The Decision That Disappeared

While working on the NxSync ERP, I noticed a recurring problem. AI tools helped the team move faster, but they also made it easier to lose the reason behind a decision. A change that made sense on Monday could look arbitrary two weeks later, because the conversation that produced it had disappeared.

This was not carelessness. Everyone was working the way the tools invite you to work. You describe a problem in a chat window, the assistant proposes a solution, you refine it together, and the code changes. The output survives. The reasoning does not.

In a small project that hardly matters. In an enterprise system it matters a great deal. When something breaks, the first question is always why is it like this? — and increasingly the honest answer was that nobody could remember, because the "why" had lived in a session that no longer existed.

The pattern, once you see it

I have spent most of my career in physical projects — buildings, automation, interiors — where the space between disciplines is where things go wrong. The electrician and the AV integrator each do competent work and the system still fails, because nobody owned the joint between them.

AI-assisted software has the same weakness in a new place. The joint is not between two trades; it is between sessions. Each conversation with an AI assistant is competent in isolation. The failure accumulates in the gaps: a requirement quietly reinterpreted, an architectural boundary moved without record, a dependency assumed that was never checked.

The speed was real. The confidence in what we had built was declining even as the output grew.

What we changed

We stopped starting with the tool and started with the problem. Before any AI-assisted work began, the business problem was written down, along with what "finished" would mean for that piece of work. Work was cut into bounded pieces small enough to review properly. Decisions that changed requirements or architecture were recorded outside the chat, where the next person — or the next session — could find them. And nothing shipped without a human looking at it against the original intent.

None of this is novel as project discipline. What surprised me was how much of it the industry quietly dropped the moment the tools became conversational.

That set of habits eventually became a product, Spine, because I wanted the discipline to be a system rather than a resolution. But the habits matter more than the software. You could keep most of them with a text file and some patience.

What I am still unsure about

I do not yet know how large a piece of autonomous AI work can safely be before it needs a human checkpoint. The honest answer is that the limit keeps moving, and anyone who claims a fixed rule is guessing.

What I am confident about is smaller and more durable: if you cannot explain why a system is the way it is, you do not fully control it. That was true of buildings before it was true of software, and it will stay true whatever the tools become.