Which processes in your company to automate with AI first

5 min read·Updated October 5, 2026

Why many AI projects start badly

Many companies start with the tool: “we have to use AI”. They buy a license, run a test and, months later, nobody uses it. The problem is almost never the technology. It’s that nobody chose a specific process, with a clear owner and a way to tell whether it improved.

Other companies do the opposite: they go straight for the most important and most complex process, like automating all customer service or the whole accounting department. It’s a big project with many edge cases, and if it fails, the team loses trust in AI.

The better strategy is more modest: start with a small, measurable, low-risk process, learn, and then grow.

Four signs of a good first process

A good candidate meets almost all of these conditions:

• It’s repetitive: it happens many times a week, always in a similar way.
• It has clear rules: a new person could learn it from a short guide.
• A mistake isn’t serious: if the AI gets it wrong, someone reviews it before it reaches a customer or money.
• It takes up the team’s time: someone spends hours today on something that doesn’t need their best judgment.

If the process depends on one person’s intuition, isn’t documented or every case is different, it isn’t a good candidate yet.

If you can’t explain the process on one page, it isn’t ready to be automated. Documenting it is already progress, with or without AI.

Good and bad candidates to start with

How to choose your first case, step by step

1. 01

   List repetitive tasks

   Ask each department to name the three tasks that take up most of their time. Don’t filter yet.

2. 02

   Score each task

   For each one, answer three questions: how much time does it take? How clear are its rules? What happens if there’s a mistake?

3. 03

   Pick just one

   The one that takes a fair amount of time, has clear rules and low risk. Resist the urge to start three at once.

4. 04

   Name an owner

   One person responsible for the result, who really knows the process.

5. 05

   Keep a person in the loop

   At first, the AI proposes and a person reviews and approves. Once results are stable, you can reduce the review.

How to know whether it worked

Before you start, measure how the process works today. Without that baseline you can’t prove any improvement. Simple metrics are enough:

• Average time to complete the task.
• Number of tasks done per week.
• Share of errors or corrections needed.
• Satisfaction of the team that uses it.

Run a short trial, for example two to four weeks, and compare. If time goes down and errors don’t go up, you have a real case to expand. If it doesn’t improve, you learned something cheaply and can try another process.

Watch out: “it saves time” isn’t enough if people spend the same time fixing what the AI did. Measure the total time, review included.

Common mistakes to avoid

• Starting with the hardest case: a small win builds trust; a big failure destroys it.
• Automating a bad process: if the process is confusing, AI just makes it confusing faster. Fix it first.
• Ignoring sensitive data: before pasting customer information into a tool, check what the provider does with it.
• Not training the team: if people feel AI is coming to replace them, they won’t cooperate. Explain that the goal is to take tedious tasks off their plate.
• Having no plan for errors: decide who fixes them and how people are told when the AI gets something wrong.

When the process needs a custom tool

Often the first process can be solved with tools that already exist. But if your workflow is unusual, for example if it lives today in spreadsheets and scattered emails, it may be worth building a simple app that connects AI to your data and keeps a person approving. It’s the kind of problem where a small MVP lets you validate the idea before investing more.