Visible is not the same as frequent

The tasks people notice are usually the unpleasant ones: the awkward export form, the invoice rebuilt by hand every time, the arrangement that takes three emails to settle. They stick in the memory because they get in the way. Whether they actually take up much time is a separate question.

Frequent tasks, by contrast, barely register. An appointment is confirmed, a record is copied from one system to the next, a notification goes to the right person. Each of these takes a minute. Because it happens twenty times a day, it adds up — and because it is so small, nobody questions it.

Before a team talks about tools, a simple inventory is worth more: which tasks repeat, how often, and who is involved. The result almost always looks different from the list of things that annoy people most.

How to tell what is worth automating

A handful of criteria show whether automation is worthwhile. None of them is a tool.

Frequency. The more often a task comes up, the stronger the case for handing it over. A task that takes ten minutes once a month is never worth it. Setting it up, testing it and maintaining it later costs more than it will ever save.

Consistency. Automation works where the input takes the same shape every time: the same fields, the same order, the same exceptions. The more variants a task has, the more rules the automation needs — and the more often it meets a case nobody anticipated.

Number of people involved. Tasks that pass back and forth between several people or systems lose time and information at every handover. This is often where automation does more good than with work one person handles alone.

Cost of a mistake. Some mistakes are a nuisance, others are expensive. Where someone retypes data by hand and a transposed digit produces a wrong invoice, reliable transfer is worth more than the time saved. The reverse holds too: where an error in the automation could do serious damage, it needs checks that eat into its benefit.

If a task scores clearly on two or more of these criteria, it is a good candidate. If only one applies, it usually is not.

What is not worth automating

The list of things you deliberately leave alone matters as much as the list of candidates.

Decisions a person has to stand behind. Whether a quote goes out, whether a complaint is justified, whether an order takes priority — decisions like these can be prepared, but not handed over. Automation can gather the information the decision needs. The decision itself stays with the team.

Workflows that keep changing. Anything done differently every few weeks cannot be automated reliably. Every change to the workflow becomes a change to the automation, and at some point maintaining it takes more effort than the manual work it was meant to replace. Here it helps more to stabilise the workflow first.

Work that only exists because of a poorly designed process. Many tedious tasks exist because a process was set up badly: data is entered twice because two systems are not connected; approvals pass through three people because nobody settled who is responsible. Automating that work makes the mistake faster, not smaller. No tool fixes it — a change to the process does.

Where automation typically fails

Most automation does not fail at launch. It runs for a while, and then one of four patterns emerges.

Nobody notices when it stops. Automation that works quietly also fails quietly. A changed password, a new version of an interface, a renamed field — and data stops arriving. Often this only comes to light when a customer asks. Every automation therefore needs monitoring that reports when something fails to happen, not only when something goes wrong.

The data arrives in a form nobody can work with. The transfer technically works, but at the other end the information sits in the wrong field, without context, or in a format someone has to clean up by hand. The work has not gone away; it has only moved.

The one person who understood the system has left. In small teams, automation often grows on the side, built by whoever knows the tools best. As long as that person is around, everything runs. When they leave or their role changes, what remains is a system nobody wants to touch, because nobody knows what depends on it.

The process was automated instead of simplified. A cumbersome workflow does not get better by running automatically. It only gets harder to change. Before automating a task, it is worth asking which of its steps are needed at all. Two or three often drop out — and what remains is simpler to automate and simpler to maintain.

What needs to be settled first

Four points should be settled before the first automation goes live. They sound organisational, and they are. The technology is rarely the hard part.

Who is responsible when it breaks. Not in general, but by name: one person who gets notified, knows the system and has time to deal with it.

Where errors go. Automation produces error messages. What matters is whether they land in an inbox nobody reads or somewhere a person sees them and acts.

What happens to the data. Which data flows where, where it is stored, who has access and how long it stays there. As soon as external services are involved, this belongs before launch, not after.

Who may trigger what. Some automations start on their own, others at the press of a button. In both cases it must be clear who can start, stop and change them — and what happens if someone triggers one twice by accident.

As long as these four points are open, a team is not ready for automation yet. That is not an obstacle; it is the actual first step.

How to start sensibly

The best starting point is a single task, not a programme. One that scores clearly on the criteria above, involves few people and would not do much damage if it failed.

For that task, the team first writes down how it is done by hand today — step by step, including every exception that comes up in practice. This description is often more revealing than any tool, because it shows which steps only exist out of habit.

Then the workflow is simplified, and only after that automated. For an initial period, the manual and the automated version run side by side, so it becomes clear whether the result is right. Before that, the team should agree on how to tell that the automation works: not by the fact that it runs, but by the right thing arriving in the right place at the end.

Only once this one task is stable and someone looks after it does the next one follow. That feels slow. It does, however, prevent a small team from running a tangle of automations after six months that nobody can see through any more.

The real work starts after launch

Setting up an automation is the smaller part. Running it reliably for years is the real work: interfaces change, workflows shift, teams grow. Building with that in mind from the start means building less, but more robustly.

How we look after existing systems over time is described under AI Operations Care.