Agents are genuinely good at a narrow class of work and genuinely bad at another, and the line between them is predictable. Here is where I have seen them pay for themselves, and where I have watched them quietly create work.
When someone asks an assistant about your business, the answer is assembled from whatever is legible about you, which is often not your website. Here is what these systems can actually use, and what they skip.
Before a funder answers your email they search your name, and increasingly they ask an AI assistant instead. What comes back decides whether the meeting happens. Here is how to run that test on yourself and read the result honestly.
Claude Opus 5 came out this month, I had an evening free, and the game I wanted to play was one I last played on a PlayStation 1 in the early 2000s. So I described it in one prompt and asked for the whole thing in a single file. The first version ran. What follows is the prompt, what it got right, what it got badly wrong, and why the second part matters more than the first.
Klarna replaced the work of 700 support agents with AI, watched quality fall, and its chief executive now calls human customer service a VIP product, like hand-stitched clothing. Across travel, banking and consumer apps the same repricing is underway, with chatbots for the mass market and people for the paying few. While the argument about AI rages on, the market has quietly made human attention a premium tier, and the thing worth defending is ordinary access to both the tool and the person.
Google searches grew right through 2025, but by early 2026 fewer than a third sent a click to the open web, and AI answers now read websites and speak for them. A dispatch from running search campaigns for small businesses in South Africa and the Philippines: how AI answers pick their sources, what is worth changing on a page, and why local search still decides most sales.
Researchers gave 640 Kenyan small-business owners free GPT-4 over WhatsApp for five months: the ones already doing well gained 15 percent, the strugglers lost revenue, and the gap traced back to which problems they asked about. As subscriptions fall to $8 a month and school tutors cost $5 a student, the next divide is guidance, and companies, schools and individuals each have one concrete job to do.
A dated chronicle of the three and a half years from ChatGPT's launch on 30 November 2022 to 900 million weekly users in 2026: the Sydney transcript, the pause letter, Italy's ban, the five-day Altman firing, and the day DeepSeek wiped $589 billion off Nvidia. Read in order, it shows how fast amazement got a new baseline, from cheering at a drafted email to shipping App Store products with no coding background.
I spend about 90 percent of my working day inside Claude Code, an AI coding agent, and at some point it became the only colleague I talked to for days at a time. On 1 July I made my first hire, a part-time chief of staff in Makati, and part of the honest reason was the AI. The research says the average user is fine and the heavy user is not; I live in the tail, and the healthiest AI setup I have found includes paying a human.
The writers demanding publishers never use AI all have publishers. The viral posts say one prompt drinks a bottle of water; the actual paper said one bottle per 10 to 50 responses. The scientists behind the 'AI rots your brain' headlines say their study shows no such thing. Every argument against using AI is made by someone who already has the thing it replaces: an editor, a tutor, a doctor. Run the numbers on every front and the case for abstaining collapses into a luxury, and the bill lands on the people with the fewest alternatives.
Paste your bio into a chat window, ask for a post in your voice, and you get prose that could belong to anyone selling anything. The reason is not that the model is weak. It is that you gave it nothing of yourself to work from. Getting your actual voice out of an AI is closer to training a new hire than writing a clever prompt, and the part nobody wants to hear is that it costs you real hours and stays your job to the end.
People can feel AI writing within a line or two now, and the moment they clock it, they trust it less. A Nuremberg study found that just labelling an ad as AI-made made people rate it less natural, less useful, and left them less willing to buy. The fix is not to hide the AI or swear it off, but to stop letting it sound like everyone's AI: train it on your own voice, your emails, your pitches, the way you actually talk, and keep your hand on the one thing it cannot do, deciding what is worth saying.
Everyone thinks AI changed marketing by writing the copy. That is the least interesting thing it did. The real change is that being specific, a different page for every ad, a different message for every kind of buyer, went from something you rationed to something you can do by default. That quietly moved the whole job, from making things to deciding which things are worth making, and most of the industry has not caught up.
When people hear that one person runs the whole marketing side, they picture someone doing a little of everything, badly. The reality is the opposite: a set of tools wired into one machine that finds customers, builds what turns them into sales, runs the ads, and keeps track of what works, all held together by one person. Here is the machine, and why the hard part was never the tools.
Sola is a travel app built only for women who travel solo. One night last week a piece of software pretending to be one of those women opened the part of the app where they find each other, and reported, calmly, that it was confused: there was a man standing there. I had looked at that screen a hundred times and never seen him.