Your opinion on AI is probably out of date
Posted on 29 July 2026
I hear it regularly. AI is overhyped, it disappoints. And I get it, because a year ago I thought the same. On paper it always looks fantastic, in influencer videos and in company demos, but the moment you try to build something real with it, it falls short. Almost everyone recognises that feeling. Except AI moves in six months what other technology takes five years to do. If your picture of it is six months old, you're describing a different product.
Five things that changed enough to deserve an update.
1. Agents do real work now
The biggest misunderstanding is that AI is an answering machine. You ask a question, you get an answer, done.
Under the bonnet sits a language model, an LLM: trained on an enormous pile of text and therefore good at predicting what logically follows from whatever you type in. That principle hasn't changed.
But a bare language model isn't what you're using anymore. Claude Code and Codex aren't models, they're agents built around one. They reach your files, they run commands and they keep going until the job is done, without you at the controls. That's where the leap is, not in the model itself.
I have one running the email marketing for House of Tones and one keeping my content pipeline going, from topic to scheduled post. Those are working hours that used to sit with an employee or an agency.
If you have work that looks the same every week, this is the moment to look at what can come off your plate.
2. AI images aren't bad anymore
The era of AI slop is over. AI slop being those slick fake images you see through in half a second: hands with six fingers, warped text, faces that are just slightly wrong.
I run my images through Higgsfield now, for video, ads, presentations and product mockups. Nano Banana Pro is my default, and when the image needs legible text on it I use ChatGPT's image model. Six months ago I opened Canva or Figma for this. Now hardly ever.
For marketing that's money straight away. You test five variants in an afternoon instead of three weeks later after a design round.
3. Voice mode replaced typing
This sounds like a detail. It isn't.
I'm sitting at my desk like always, I just barely type anymore. My girlfriend hears me talking to my computer all day and occasionally bursts out laughing when it gets a bit heated.
There's a simple reason underneath: you speak roughly three times faster than you type. But the real gain is elsewhere. I get rough ideas out of my head out loud, exactly as messy as they are in there, and then AI helps me turn them into something concrete and structured.
The tool that does this for me is Wispr Flow. You speak, and your words land as clean text in whatever window you happen to be working in. Mail, chat, an editor, it doesn't matter which app. It strips out your ums and half-finished sentences by itself. It runs on Mac, Windows and your phone, and it handles Dutch without any fuss. Of everything I've tried over the past six months, this is the one I'd least want to give up.
If you still only use AI through a keyboard, you're leaving a lot on the table.
4. The desktop app, not the browser
Plenty of people I speak to still work in the browser, including people who have been at this a while. Fine for one-off questions. It just keeps you small.
The desktop apps reach the files on your computer. They see what you're working on without you spelling it out every time. That sounds like a technical difference, but it changes what you can hand off. Try it for a week.
5. Prompt engineering matters less
Two years ago you had to tell a model step by step how to approach something. And often you gave it a role on top of that: you are the best marketer in the country, you are an award-winning brand copy strategist. A whole subculture was built on that, with tricks, formulas and templates.
You don't need to hand out that role anymore. Current models are sharp enough to get what you mean in one go.
What does count is that you know what your goal is. The more concretely you work that out, the better what comes back. "Write an email" works worse than "I want a hesitant customer to finish reading and know exactly which single step to take". The route there, the model works out by itself.
What this means
People who say AI disappoints are often right about the version they tried. That version just doesn't exist anymore. What didn't come out of it then, does now.
If your picture of it is six months old, you aren't hopelessly behind. You're simply due an update.
I build this kind of automation for small and mid-sized businesses in West Friesland. If you want to know what's there to gain in yours, send me a message.