Where AI automation is worth it and where it is not: a four-question filter

Most failed automation projects did not fail technically. They failed because the wrong process was chosen. The filter below takes ten minutes and tells you, broadly, whether it makes sense.

Question 1: does it repeat often?

Automation takes a fixed amount of work to build. To be worth it, the process has to repeat enough. A process that happens three times a month rarely justifies the effort, however annoying it is.

Rough threshold: above 20-30 repetitions a month it starts to be worth it. Below 10, almost never.

Question 2: can the rules be stated in words?

If a new colleague could learn the process from a list of rules, it can be automated. If the answer is “it depends, you get a feel for it”, then the process contains judgement written down nowhere — and that is where automation produces plausible mistakes, which are the most dangerous kind.

It does not mean it is impossible. It means the first stage is writing the rules down, not building anything.

Question 3: what happens if it gets it wrong?

Split processes into three:

• The mistake is caught immediately and fixed cheaply — excellent candidate. E.g. a draft reply a human reads before it goes out.
• The mistake is caught late but can be repaired — automate, with checks.
• The mistake goes straight to a client or an authority — automate only with mandatory human approval.

The practical rule: the bigger the consequence, the more human in the loop. That is not a limitation of the technology, it is risk management.

Question 4: does the data exist in usable form?

An agent that must answer questions about orders needs access to orders. If the information lives in someone’s head, in three different spreadsheets and in a notebook, your first problem is not the AI.

What works well, in practice

• Triaging and classifying incoming messages — easy to verify, high volume.
• Extracting data from documents with a consistent format: invoices, delivery notes, forms.
• Draft replies to frequent questions, reviewed by a human.
• Recurring reports assembled from several sources.
• Filling documents from templates, with automatic checks.

What works badly

• High-stakes decisions without review — risk assessments, approvals, penalties.
• Processes whose rules change weekly; you end up maintaining the automation more than you saved.
• Talking to upset customers. An unhappy client wants a person, and a perfectly polite automated reply makes it worse.
• Anything relying on data you do not have.

What a healthy start looks like

One flow, the most repetitive and the lowest stakes. Measure first: how many cases a month, how long one takes. Build a prototype on that flow. Compare after a month. Only then discuss the second.

It is boring and unspectacular. But it is the difference between automation that stays in service and automation that gets talked about once, at launch.

← Back to the blog