AI Doesn’t Break. It Drifts.

The bigger an AI system gets, the more machinery it needs just to keep running. More rules. More edge cases. More places for it to quietly drift off course.

I run four AI agents for my own business, every day. The model was never the hard part. The hard part is drift, and drift is worse than breaking, because nobody notices it until a decision has already been made on it.

What drift actually looks like

My system keeps every version of every fact it has ever been told: old prices, old splits, old decisions. The superseded versions outnumber the current ones. Left alone, an agent answers with whatever it finds first, and what it finds first is often last month's truth. It says it with complete confidence. Nothing crashes. The answer is simply wrong.

The fix was not a better model. It was a ledger: one place that is allowed to say what is true right now, which everything else has to check before stating anything as fact.

Rules drift too. Every fix adds an instruction, and instructions pile up. One of my agents grew a rulebook heavy enough that it kept running out of working room halfway through a task. So every rulebook now has a size budget, checked each time a session starts. This morning that check flagged one of the four at roughly twice its budget. The loop does not end. It only gets watched, or it doesn't.

A launch is not a system

I have watched this play out at scale. A major AI platform was built well, and at least SAR 10 million went into the initial build alone. But without a standing team of AI architects and engineers tending it after launch, it went stale fast. Not because the model got worse. Because nobody was maintaining the machine underneath it.

Ten million riyals buys you a launch. It doesn't buy you a system that stays true a year later.

Nobody selling you AI leads with that. They lead with the launch, not the six months after.

What maintenance means in practice

None of it is glamorous. In my own system it comes down to three habits.

One source of truth. A single record decides what is current. Everything else is history, and is treated as history.

A budget on complexity. Every new rule has a cost, and the system measures that cost instead of pretending it is free.

A cold read on a schedule. A separate agent whose only job is to read the others' rules with fresh eyes, plus a monthly maintenance session that sits on the calendar like any other meeting.

Each of these is dull. That is the point. Maintenance is the work nobody wants to own, which is exactly why it decides whether an AI system is still worth anything on day one hundred.

The question to ask before you buy

Not what it does on day one. Who is maintaining it on day one hundred?

That is the question AI.me is built around. Each client gets an AI second brain tailored to them, and a relationship manager who keeps it true. You drive. We handle the maintenance.

AI doesn't break. It drifts. Someone has to be watching.

Dennis Noel Gomez is the founder of AI.me, an applied AI solutions provider in Riyadh, Saudi Arabia. He builds working executives an AI second brain and runs one himself. These notes are written from the builds as they happen.

This article expands a LinkedIn post.