Strip away the branding and the automation does something specific: it tests thousands of combinations of headlines, images, audiences and bids, then shifts money toward the combinations producing whatever result it was told to chase. At that job it is genuinely better than any human, because no human can run ten thousand small experiments a day.
The catch sits in the phrase whatever result it was told to chase. The system pursues the goal in its settings with complete literalness. It has no idea whether that goal makes the business money.
The AI does not know that a bakery loses money on small custom orders. It does not know a plumber's vans stop at the county line, that Tuesday appointments go unfilled, or that one product carries triple the margin of another. It knows exactly what the settings and the conversion tracking tell it, and nothing else. Every default setting left untouched is a question answered without the owner realizing they were asked.
A plumber's account shows the shape of the problem. The campaign goal was contact-form submissions, so the AI hunted for cheap form fills and found them: people outside the service area, price shoppers, even job applicants. The cost per lead looked wonderful. The cost per booked job was dreadful. Once the account tracked phone calls and booked appointments instead, the same automation started finding a different, better kind of customer. The machine did not improve. The instructions did.
Location. Target people physically in your service area, not people interested in it, and draw that area honestly.
Conversions. Track the action that means money, a phone call, a purchase, a booked appointment, not visits to a thank-you page.
Budget. Daily budgets are allowed to overspend on busy days, so judge the monthly total, and set it at a figure you could lose without pain while the system learns.
Placements. Decide where the ads may appear. A local service business rarely needs its budget spread across mobile games and video partner sites.
Recommendations. Switch off automatically applied recommendations. Review the platform's suggestions once a month instead of letting them install themselves.
Every ad platform produces recommendations, and most point the same direction: broader targeting, bigger budgets, more automation. Some are genuinely useful. All of them come from the seller. Treat them the way you would treat a waiter recommending the most expensive dish: possibly sincere, worth hearing, and no substitute for looking at your own numbers first. The question to ask of any recommendation is simple. What evidence in my own results says this will help?
Read the search terms report and block anything irrelevant.
Compare the platform's reported conversions with your till, calendar or booking system.
Check where the spend went geographically.
Note your best and worst ad, and ask what separates them.
Make one change, write down why, and leave the rest alone for a month.
Owners who keep this routine end up knowing more about their account than many people who manage ads for a living, because they check every claim against a business they understand completely.
A sensible tipping point: when monthly ad spend passes the level where a modest efficiency gain would cover a professional's fee, or when the ads become a channel the business genuinely depends on. At that point it is worth seeing how a specialist adwords management agency runs small accounts, with tracked calls, weekly search-term reviews and budget decisions explained in plain numbers, if only to know what good management looks like before hiring anyone.
AI advertising tools make excellent employees and terrible owners. They work fast, never sleep and follow instructions with perfect literalness. Give them clear instructions, who a good customer is, what a sale is worth, where the business actually operates, and they will earn their keep. Skip the instructions and they will still spend the money. That part is fully automated.