31 August 2026
We scope and discuss a lot of AI projects nowadays, and the same picture repeats: most of the work is not AI. A product with a model inside still needs everything ordinary products need, and building that is where most of the effort goes. This year the typical request is an agent that answers calls, schedules jobs, builds a product catalog, or sorts documents. Every brief mentions AI. Very few of them say where exactly the work gets stuck today. Break any of these ideas into tasks and the list fills up with plain software: collecting and cleaning data, business rules, calendar logic, integrations, an admin panel where a human can see what happened. That part is deterministic. It either works or it does not, and we build and test it like any other software. AI earns its place in a much smaller slice. Reading messy human input. Classifying things that resist strict rules. In those spots a model beats hand-written code, and we use one. Everywhere else it adds cost and wrong answers to a problem code already solves. So when you bring us an AI project, expect a boring first step. We will find the real bottleneck and tell you which part needs AI at all. Sometimes the honest answer is ten percent of the whole workflow. Most of the work is still not always AI, and while not so exciting that is not a bad thing. Plain software engineering is predictable, and we have been doing it for over twenty years. How we use AI: https://abz.agency/how-we-use-ai
