What not to automate: the jobs AI should skip

Not every task worth doing is worth automating. That sounds obvious until you're sitting in a meeting and someone says, 'AI could do that.' Then you nod along, and six months later you're paying for a tool that saves 3 hours a month.

The question isn't whether AI can do something. It's whether automating it makes your business run better. Those are different things.

The jobs that look good but aren't

Start with frequency. If something happens once a month or less, don't automate it yet. A dental clinic in Kelowna might use a tool to scan patient consent forms and file them into the right folders. But if they get 40 forms a month and it takes 6 minutes per form, that's 4 hours a month. The setup alone—connecting the tool to their filing system, testing it, training someone to watch it—takes 20 to 40 hours. You break even in five to ten months. That's not terrible. But if they get 4 forms a month, you never break even. Do it by hand.

Next, look at judgment calls. A property manager in Surrey might think about automating tenant screening. An AI tool could flag applications that don't meet basic criteria—income, credit score, references. That's useful. But the decision to approve or reject a tenant? That stays with a person. It involves reading between the lines, understanding local context, and knowing what the lease actually requires. An AI can narrow the pile. It shouldn't make the call.

Then there's relationship work. A plumber in Richmond has customers who call him back because they trust him. If he automates his follow-up texts to sound like a chatbot, he loses that. If he automates them to sound like him, he's just speeding up something he was already doing—and he still has to write the first message. The gain is small and the risk is real.

Finally, skip automating anything where a mistake is expensive or hard to catch. An accounting firm in Burnaby handles client tax filings. They could use AI to draft sections of a return. But a filing error costs the client money and the firm its reputation. The person still has to read every word. The AI saves maybe 20 percent of the time. Is that worth the setup cost and the risk that someone gets sloppy because the tool is doing most of it. Often not.

What automation actually costs

Here's what people miss: the bill for automation isn't just the software.

It's the setup. Someone has to connect your AI tool to your existing systems—your invoicing software, your customer database, your email. That's 10 to 40 hours depending on how messy your systems are.

It's the testing. You run the tool on real data. It makes mistakes. You adjust it. You test again. That's 5 to 20 hours.

It's the person who watches. A 12-person software company in Gastown might automate their lead intake form so that new prospects are automatically added to their CRM and sent a welcome email. But someone still has to check that the tool is working. Is it catching all the leads. Is it filing them correctly. Is it sending emails to spam. That's 2 to 5 hours a week, forever.

It's the maintenance. Your tools break. Your systems change. Your AI tool needs updates. That's another 2 to 10 hours a month.

Add it up. A tool that saves 5 hours a week sounds great. But if it costs 60 hours to set up, 15 hours to test, and requires 3 hours a week of watching, you're at 150 hours of work before you see a net gain. At a loaded cost of $50 to $100 per hour, that's $7,500 to $15,000 in real money.

That doesn't mean don't do it. It means know the number before you start.

The jobs worth automating

The best candidates are repetitive, high-volume, low-stakes, and predictable.

Repetitive means it happens the same way every time. A distributor in Calgary receives orders from customers. Each order needs to be checked against inventory, priced, and sent to the warehouse. That's repetitive.

High-volume means it happens often enough that the setup cost pays back. If the distributor gets 200 orders a day and each one takes 8 minutes to process, that's 27 hours a week. Automating it makes sense. If they get 3 orders a day, it doesn't.

Low-stakes means a mistake is annoying, not catastrophic. If the tool misfiled an order and someone catches it before it ships, no real harm. If the tool approved a $50,000 credit decision, that's different.

Predictable means the rules don't change. If your business follows the same process every time, automation works. If you're always making exceptions, automation breaks.

And crucially: a person still checks the output. The distributor's tool processes orders automatically, but a human reviews the batch before it goes to the warehouse. That's the safety net. That's what keeps you from shipping the wrong thing to the wrong place.

Where to start

• Write down three tasks your team does every week that feel repetitive and take more than 2 hours combined. For each one, count how many times it happens and how long it takes. Be honest about the numbers.

• Pick the one task where the rules are clearest and a mistake would be least expensive. Ignore the one that seems most painful. The easiest win matters more than the biggest problem.

• Talk to the person who does that task. Ask them what could go wrong, what exceptions they handle, and what they'd actually want the tool to do. Write down what they say. That's your spec.