What to automate with AI first.

What to automate with AI first is the task you could write down on one page: it comes round again and again, its inputs and rules are clear, you can check the result at a glance, and a mistake is cheap to catch. Keep a person on anything that needs judgement or a conversation with a client. That is the test we run before we build anything.

What should you automate with AI first?

Run every candidate task through four questions. A task that passes all four is a good first automation. A task that fails one usually needs work before it is ready. A task that fails two is usually not worth automating yet.

Does it repeat? Automation pays back every time the task runs. Something done constantly earns its cost. Something done twice a year rarely does, however tedious it feels.

Can you write it down? If you can list the inputs, the rules and what finished looks like, a system can follow them. If the honest answer is "it depends who's asking", the rule lives in someone's head, and it has to come out of there first.

Can you check the output quickly? Somebody will look at what the system produces, especially early on. If checking takes as long as doing the work, nothing has been saved.

Is a mistake cheap to catch? A wrong draft that a person reads before it goes out is cheap. A wrong payment, or a wrong message sent straight to a customer, is not. Start where errors are caught before they cost anything.

Why does the tidy version of a task matter so much?

Mercor published a study this month that makes the point better than I can. On four simplified month-end accounting tasks, a frontier AI model scored 100%. Twelve licensed accountants averaged around 37% on the same tasks.

Read one paragraph further and the picture changes. On Mercor's full benchmark of 160 accounting tasks, the same model scores 54.5%, and 58% of the tasks are not fully solved by any model. The authors say the simplified version "stripped out the job": no colleagues, no supervisor, no client changing their mind. (The study, with its caveats)

That gap is the whole lesson. AI is very good at the tidy version of a job, where everything it needs has been written down. It is much weaker at the messy version. So the first job in any automation project is not building. It is writing the task down until it becomes the tidy version.

We call that the Study step. We learn where the hours go, what repeats and what a person still has to decide. Sometimes it ends with a plan. Sometimes it ends with us telling the client a task is not ready to automate, and why.

What should stay with a person?

The pipeline we built for Chrvoje Engineering, an engineering education channel, shows where the line sits. The client types one video idea. The system writes the script, plans every shot, generates the shots and the voice, makes the thumbnail and publishes, along with the blog post and the article.

The client approves three things: the script, the shot plan and the thumbnail.

Those three are not random. The script decides what the film says. The shot plan decides what it shows. The thumbnail decides whether anyone clicks. Each is a judgement about the channel and its audience, and each is cheap to check before anything expensive happens. Everything between those points is repeat work, and that is where the system runs on its own.

How did this play out on real projects?

On Chrvoje, the repeat work was making the film, and the decisions were few and clear. That made it a good first automation. The channel is now past two million organic views across formats, with no paid promotion.

The second is a custom AI-assisted ERP we built for an Australian business. The client is confidential, so I will say only this: the AI sits in the steps that repeat, not across the whole system.

The same test picked the work both times. Repeat work, written down, quick to check, cheap to fix.

When is automating the wrong call?

Three kinds of task fail the test more often than people expect.

Tasks that change every time. If every instance is different, there is nothing to write down, and a system has nothing to follow.

Tasks where the rule is "it depends". When the answer depends on who is asking, the knowledge is in a person's head. Get it out and onto paper first, or leave the task with that person.

Tasks done rarely. The build costs the same whether a task runs often or hardly at all. Rare tasks almost never pay it back.

In each case "not yet" is a fine answer. Writing a task down often fixes the process before any AI is involved, and that alone is worth doing.

Questions people ask before their first automation

Do I need an in-house AI team? No. The way we build it, the automation sits inside the tools you already use, and you get the source and the documentation at handover.

What does it cost? It depends on the task, which is why we agree a fixed price against a written scope before building anything. The Study step is how that scope gets written.

Can it run without anyone checking? Not the way we build it. The system does the repeat work, and a person signs off wherever a mistake would cost something.

Which business processes should I automate first? The ones that pass all four questions. In most businesses that means quotes, reports, orders, data moved between tools by hand, or content made to the same pattern again and again.

What I would tell someone commissioning their first automation

Write the task down before you ask for a quote. Inputs, rules, and what finished looks like. A quote against a vague task is a guess, and you pay for the guess later.

Name who approves what. Decide which points a person signs off before the build starts. Adding approvals afterwards means rebuilding the parts around them.

Start with one task, not a platform. One task that works teaches you more than a system that half-does ten. The second automation is much easier once the first one is running.

Ask where it stops and asks you. Every good automation has points where it waits for a person. If the answer is "it never does", be careful with what you hand it.

Get the source and the documentation. You should be able to change, move or hand over what you paid for. If you can't, you have rented it, not bought it.