If you had told me five years ago that an AI would quietly fix a piece of a client's software overnight, test it, and leave a short note explaining what it changed and why, I would have smiled politely and gone back to work. Today that is a Tuesday. The last twelve months have been the most exciting stretch of my career, and I want to explain why, without the hype and without pretending the hard parts do not exist.

The pace is the story

People keep asking me what changed. My honest answer is: the calendar. Not one model, not one product launch, but the speed at which the thing you saw in a flashy demo last spring becomes the thing running your business by autumn.

One number captures it. There is a well-known test where an AI is handed real, messy bugs from real software projects and asked to fix them properly. A year ago the best systems solved about 60 percent. Stanford's 2026 AI Index reports they are now close to 100 percent. Read that again. One year. The test is imperfect (some of it is too easy now, and some AI systems have effectively seen the answers), so "near 100 percent" does not mean "solved." But the direction is unmistakable, and the slope keeps getting steeper.

The race at the top is a tie, which is good for you

I used to keep a running list of every new AI model with its release date. I deleted it, because it goes stale faster than milk. Here is all you need to know: the top four AI labs are now practically neck and neck. On the most-watched public ranking, all four sat within a hair of each other in March 2026, and the gap between the best American and best Chinese systems has shrunk to under 3 percent.

When the leaders are that close, raw intelligence stops being what you pay for. Price, reliability and fit for the job take over. That is great news for anyone buying software: we pick the right model for each task instead of betting your company on one vendor. And for clients who need their data to stay inside their own walls, there are now free, downloadable models good enough to run in-house that still feel modern.

The story I actually want to tell: we closed the loop

Here is the part I care about most, and it is a real story, not a slide.

One of our clients is Benson AI. We did not just build features for them. We set up their entire software project so that a person who has never written a line of code could add to it safely. Think of it as building a kitchen where the knives are sharp but the layout makes it very hard to cut yourself: the rules of the house are written down where the AI can read them, every change is automatically tested, nothing goes live if a test fails, and a human still signs off where it matters. Boring, careful plumbing. The boring plumbing is the whole game.

The result still surprises me when I say it out loud. The founders at Benson, who are not engineers, now build features for their own product using their own Claude. They describe what they want in plain language, the AI builds it inside the rails we set, and the safety checks catch what they are there to catch.

Then we went one step further. When a user of the app reported a bug or asked for a small change, our "autopilot" ran the whole loop on its own: it read the request, made the change, tested it, and shipped it to users. No developer in the middle. We closed the loop.

I want to be careful here, because this is exactly the kind of claim that gets oversold. We did this only for small things: a typo in the app, a form that should reject bad input, a button in the wrong place. The kind of change that is annoying to queue up and easy to verify. We did not point autopilot at big features, and we would not. Big changes still need a human who owns the outcome. But for the long tail of small stuff, the path from "a user complained" to "it is fixed and live" now runs by itself, and it holds because the rails hold.

Why it works, and where it breaks

Google Cloud's 2025 DORA report, based on nearly 5,000 technology professionals, said it better than I can: "AI doesn't fix a team; it amplifies what's already there." Ninety percent of those people already use AI at work. Well-run teams get gains that compound. Chaotic teams get faster chaos. The Benson loop did not work because the AI is magic. It worked because everything around it was solid.

That is also the uncomfortable subtext of the big surveys. McKinsey's latest State of AI report found that 88 percent of organizations now use AI somewhere in the business, close to saturation. But fewer than four in ten say it has moved their bottom line, and only about 6 percent say it moves it meaningfully. I read that gap as opportunity, not disappointment. The tools work. The bottleneck is how companies are organized. The ones that redesign how they work now will pull ahead of the ones still running pilots.

I will not lie: something got lost

Here is the part I did not expect to write. This shift took some of the pure happiness out of the job for me. I got into software because I love solving problems, the satisfying kind where you stare at a wall for two hours and then the wall moves. Now, in most cases, the AI moves the wall before I have finished my coffee, and it does it better than I would have. There is a small grief in that, and I think anyone who pretends otherwise is either not a builder or not being honest.

What I have found on the other side of it is a different kind of joy. Less of the puzzle, more of the outcome. I get to spend my day on the problems that actually needed me, and I get to watch people who could never build before build something real. That is not the happiness I signed up for fifteen years ago. It might be a better one.

The skills that are quietly becoming everything

If AI does more of the typing, what is left for people gets more human, not less. Describing clearly what you want is now a superpower, because a vague request produces confident nonsense and a precise one produces something you can use tomorrow. Judgment about what to build, and what to refuse to build, matters more when building is cheap. Communication matters more, because half of my week is turning "make it nicer" into an actual definition of done. And taste, the ability to look at what the AI produced and know whether it is good or merely finished, is the difference between a product that lasts and a product that becomes a headache in six months.

A small example. Last month the hardest thing I did was not technical. It was sitting with a client and rewriting one paragraph of a feature request until it was impossible to misunderstand. The AI did the rest in an afternoon. The paragraph was the work.

And for developers specifically, one word matters more than any technology: responsibility. Ownership. Taking a thing from a fuzzy idea all the way to "it is live, it works, and I will answer for it." Writing code is getting cheaper by the month. Owning the result is not. The developer who quietly takes things end to end and gets them done has the biggest edge going, and it is an edge no AI is coming for.

Where this leaves us

I am not going to pretend AI never produces confident nonsense or that every pilot project escapes the lab. It does, and some do not. But those are engineering problems now, not mysteries, and they are getting solved faster than anyone predicted.

So let me be openly optimistic about where this goes. Within a few years I expect the gap between having an idea and having a working product to shrink to days, then hours. A dentist in Celje will describe the booking system she always wanted and have it running by the weekend. A two-person startup will ship what used to take a team of twenty. Software will stop being a scarce, expensive thing that only some companies can afford and become something closer to electricity: everywhere, cheap, and mostly invisible. The boring work of keeping systems alive will fade into the background, and the interesting work, deciding what should exist and for whom, will be what people are paid for.

I think that is a genuinely better world, and not just for developers. Most of the good ideas on this planet die because nobody nearby could build them. That excuse is going away.

At Indigo Labs we have always believed the tool matters less than the craft around it. That belief has never paid off more than it does right now. Build carefully, measure honestly, own what you ship, and this is genuinely the most fun software has been in three decades.

Until next time, keep shipping.