrives by email in any format, in spreadsheets kept in different ways, or in several systems without a common identifier, it first has to be put in order. That is often a bigger part of the work than implementing the automation itself.

You know the cost of the current way of working

Without measurement you cannot assess the return. Count people's working time, the number of errors, the cost of corrections and the delays. The same measurement will later serve to check whether the implementation delivered the expected effect.

Someone owns the process

An automated process still needs looking after: products, suppliers, regulations and document formats change. If there is no one responsible for how it runs and what it delivers, after a few months the automation starts producing errors that someone discovers only through their consequences.

When not yet

The process is not in order

Automating a mess gives you a faster mess. If the stages of the process change depending on the person, if there are informal workarounds, and tasks circulate between departments without clear responsibility, it is worth first writing the process down, removing unnecessary steps and agreeing the rules. After that kind of tidying up, some of the problems disappear without any tool at all.

There are more exceptions than rules

If most cases require manual intervention, the automation will handle a small share of the work and people will get an extra task: keeping an eye on the system. A good warning sign is the answer "it depends" to a question about the rules.

The process is about to change

Before implementing a new ERP system, changing the offer or reorganising the department, it is worth holding off on automation. An automated solution rigidly mirrors the current state and has to be rebuilt as soon as that state changes.

There is no measurement and no knowledge of the current state

If nobody knows how long the process takes and how often mistakes are made in it, there is nothing to compare against after implementation. In that case a sensible first step is a two-week measurement, and only then a decision.

How to check this in practice

When assessing a candidate for automation, I ask a few questions. How many times a month is the process performed? How long does a single run take? What percentage of cases requires a manual decision? Where does the data come from and in what form? Who is responsible for the process? What happens if the automation makes a mistake, and who will notice? The last question is often skipped, yet it decides how cautiously the implementation needs to be approached.

The answers let you put the processes in order of priority. To begin with, it is best to choose one that has high volume, clear rules and a low cost of error, for example transferring data between systems or preparing repetitive reports. Processes with a high cost of error, such as payments or decisions that affect the customer, require human control at the important stages, also after implementation.

Risks worth keeping in mind

Automation changes the way people work. Some tasks disappear, some change, and supervision of the system appears. It is worth talking to the team before implementation, explaining the purpose and asking about their experience with the process. Employees know the exceptions and workarounds better than the documentation does, and their knowledge determines the quality of the rules put into the system.

The second risk is dependence on a single tool vendor or on the one person who built the solution. The solution should have a description of how it works, access for more than one person, and a plan for what to do in the event of a failure, that is a way of carrying out the process manually.

The role of artificial intelligence

AI tools widen the range of what can be automated, for example to reading documents in different formats or initial classification of requests. The same criteria apply here too, however. A language model will not fix a process that has no owner, nor will it replace data that is missing. There is one more issue: the outputs of AI systems require quality control, and therefore rules on when a human checks the result and how accuracy is measured. Without that, it is hard to judge whether the solution is working correctly.

Where to start

Pick the one process that gets in the way most and meets the criteria above. Write down its steps, measure the time and the number of errors, simplify whatever can be simplified, and only then decide which elements to automate. This order takes a little more time at the start, but it lets you avoid costly fixes after implementation.

If you want to assess how ready your processes are, take a look at the Automation and AI page or try the maturity test.