The short answer
Automate the task that scores highest on four things: how often it happens, how cheap a mistake is, whether the data is already accessible, and whether two people would agree when it is done. Inbox triage, lead follow-up, and report assembly usually win. Skip anything scoring lowest on any single measure.
Why do first projects fail?
Most first automation projects fail because of what was chosen, not how it was built. A company picks the most painful problem in the building, usually something messy, high stakes, and understood by exactly one person. That's a hard place to learn, and when it stalls, leadership decides AI doesn't work here.
A first project has two jobs. It has to return real hours, and it has to teach your team how this kind of work goes. So you want something a bit boring, where a mistake is cheap to catch. Save the ambitious one for the third project, when you know what you're doing.
That can feel underwhelming when competitors are talking about agents that run the whole company. Ignore it. A dull system that runs every day builds trust with your staff, and you'll need that trust for the harder projects later.
How does the scoring work?
List every repetitive task you can think of. Ask the people doing the work instead of their managers, because the real list lives with them. Then score each task from 1 to 5 on the four measures below, where 5 is always good for automation. Add the scores for a total out of 20.
The scale is a thinking tool, so don't argue over a 3 versus a 4. Have two people score separately and compare. The disagreements are the useful part, because they show you where nobody really knows how the work gets done.
- Volume: 5 means many times a day, 1 means a few times a month
- Cost of a mistake: 5 means cheap and easy to catch, 1 means lost money, a legal problem, or a lost customer
- Data availability: 5 means it already sits in a system you can reach, 1 means it lives in someone's head
- Definition of done: 5 means two people would always agree, 1 means it depends who you ask
What does each score really measure?
Volume tells you whether the payoff is real. A task that takes four hours once a quarter feels awful and isn't worth automating. A two-minute task done all day long is where the hours are. Count before you guess, because people tend to remember the painful task, not the frequent one.
Cost of a mistake tells you how much supervision you'll need. Sorting a shared inbox wrong means someone forwards an email. Sending a customer the wrong price means a phone call and an apology. Anything that spends money, messages a customer, or changes a record of truth can still be automated, but it needs a person in the loop at first, and that lowers the return.
The last two measures are the ones people skip. Data availability exposes the task that looks easy until you learn the information sits in a vendor system with no export. Definition of done exposes the task nobody can actually describe. If you can't write down what correct looks like, you can't test it, and if you can't test it, you can't trust it.
What is the rule for a low score?
The total matters less than the floor. Anything with a 1 in any column waits, whatever its total. A task with huge volume and a catastrophic cost of failure isn't a first project. It's a later one, with controls around it. A task with no reachable data isn't an automation project yet. It's a data project, and that's worth knowing before you spend money.
Low scores also tell you what to fix. A 1 on definition of done usually means you need to write down the process, and that exercise alone can save hours whether or not you automate. A 1 on data availability points you toward getting your records into a place you control.
Which tasks usually win?
Certain tasks keep coming out on top because they have high volume and cheap mistakes. Inbox triage is the classic. One of ours sorts a shared mailbox into eight categories and routes each message to the right person, and a wrong guess costs somebody a few seconds.
Lead follow-up scores well too, since the first useful reply often decides who wins, and a person can review the replies at the start. Then there are the numbers you assemble by hand every Monday, which are predictable and easy to check against last week.
- Sorting and routing a shared inbox
- First replies to new leads, with a person approving at first
- Weekly and monthly report assembly
- Reading invoices and forms into records, with low-confidence items held for review
- Cleaning up duplicates and blank fields in a CRM
What should you do after you pick one?
Before you build, record the current numbers: how many per week, how long each takes, how often it goes wrong. Without a baseline you can't show the project worked, and you'll end up arguing from anecdotes. Then write one sentence that defines done, and collect a set of real examples for testing, including the ugly ones.
Check whether a tool you already pay for can do it. If it's simple, a Zapier or Make flow may be all you need. A custom single workflow usually takes two to four weeks from signed scope to live, and most of that goes to connecting systems and testing on real data. Run it with a person watching, measure it against your baseline, and let those results choose the next project.
Common questions
What are the best business processes to automate with AI?
Repetitive tasks that involve reading or sorting text, with cheap mistakes and accessible data. Inbox routing, lead response, document extraction, and report assembly are common starting points.
Should we automate the task our team hates most?
Only if it scores well. A hated task that happens rarely, or that nobody can define, makes a poor first project. Run it through the same four measures as everything else.
How many tasks should we automate at once?
One. You want a measured result before you start the second. A single workflow in production will teach you more than three stalled pilots, and it will show you what your data really looks like.