The usual order is backwards.
Somebody reads about a tool, subscribes, uses it enthusiastically for three weeks, and quietly stops. Nothing broke. It was simply never attached to work that recurs, so there was never a moment when not using it caused a problem.
Start from the work instead, and the tool becomes obvious.
How do I find the right job to automate?
Look for four properties together. Any one of them alone is not enough.
It repeats. Weekly at minimum, ideally daily. A quarterly task takes longer to automate than to keep doing.
It follows the same steps every time. If it requires a judgement call halfway through, automation will make it worse rather than faster.
Being slightly wrong is survivable. Because it will occasionally be slightly wrong, and you need that to be an inconvenience rather than a lost client.
Somebody currently hates it. This matters more than it sounds. Automating a job nobody minds saves time. Automating a job somebody resents saves time and returns attention to work that needs a human.
Write down everything that takes more than fifteen minutes and happens every week. The ones ticking all four are your list, in order.
What usually tops that list?
The same handful, across nearly every business I have looked at.
Moving information between systems. Enquiry arrives in one place, gets typed into another, gets typed into a third. Every retype is time plus a chance of error, and this is the single most common thing worth fixing.
Acknowledgements and confirmations. The message saying we received this and somebody will call. Universally valuable, almost never done, and an afternoon of work.
Follow up reminders. Not automated messages to the customer, a reminder to a person that this needs chasing on Thursday. Most lost enquiries are not rejections, they are people who were busy that week.
Recurring reports. The Monday numbers that somebody assembles by hand from three places. This is often several hours a month for one person.
Getting files into the right folder with the right name. Unglamorous, and it compounds.
What should I leave alone?
Anything where the human contact is the product.
The first call after an enquiry. Anything to do with a complaint. The conversation that closes a sale. Automating these does save time, and it costs you the thing customers were actually buying.
I would also leave alone anything that only you know how to do. Automating a process you have not written down produces a system nobody understands, and when it breaks, and it will, nobody can fix it.
Write the process down first. Half the time you improve it enough that the automation becomes unnecessary.
Does this need AI, or just software?
Mostly just software, and this is worth being clear about because the word is attached to everything at the moment.
Moving data between systems, sending confirmations, scheduling reminders, assembling reports: none of that needs a language model. It needs an integration, and those have existed for years, work reliably, and are cheap.
Where AI genuinely adds something is where the input is messy or written in prose. Summarising long documents. Sorting incoming messages by what they are about. Drafting a reply for a person to check. Pulling structured information out of an unstructured email.
Use the simplest thing that works. A tool that does one deterministic job will outlast a clever one that occasionally invents an answer.
How do I know if it worked?
Measure the task before you change it.
How long does it take now, how often does it happen, how often does it go wrong. Five minutes of counting, and without it you will never know whether the automation helped, because the before state exists only as an impression.
Then check again after a month. The usual finding is that it saved less time than expected and removed more irritation than expected, and the second is often worth more.
What is the most common way this fails?
Building something nobody owns.
An automation runs quietly until it stops running quietly. If no person is responsible for noticing, it can be broken for weeks before anyone realises the confirmations stopped going out.
Every automation needs a named owner and something that tells them when it fails. That is a five minute setup and it is the difference between a system and a liability.
What would I do this month?
Pick one thing. The retyping, probably, because almost every business has it.
Fix that one, properly, with an owner and a check. Then look at the list again in a month.
Businesses that automate one thing a quarter end up transformed in two years. Businesses that try to automate everything in one month end up back where they started with a subscription.
We start from the job that eats somebody’s Tuesday rather than from a tool. Tell us what you are working on.
