The Determinism Problem: Why Regulated Industries Can’t Just “Add AI”
Everyone wants an AI orchestration layer. In a pharmacy, a bank, or anywhere a regulator can subpoena your logs, the hard part isn’t the model — it’s the parts that are not allowed to be probabilistic.
I spent a good chunk of last year leading product and architecture for a Fortune 10 retail pharmacy’s omnichannel customer communications. The brief sounded like every other AI brief in 2025: make it smarter, make it conversational, make it feel like one company instead of twelve disconnected systems. The reality was more interesting, and more constraining.
When you move a message about a prescription, you are not just moving content. You are moving something a regulator can audit, a lawyer can discover, and a patient can be harmed by if it’s wrong. The question stops being “can the model do this?” and becomes “can we prove” what the system did, every time, forever?
Orchestration is a spectrum, not a switch
The trap is treating AI as a binary — either the system is intelligent or it’s dumb. The useful frame is a spectrum of determinism. Some decisions must be 100% deterministic: eligibility, dosage references, regulatory disclosures. Some can be probabilistic with a human in the loop. And some — tone, sequencing, channel preference — can be fully delegated to a model because the blast radius of a mistake is small.
The architecture’s job is to draw those lines explicitly, not to let them blur because the demo looked magical.
What that looks like in practice: a deterministic core that owns the things you’ll defend in front of a regulator, wrapped in an orchestration layer that can be as clever as you like — but that can never overwrite a decision the core has made. The model proposes. The core disposes. Every handoff is logged.
The unglamorous payoff
This isn’t the architecture that wins the keynote. It’s the one that survives the audit. And in regulated industries, surviving the audit is the product. The teams that win the next five years won’t be the ones who added AI fastest — they’ll be the ones who were precise about where intelligence was allowed to live.