Keyboard
I spend more than $600 a month on AI subscriptions. Claude, ChatGPT, Gemini, and a few others I'd rather not list in a newspaper. I use them for client work, research, writing and building systems. Every day. For hours.
Last month, for the first time, I used one of them to plan a family vacation.
We took our three kids to the Florida Keys, and I wanted to see what locals do, not what the brochures sell. So I asked. The answers came back fast, organized and confident. State parks with mile markers. Beaches with no admission fee. Restaurants described as local favorites and hidden gems. It looked like a plan. It could have been anybody's plan.
That's a vacation story, but it belongs in the business section. Travel planning should be AI's easy mode. There are millions of reviews and menus online, and the stakes are a sandwich. If there's any task where the first answer ought to be good enough, it's this one.
Now think about your business. Nobody has posted reviews of your pricing, your best customer's quirks or why your last hire didn't work out. The context that makes an answer right for you lives in your head and your inbox, not on the open web. Your business is harder for AI to get right than any vacation. If it misses on a trip, with the whole internet as its study guide, it will miss by more on the work only you understand.
When the answers disappoint, most people decide AI is overhyped or that they need a better model. That second idea is the conventional wisdom, and the industry is happy to sell it.
I was already paying for the top tier. It didn't help, because the first answer is almost never the answer.
Two of the restaurants it recommended were closed. Each restaurant's own home page said so. So did its Google listing. The tool had searched the web, but the fact that mattered most sat one click away, and it handed me the recommendation anyway, with total confidence.
It isn't just me. This January, a Tasmanian tour company let an outside firm fill its website with AI-written articles. One of them sent travelers hunting for hot springs that don't exist. The owner's summary to reporters was that the AI had "messed up completely." His company's name was on the page.
Merriam-Webster has a word for this. The dictionary named "slop" its 2025 word of the year, defining it as low-quality digital content, usually churned out in bulk by AI. The slop that costs you is quieter. My restaurant list had lines like "no pretense" and "worth checking current hours before you go." It read fine. It could have described any town in any week. Slop isn't always wrong. It's just not about you.
The first answer fails in four predictable ways, at the dinner table and at work.
First, it answers the average person. I have three kids, a budget and no patience for tourist traps. The first list was written for anyone who typed "Marathon, Florida." Ask for a marketing plan and you'll get one for the average business, which is to say nobody's.
Second, it sounds sure without checking. Speed came first, and verification never came at all. At work, that's the market figure nobody traced or the competitor price from last year.
Third, it won't volunteer what it left out. It never mentioned that September is the off-season in the Keys, when plenty of places close for a week or more, until I pushed back. The gaps in a first answer are silent.
Fourth, it sounds finished, so you stop checking. Researchers at Microsoft and Carnegie Mellon surveyed 319 people who use AI at work and found that the more people trusted the AI, the less they questioned it. The ones who trusted their own judgment checked more.
A team of MIT-affiliated researchers gave close to 1,900 people an image and asked them to recreate it with an AI image generator. Some got an older version, some a newer one. The newer tool won. Then the researchers fed the prompts written for the old tool into the new one. Same words, better machine. The results improved only about half as much. Half the improvement came from the people, not the machine. Through trial and error, they had learned to ask for more.
The same study found that when software automatically rewrote people's prompts, it wiped out more than half the gain. You can't hand off the adjusting. And the adjusting paid off most when people had a clear target. In your business, you're the only one who knows what the target looks like, which means you're the only one who can steer toward it. The only way to use AI right now is to iterate.
Iteration isn't fact-checking, though you'll do some of that. It's steering. Each round, you tell it something only you know, and the answer moves a little closer to you.
That's what happened on the trip. I told it to research instead of guessing. I told it we didn't want high-priced tourist traps. Round by round, the suggestions stopped looking like a travel guide and started looking like our family.
At work, skipping that step costs real money. Researchers at Stanford and BetterUp Labs surveyed about 1,150 office workers and found that 40% had received AI-generated "workslop" from a coworker in the past month. Each one took nearly two hours to untangle. Half said they now saw the sender as less capable. Skipping the iteration doesn't save the time. It moves it to whoever receives the work. Sometimes that's a coworker. Sometimes it's a customer driving to a hot spring that isn't there.
So decide what a good answer looks like before you ask. "Restaurants in Marathon" is a topic. "Cheap places locals take their kids" is a target. The same goes for email. "Write a follow-up" gets you a template. "Follow up with a client who went quiet after our proposal and cares most about response time" gets you something you might actually send. Then treat the first response as a draft. Tell it what it couldn't know. Ask what it isn't sure about. You already do this with every contractor and new hire. Do it with the chat window.
One caveat about that $600. My setup isn't typical. My tools run on the highest usage plans and connect to my email, calendar, files and the systems I've built for clients. For everyday use, a $20 plan goes a long way. If you want to build systems and do real client work, it won't take you very far. But no plan at any price turns a generic first answer into yours.
The subscription buys access. Iteration buys the answer.
The first answer is written for everyone who might have asked. The best answer is written for you, and you're the only one who can get it there.
Tyler Kelley is the Co-founder and Chief Strategist of SLAM Agency, helping organizations use AI to build visibility, strengthen relationships, and equip teams to deliver results that matter in an AI-driven future.