How I decide whether a process is worth automating with AI

These days, almost any process can be “hooked up to AI.” The question is whether any savings are left once you subtract the time spent on review and corrections. I describe four questions and one formula I use before we build anything.

Who it’s forFor business owners, managers and specialists without a technical background, as well as developers who get requests along the lines of “let’s do this with AI.”

“Can we automate this with AI?” The answer is almost always “yes.” So I ask a different question: once it’s automated, will we spend less time on it than we do now? That’s where the answer can be surprising.

Four questions to start with

Is the process written down?

If nobody can describe, step by step, how the task is done today, automation won’t bring any order to it. First we write the process down, then we automate it.

How often does it repeat?

A task done once a month rarely justifies building automation. A task done dozens of times a day, even a short one, usually does.

What does a mistake cost?

A wrong category on an email in the inbox is a minor thing. A wrong amount on an invoice or a bad answer to a customer means actual losses. The more expensive the mistake, the more time you have to spend on checking.

Is checking faster than doing?

This is the question that decides it most often. If reading and fixing the output takes almost as long as doing the task from scratch, automation makes no sense.

A simple calculation

Before we build anything, I calculate the savings in hours per month. The formula is deliberately simple:

type Process = {
  perMonth: number; // how many times a month
  minutesManual: number; // how long it takes manually today
  minutesCheck: number; // how long it takes to check the AI output
  errorRate: number; // share of outputs that need fixing (0–1)
};

const hoursSaved = (p: Process) =>
  (p.perMonth * (p.minutesManual - p.minutesCheck - p.errorRate * p.minutesManual)) / 60;

// Example: categorizing customer requests
hoursSaved({ perMonth: 600, minutesManual: 3, minutesCheck: 0.5, errorRate: 0.1 }); // ≈ 22 h

The numbers in the example are illustrative, but the structure of the calculation is not. It shows right away that review time and the correction rate can eat up all of the savings.

Where I actually start

Not with AI. First, I check whether part of the work can be automated with plain rules: email filters, an integration between two systems, a simple script. I bring in a language model only where you need judgment on unstructured text: a message, a ticket description, a document.

Good first use cases are usually:

categorizing and routing messages to the right person
summarizing long threads before a meeting
extracting data from documents into a form that a person only has to approve

Decision: I automate with AI only where checking the output is at least several times faster than doing the task.
Why: Only then do the savings survive contact with the actual error rate.
Costs: Some interesting ideas get ruled out from the start. Some processes have to be cleaned up first, which takes time in itself.
Revisit when: models make noticeably fewer mistakes on this type of task, or the cost of a mistake changes.

The bottom line

Good AI automation starts with a sheet of paper, not with choosing a model. Write the process down, measure how long it takes, and estimate how long checking the output will take. If the numbers don’t add up, no model will fix that.