What changed in AI tooling over the last six months

It is hard to talk about what changed in AI without drowning in the hype, so let us be disciplined about it. The question we care about is narrow: of everything that happened in the last six months, what actually changed how we build for small and mid sized businesses. Most of it did not. A few things did.

The most useful shift is that the better models got noticeably more honest about uncertainty. A year ago, getting a model to reliably say "I am not sure about this one" took real effort and a lot of careful wrapping. It is now far more workable. That sounds modest. It is not. For the workflows we build, knowing when the system is guessing is more valuable than squeezing out another point of accuracy, because it lets us route the doubtful cases to a person instead of letting them slip through. Honesty about uncertainty is what makes automation safe to leave running.

The second shift is cost and speed. The price of running these models through a workflow has kept falling while quality has held or improved. Six months ago we sometimes had to talk a client out of an idea purely because the running cost did not justify the benefit. That conversation happens less often now. Things that were borderline are now comfortably worth doing, which quietly widens what we can build for a sensible budget.

What did not change

The third is that handling documents and messy structured inputs got better. Models deal with real world documents, the slightly crooked scan, the inconsistent layout, the form someone filled in their own way, more reliably than they did. That matters enormously for back office operations, where the input is almost never tidy.

Now the honest counterweight. The fundamentals did not change at all. Clean data still decides whether any of this works. A model pointed at a mess still produces a confident mess. You still need a real process around the model, with clear inputs, sensible handling of the uncertain cases, and a person in the loop where the stakes are high. Nothing in the last six months removed the need to think carefully about the workflow. If anything, cheaper and faster tools make it tempting to skip that thinking, which is precisely when projects go wrong.

So our advice to clients has barely moved, even as the tooling under our feet has improved. Start with one workflow. Get the data right. Build something you can switch off. Measure whether it actually works. The tools have made all of that a little easier and a little cheaper, and that is genuinely good news. But the discipline that makes AI useful in a real business is the same as it was. The improvements are real. They are just improvements to the engine, not to the need for someone who knows how to drive.