AI vs automation: how to choose for everyday work
Guy PavlovPublished
- You can write the rule down: when this happens, do that
- Use ordinary automation. It follows your rule every time, so you always know the result.
- Someone has to make a judgement, such as sorting a customer's question or reading an invoice
- Use AI for that step, and have a person check its result.
- People do the task a different way each time, or you don't do it often enough
- Use neither yet. Agree on one way of doing it first, or keep doing it by hand.
Most small and mid-sized firms have a few tasks that someone does every week: copying enquiries from the website form into a spreadsheet, sending reminders for unpaid invoices, or answering the same customer questions by email. You may already pay for an AI assistant and wonder whether it should take one of them over. In a Statistics Canada (opens in a new tab) survey, 19.2% of businesses said they used AI to produce goods or deliver services over the previous 12 months, up from 12.2% a year before.
To choose, ask whether you can say in advance what should happen each time. If you can write the rule down as “when this happens, do that”, ordinary automation does the job, and you'll always know what the result will be. That's software that follows your rule, such as a Zapier or n8n workflow or a spreadsheet formula. If the right move depends on someone's judgement, use AI for that step and have a person check the result. Many tasks have steps of both kinds, and one workflow can run both.
Ordinary automation when you can write the rule down
Picture the invoices you send each month. The rule for chasing them is easy to write down: seven days after the due date, email a polite reminder; at fourteen days, send a firmer one; when the payment arrives, stop. Your accounting software or a workflow can follow that rule for every invoice. Once you've tested it, nobody has to remember who owes what or check each reminder before it goes out.
When a workflow like this breaks, the cause is usually something you can see, such as an expired login to your accounting software or a client's email address that bounces. You fix that one thing, and the reminders go out again.
AI for the step that needs judgement
The replies to those reminders are different every time. One client says they paid by cheque last week, another disputes a line on the invoice, and a third asks to pay in two parts. Someone has to read each reply and decide what happens next, and that judgement is the step to give to AI.
An AI tool can read each reply, sort it and draft an answer. It can also get a name or a figure wrong, so a person in your firm checks what it produces before it goes anywhere. If client details would go into the tool, read my six checks before an AI tool touches client data first.
Three projects with steps of both kinds
| Project | Steps you can write a rule for | Steps that need reading or judgement | What changed |
|---|---|---|---|
| Invoice processing for an in-house finance team | Sending each matched invoice on as a record for review, and each unmatched one to accounts payable | Reading each incoming invoice and pulling out its fields | The team checks finished records and resolves exceptions, instead of keying in and matching each invoice by hand |
| SEO audit automation | Running nine tools on the site and placing each screenshot in the agency's slide template | Drafting the findings for each audit screen, which a person reviews before the deck goes out | Each audit took 4 to 7 hours of hand work per site and now takes under an hour of review |
| A multilingual customer service agent | Answering predictable questions through set paths, with the same answer every time | Answering anything outside those paths from a search of the reference material, in the visitor's language | One assistant serves visitors in more than one language |
When to use neither
Hold off on both kinds of automation while any of these is true:
- People do the task a different way each time. Agree on one way and write it down first, or you'll be keeping up a set of rules for each version.
- You don't do the task often enough. Building and testing a workflow can take longer than the hours it would save.
- You can get rid of the task by changing how the work is done. If clients keep sending invoices back because they're missing a purchase order number, add a box for it to your invoice template, and nobody has to fix and resend them.
Still unsure about a task? Go through these four questions with whoever does it every week, and you'll know whether it's worth setting up.
Getting help with the choice
The right choice saves money twice: you don't pay for AI where a rule would do, and you don't build what a tool you already have can handle. In an AI Adoption Review, I map up to three of your tasks, check the tools you already pay for, and give you a written recommendation for each one. Each recommendation comes with a business case, so you know what the change is worth before you spend on it.
To see which area to work on first, take the free AI readiness self-check.
