AI Is Splitting the Engineers – and the Answer Is Architecture
“I can’t put this into production like that.”
The sentence came up in an architecture review. Not from fear. Not from ideology. From engineering instinct.
“I can’t guarantee the behaviour.”
And right there begins the real discussion about AI. Not around ethics. Not around copyright. Around system behaviour.
Two Mental Models Collide
Classical engineering culture rests on an implicit promise: if we specify cleanly enough, the system becomes controllable. Same input, same output. Deviation is a bug.
That model is powerful. But it is not universal.
In radar data processing, that promise never existed. A target is not “there” or “not there”. It is detected with a probability. You define thresholds, model false alarms, build filters. You accept uncertainty – and control it indirectly.
AI Is Closer to Radar Than to Classical Software
An LLM is not a deterministic code block. It is a probabilistic core. It approximates. It responds sensitively to context.
This unsettles those who think of systems as fully specifiable. And it is entirely reasonable for someone to say: “I can’t put this into production.”
If you try to treat a stochastic system as a deterministic one, it becomes unstable.
The Fault Lies Not in the Model – but in the Architectural Understanding
The real problem is not that AI is non-deterministic. The problem is running it without deterministic embedding.
A probabilistic core without controlling structure produces chaos. But a probabilistic core with a deterministic frame is highly effective.
And that is precisely where architecture begins.
The Deterministic Spine
In my own systems – the Policy Monitor, for example – the AI module is never the system. It is a component.
The spine is deterministic:
- clearly defined inputs
- controlled prompt structures
- validation logic
- formal output formats
- reproducibility rules
- logging and audit trails
The AI is allowed to interpret. But it is not allowed to define the system boundaries. That is the architecture’s job.
This Is Not a Compromise – It Is Classical Systems Architecture
Probabilistic subsystems are stabilised by deterministic control layers. Exactly as in radar systems, autonomous driving functions, sensor fusion, risk assessment models.
Nobody would wire a radar directly to a launch mechanism. You build decision logic in between.
Why would you deploy an LLM without a spine into production systems?
The Real Split
The split does not run between “AI fans” and “AI sceptics”.
It runs between:
“I need full determinism” and “I can contain non-determinism architecturally.”
Those who only know deterministic systems experience AI as a threat. Those who understand probabilistic systems see a new class of building blocks.
Whoever says “I can’t put this into production like that” is right.
The answer is not: “Then don’t.”
The answer is: “Then build the frame around it.”