Most business owners we talk to have tried AI once. Somebody opened ChatGPT, asked it a question, got a decent answer, and went back to work. Nothing about the week changed.
That isn't a failure of the tool. A chat window is a blank page, and a blank page won't tell you what to do with it. So here's how to find the jobs worth handing over, and what it looks like when one of them sticks.
Start with your week, not with the tool
The common mistake is starting from the tool. You read that AI can summarize a meeting, so you go looking for a meeting to summarize. That buys you a party trick and a subscription.
Start from your own week instead. Think about what you or somebody on your team did more than twice, and what they'd describe as typing something up, or checking something, or moving it from one place to another. That's the list.
Then narrow it. A job is worth handing over when all four of these are true.
- It happens often. Once a quarter isn't worth the setup. Twice a day is.
- It's boring to whoever does it. If they'd be relieved to lose it, that's a good sign.
- Somebody can check the answer in seconds. A draft you skim beats an answer you have to verify from scratch.
- Getting it wrong is cheap. Start where a mistake costs a minute, not a customer.
That last one matters more than people expect. AI is confident when it's wrong, so the first job you give it should be one where a mistake is obvious and easy to fix.
The four jobs AI is good at
Whatever you point it at, it's doing one of these four.
- Reading. A document, a photo of a form, a supplier's PDF. The information already exists and it's stuck in a shape software couldn't use before.
- Sorting. Taking what arrives in no order and putting it where it goes. Enquiries, applications, tickets, email.
- Drafting. A first version of something a person finishes. A reply, a summary, a description, the boring half of a quote.
- Judging. Looking at something and saying which bucket it belongs in, and why.
If the job you're thinking of isn't one of those four, AI probably isn't the answer. Ordinary software often is, and it's cheaper and more predictable when it fits. The difference between the two is worth knowing before you spend anything.
What it looks like in a real week
- Paper that arrives and has to be typed in. Work orders, delivery slips, forms filled in by hand. A photo goes in, the fields come out, somebody checks them.
- Enquiries that need sorting before anyone answers. What it's about, how urgent it is, who should take it.
- Notes nobody has time to write up. A recording of a site visit becomes a summary and a list of what to do next.
- Long documents somebody has to search. A contract, a manual, a policy. You ask a question and it points at the paragraph.
None of them are impressive on their own. Each one gives somebody two or three hours a week back, and those add up.
Where a chat window stops being enough
All of that works in a chat window while you're the one doing it. You paste something in, you read the answer, you use it.
It stops working when it has to happen every time, without you. You'll know the moment: somebody is pasting the same prompt forty times a week, or the job only gets done when one particular person remembers to do it.
That isn't an AI problem. It's the same problem a spreadsheet has when four people need it at once. The work is real and it repeats, so it belongs in software, with AI doing the one part that needs reading or judging and normal code doing the rest.
That's what custom software is. Most of it is ordinary, and the AI is one component inside it rather than the point of it.
What it costs
Two different numbers, and mixing them up is how budgets go wrong.
A subscription. A business seat on ChatGPT or Claude runs $20 to $25 US per person per month, and both want a minimum of two seats. For a team of five that's about a phone bill. What you're buying is a better chat window. Nothing about how your business runs changes.
Building it into the work. This is where the hours come back, and it's a project rather than a subscription. Builds here start at $5,000, quoted in full before anything is written. Automation work on its own is often less. Larger or more custom projects are quoted separately.
Start with the subscription. Give it a month. If one job turns out to be worth doing every single day, that's the one worth building properly.
How to tell whether it's working
Measure the time before you start.
Write down how long a job takes today and how often it happens. A month later, take it again. If nobody can tell you what changed, nothing did, and you're paying for a habit.
The number worth watching isn't how clever the output looks. It's whether the person who used to do that job got their hours back, and what they did with them.