AI Can't Do the Last 20%

Currently Available: Need a skilled Software Developer for your next project?

Hire Me Today

Categories

LLM Productivity Software Development

AI Can’t Do the Last 20%

July 27, 2026 by Vincent Schmalbach

I've built a lot of software, both my own side projects and client work as a freelancer. AI tools have made me much faster. Before AI, I spent much more time writing each feature and working through every problem myself.

Now, when I know what I want to build, I can explain it to an LLM and get a huge amount done in a short time. It can set up the project, build features, fix bugs, and help me work through problems. That can make a complete project feel much closer than it is. But an LLM has almost never taken one of my projects all the way to the point where I could call it finished.

AI can race through the first 80 percent, or what feels like 90 or 99 percent, while the last few percent still take a lot of focused work. The point is that AI gets me very far, very quickly. Then I still have to work on the details, even though I told the AI about those details upfront.

AI Makes the Easy Part Faster

At the beginning of most projects, there is plenty of work I already understand. I need to set up the application, build the main screens, connect the data, and add the normal behavior. Because I've been building software for years, I usually know what to ask for. When the output is wrong, I can spot the problem and give the LLM better instructions. That lets me turn an idea into a working demo much faster than before AI.

This was true before AI too. Once I knew what I wanted to build, I could often make quick progress on the main features. Finishing took longer. That was when I found the strange user input, the awkward integration, the deployment problem, the security issue, the background job that sometimes failed, or the design detail that only felt wrong after I used it for a while. AI gets me to those problems sooner. It doesn't make them disappear.

AI is useful during that final stretch as well. It can write tests, inspect logs, suggest fixes, review my code, and help polish the interface. But I cannot walk away and expect it to know when the project is finished. It doesn't know which compromise a client will accept, which small workflow problem will annoy users every day, or which bug matters most. I have to notice those things, make the decisions, and check that everything works.

Starting Is Easier Than Finishing

This is where AI can make me lazy. When I have a working demo in front of me, it is easy to feel like most of the work is behind me. Starting another project is more fun than going back to fix awkward details in the current project. AI can help me start projects faster than I can finish them. Every easy start leaves another last 20 percent waiting for me.

I notice this most with side projects. I can get several ideas far enough to look promising, then find that every one of them still needs quite a bit of polishing and manual work from me.

Client work follows the same pattern, only with higher stakes. I cannot hand over a pile of code or an impressive demo and call the job done. The software has to solve the problem the client hired me to solve. It has to behave correctly, survive everyday use, and be understandable when somebody else works on it later. Getting the main features built faster is valuable, but all the finishing work still has to happen.

So AI saves me an enormous amount of time. It also gets me to the part that needs my attention sooner. But I cannot let that speed make me lazy, because finishing is still my job.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com