When not to automate: the most dangerous words are "the pilot worked"

The pilot always works. The real question arrives three months later: who fixes it when it breaks at three in the morning, and will anyone even notice?

AA Abdelilah Arahal
5 min read Updated 20 September 2026

The most dangerous sentence in automation projects is not "the pilot failed". It is "the pilot worked".

Pilots always work. They are built in ideal conditions, on clean data you selected yourself, with the engineer who built it sitting in front of it. And a decision made on that success ignores the only question that matters: who fixes it eight months from now, when its author has left, and it breaks at three in the morning on a public holiday?

Five things not to automate

A task whose rules keep changing. Automation freezes a decision at a moment in time. If the rule changes every two months, you are building something you will rebuild forever.

A task that happens five times a year. Two hours a year does not justify a system that needs maintaining. The correct comparison weighs time saved against build time plus maintenance plus time lost when it breaks.

A task whose errors are not detectable. This is the dangerous one. An automation that fails loudly is annoying but safe. One that fails silently accumulates damage for months.

A task whose value is the relationship, not the output. A follow-up message to an important client is not text being delivered; it is a signal that you set aside time. Automating it succeeds technically and fails at its purpose.

A task you do not understand well. Automating a vague process produces vagueness faster. If you cannot explain the steps to a new colleague, you cannot explain them to a machine.

The hidden cost, in numbers

When weighing an automation, most people compare build time against time saved. That equation is incomplete.

Where the effort actually goes over an automation's life
Initial build25%
Fixing after first contact with reality20%
Maintenance across the year35%
Handling cases nobody anticipated20%
An approximate split from practical experience rather than a study. Exact proportions vary; the constant is that building is under a third of total effort.

Building is a quarter of the work. Anyone deciding on the build estimate alone is wrong by a factor of four.

The test before you start

Prompt
I am considering automating the following task: [describe it in three sentences]. Do not encourage or discourage me. Ask the questions that reveal whether this is a good idea: 1. Questions about how often it repeats and how stable its rules are. 2. Questions about what happens when it breaks, and how I would know. 3. Questions about exceptional cases I may not have considered. 4. Questions about who maintains it in six months. After I answer, give me a clear recommendation: automate, do not automate, or automate only a specific part. And justify it.

That last line matters. Most tasks are not all-or-nothing: the tedious part gets automated and the sensitive part stays human.

The third way: half automation

The choice is not between fully manual and fully automatic. In most cases the best design is the machine prepares, the human decides.

Model Who does what Suits
Fully manual The human does everything Rare and high impact
Machine suggests Prepares a draft, human edits and sends Most real cases
Machine acts with review Executes and reports, human reviews periodically Frequent and low risk
Fully automatic No human involvement Very frequent, cheap failure

The second row is the one whose value is consistently underrated. Saving eighty percent of the time while keeping the decision human gives you most of the benefit at a fraction of the risk.

In closing

Good automation is measured by what it removes from your work, not by what it adds to it. A system saving two hours a week and consuming three in maintenance is not automation. It is an expensive hobby.

Start today with an audit: list every automation running for you now, and beside each one the name of whoever would fix it if it broke today. The blanks you find are your real risk.

And if you want a clear decision framework for your team, separating what to automate from what to leave alone, that is what we build in the training programmes.

Common questions

What is the most important question before automating something?
Who fixes it when it breaks, and how will they know it broke. If neither has a specific person's name as the answer, the automation will become a liability rather than a gain.
How do I calculate whether an automation is worth it?
Compare time saved against build time plus maintenance plus time handling failures. Building is usually under a third of total effort, so anyone estimating from the build alone is wrong by a multiple.
What is the most dangerous kind of automation?
One that fails silently. An automation that stops loudly is annoying but safe; one that produces wrong results with no notification accumulates damage for months before anyone notices.
What is the alternative to full automation?
Half automation: the machine prepares a draft or recommendation and a human reviews and decides. That captures most of the time saving while keeping human judgement in the loop, and it suits most real cases.
Should I automate client communication?
Where the value is in the output, yes. Where the value is in the relationship, no. A follow-up to an important client is valuable because you gave it time, and automating it succeeds technically while failing at its purpose.
Share

No comments yet

Leave a comment

Never published. Used only if I reply to you directly.

Comments are read before they appear.

Related reading

4 min read

Will AI replace accountants?

Entries, reconciliation, transcription: all heading for automation. But signing a tax return is accountability, and accountability does not automate.

Start with a short call

Fifteen minutes to understand what your team does and what you want to change. If training is not the right answer, I will say so.