What should we automate first?
A scoring method you can run in a meeting: three questions per task, and the honest list of things that should stay human.
The most expensive mistake in business AI isn't picking bad technology — it's pointing good technology at the wrong job. The wrong job is usually the impressive one. The right job is usually the boring one nobody wanted to mention in the meeting.
Score each candidate on three questions
How often does it happen? Volume is what pays for automation. A task done two hundred times a week justifies real investment; a task done twice a quarter almost never does, however painful it is. Count honestly — the free workflow review we run exists mostly because nobody's internal estimate of volume survives contact with an actual count.
How rule-shaped is it? Could you write down how to do it, such that a capable temp could follow the instructions on their first morning? Then an agent can do it. If the honest answer is "it depends, you develop a feel for it" — that's judgement, and judgement stays with people. (Often a task splits: the reading, matching and filing is rule-shaped; the final call isn't. Automate the first part, hand the decision to a person with everything prepared.)
What does a mistake cost? Not "must be perfect" — your current process isn't perfect either, you just don't see its errors. The question is whether a mistake is cheap and reversible (a misfiled document you can re-file) or expensive and public (a wrong quote sent to a customer). Start where errors are cheap, add human checkpoints where they aren't.
Score each task low, medium or high on all three. High volume, highly rule-shaped, cheap mistakes: that's your first project. In most businesses it's something like reading incoming documents and putting the details where they belong — unglamorous, constant, and quietly eating a salary.
What not to start with
Anything customer-facing where tone carries the relationship. Anything requiring judgement nobody has written down. Anything that's broken for reasons other than effort — automating a bad process gets you the same mess, faster. And anything chosen mainly because it would look good in a board slide. The board will be considerably more impressed by "we got four hundred hours a year back" than by a chatbot.
Then make the first one count twice
The first project builds more than itself: the secure connections to your systems, the logging, the approval patterns. Done properly, the second workflow costs a fraction of the first because it inherits all of that. Which is the real reason to start small and start correctly — you're not buying one automation, you're laying the rails for the next five.
How this was made
Drafted by our research agent from the sources above, then edited, fact-checked and approved by a human before publishing. We publish this note on every article because we think you should know.
More articles
Don't marry the model
Every AI pitch leads with a brand name. But models age like phones, not buildings — and the one question that matters is what happens the day you want a different one.
What does an AI agent actually cost?
The AI is the cheap part. Here is where the money really goes, and the three numbers to ask any supplier for before you sign.
Start with the inbox
Asked what their first AI project should be, most businesses reach for something impressive. The right answer is usually sitting in the shared inbox, arriving as you read this.