Research article · Global Business & Economics Journal · July 2026
AI Is Not Software
Fifty years of software assumptions — deterministic, obedient, nearly free to scale — do not hold for probabilistic systems. Planning as if they do is how AI programs quietly go wrong.
Read it at Global Business & Economics Journal ↗What it argues
- What it can reach: the blast radius of an AI system is defined by its access, not its accuracy.
- What you will believe: decide in advance which outputs are advisory and which are allowed to be authoritative.
- What you can reproduce: a system you cannot re-run and explain is a system you cannot defend.
- What it costs to serve: inference is a recurring unit cost, unlike the marginal cost of traditional software.
- How much of it you actually build: most organizations should assemble and govern far more than they train.
Why I wrote it
Traditional software gives the same answer twice. It does exactly what it is told. Once it is written, serving one more user costs almost nothing. Every planning habit in enterprise engineering — estimation, testing, capacity, procurement — rests on those three properties, and probabilistic systems break all three at once.
So I proposed five operational boundaries instead of a maturity model. Boundaries are decisions a leadership team can actually make in a room: where the system reaches, what its output is permitted to mean, what evidence it leaves behind, what a served answer costs, and how much of the stack is genuinely yours to build.
Teams that set those boundaries early end up shipping less impressive demos and far more durable systems.