When you should not automate a process

Better not to automate a process if the task is rare, constantly has exceptions, requires judgement, or if a program alone would decide about people. In those cases set-up and upkeep cost more time than the workflow saves, or you lose control of the result. Five check questions settle this before you start, and a simple calculation shows whether it pays off.

What a workflow is (a trigger starts a fixed sequence of steps in your programs) is explained in the article on the basics. This one deals with the other side.

What the literature says and what it does not

First, the state of the evidence, because it is thin. I do not quote a failure rate for automation projects, since I found no reliable one. The literature review by Wewerka and Reichert evaluates 63 publications on so-called robotic process automation (RPA), meaning software that imitates how people operate programs. The authors note that the literature covers only successful projects and that failed ones remain to be studied. Syed and colleagues write in another review that sound theoretical foundations are lacking for judging its application and development objectively. I have read only the abstract of that work.

What is documented, by contrast, is which features the evaluated studies name for suitable tasks: many repetitions, fixed rules, a stable environment and little exception handling. The simpler the task, the better suited it is. A consultancy (noventum) describes it similarly: tasks with many exceptions, frequent changes or inconsistent data sources cause problems. That is one provider’s opinion, not a measurement.

One caveat: these studies concern RPA. Tools such as n8n usually connect programs through interfaces. That the criteria apply equally there is the author’s assumption, which I find plausible but cannot prove.

Rare tasks: calculate first

A workflow has to be built, tested and maintained. If the task is rare, it never earns that back. The following example uses invented data and author’s assumptions, not measured values. Suppose set-up takes 12 hours in both cases, upkeep 2 hours a year, and the workflow removes 40 percent of the working time.

For the rare task, upkeep alone costs more time than the workflow saves, and the set-up never comes back. A good checklist is enough there. For the frequent task, the workflow pays off within the first year, as long as the assumptions hold. Change them to your own values, and be honest about upkeep, because programs and access rights change. One of the studies evaluated by Wewerka and Reichert describes upkeep as potentially costly and tedious, in contrast to the cheap set-up.

Many exceptions

The workflow has to either know each exception or hand it to a person. When the exception becomes the rule, for instance because customers send free-form enquiries by email and almost every one looks different, most of the work lands by hand with your office staff, and the workflow becomes a detour. Count beforehand how many of the last 20 cases follow the pattern. How many exceptions are still tolerable I cannot say. The calculation above answers that for your figures.

Judgement and discretion

Goodwill gestures, discounts, the order of emergency call-outs: here a person decides, with knowledge of the customer and the situation. In the literature, two of the evaluated studies criticise that RPA cannot make decisions. A workflow can prepare such tasks by assembling documents, but it should not decide.

There is a second problem. In 1983 Lisanne Bainbridge described the “ironies of automation” using industrial plants as her example: people are left with what cannot be automated, and skills that are no longer practised decline. According to the studies she cites, a person can keep up attention to things where almost nothing happens for only about half an hour. Her paper is about plant operators, not offices. Applying it here is my assumption: an approval where nothing ever stands out will eventually be waved through. So plan spot checks and do not let those responsible unlearn the work.

The middle way: prepare instead of complete

Between “everything by hand” and “everything automatic” lies a part that often pays off. The workflow collects, sorts and provides a draft, and a person decides and approves. That leaves the judgement with you, and the calculation still comes out better than with purely manual work. An invented example: when an enquiry arrives through the form, the workflow files the data in the customer record and creates a draft reply. Whether you accept, and at what price, your office decides. The saving is then smaller than the 40 percent in the worked example, because the decision remains manual work. Use a lower figure in your calculation.

Operation also belongs in the calculation. Who maintains the workflow if the person who built it is absent? If there is no answer, the task is not yet ready for automation, even if it meets every other criterion.

Legally sensitive tasks

Article 22 of the General Data Protection Regulation gives data subjects the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning them or similarly significantly affects them. There are exceptions: the decision is necessary for a contract, a legal provision allows it, or the person gives explicit consent. For contract and consent, the controller must at least allow a person to intervene, the data subject to express their point of view and to contest the decision. A workflow that sorts out job applications or customer enquiries on its own therefore belongs first with the data protection officer or a lawyer. The same goes for deadlines, tax matters and contracts.

The wording above is the version in the Official Journal of 2016, and as far as I know it is still unchanged as of November 2026. The European Commission has proposed amending Article 22 as part of a “Digital Omnibus”, and the German Bundesrat commented on it on 27 March 2026. I found no adopted text, so check the current position before you decide. Your data protection officer, tax adviser or lawyer clarifies the individual case, not this article.

Five check questions before you start

The graphic shows the questions in the order in which you ask them. A No does not always mean “do not automate”. Often it is enough to simplify the task or leave the decision to a person.

The last question is about responsibility: a workflow that silently stops is worse than none, so it must be settled before the start who notices a failure and fixes it.

What you can do now

1. Write down five recurring tasks in your office, with cases per month and minutes per case.
2. Calculate each task as in the table: manual work per year, estimated saving, set-up, upkeep.
3. For the last 20 cases, count how many deviated from the pattern, and mark every one where someone had to use judgement or decide about a person.
4. Before any workflow that touches personal data or makes decisions, ask your data protection officer.