The Three Foundations
Deterministic workflows, governance, and observability — what they mean, and why every system I build has all three.
Every AI product gets sold on what it can do. The more useful question is what holds it up.
There are three things underneath everything I build. They aren’t features, and you won’t see them in a demo — they’re the structure the features sit on. If one of them is missing, the other two can’t cover for it.
Here they are in plain language.
1. Deterministic workflows
Deterministic is a long word for a simple promise: the same question, asked of the same information, always gets the same answer.
That sounds like it should go without saying. With AI, it doesn’t. A language model is a probability engine — ask it the same thing twice and you can get two different answers, and nobody, including me, can fully explain the difference. That’s tolerable when it’s writing a sentence. It’s not tolerable when it’s deciding whether your file needs attention.
So in the systems I build, the deciding is not done by AI. It’s done by rules, and the rules are plain settings written in a file a person can read:
- A file with no activity for thirty days is flagged.
- A required document that isn’t there is flagged.
- An assessment past its due date is flagged.
That’s subtraction and comparison. It’s arithmetic. It doesn’t need a language model, and putting one in the middle of it makes it less reliable, not more.
The AI still has a job — a real one. It takes what the rules found and writes it in readable English instead of a table. It reads a messy document and pulls the date out. It drafts the note a person is about to review. Those are things it’s genuinely good at, and where an awkward sentence costs you nothing.
What this buys you: you can check the work. You can look at a flag and see exactly which rule produced it and why. You can change a threshold from thirty days to fourteen and know precisely what will change. And when someone asks why the system said what it said, the answer is a sentence, not a shrug.
2. Governance
Governance is the boring word for who is allowed to see what, who is allowed to approve what, and what the software is allowed to do on its own.
Most organizations have this written down somewhere. The difference that matters is whether it’s a document or a mechanism. A policy says people shouldn’t look at records outside their role. Governance means the software won’t let them — and won’t let the AI either.
Three things it decides, in practice:
Who sees what. Someone in billing has no business opening a full case file. A family member should see their own relative and nothing else about the building. The check happens before the information is fetched, not after — so there’s nothing sitting in the page waiting to leak. A system that gathers everything and then hides most of it is a system one bug away from showing you all of it.
Who approves what. Not every action deserves the same signature. Attaching a document to a checklist is not the same as changing a deadline or a care level. The bigger the consequence, the more deliberate the approval — and the approval gets recorded with a name on it.
What the AI may do by itself. In most of what I build, the answer is nothing that changes your records. It reads, it flags, it drafts. A person decides. Where that ever changes, it changes one narrow action at a time, deliberately, and never because something demoed well.
What this buys you: the damage any single mistake can do is bounded before it happens, rather than investigated afterward.
3. Observability
Observability means you can see what the system did — and what it declined to do — without taking anyone’s word for it.
From the first day it runs, the system keeps a record: what it looked at, what it flagged, who opened it, who approved what, and who was denied access to what. That last one matters more than people expect. A system that only records its successes can’t demonstrate that its rules are working. The denied request — someone reached for something outside their role and got nothing — is the proof that the governance layer is load-bearing rather than decorative.
The record deliberately holds identifiers, decisions, and counts. Not case narratives, not clinical detail. A log stuffed with sensitive content is just a second copy of the sensitive content, in a place with weaker protections that nobody thinks of as a filing cabinet.
What this buys you: two things, actually.
The first is that you can tell whether the thing is working. Not a feeling — a number. “It has flagged this correctly for four months” is a sentence you can support.
The second is the one clients tend to appreciate later. The day someone asks you to prove how you operate — a surveyor, opposing counsel, a family member with a fair question, an insurer — the record already exists. It was produced continuously as a side effect of running, rather than reconstructed out of email and memory under time pressure.
Why all three, and not two
They interlock, and the gaps are specific.
Without deterministic workflows, your log faithfully records decisions nobody can reproduce or explain. You have evidence of something you don’t understand.
Without governance, the log becomes a careful record of things that shouldn’t have happened in the first place.
Without observability, the first two might be working perfectly and you have no way to know. You’re trusting an assurance, which is exactly the position you were trying to get out of.
Together they answer the question I think is the only one that matters when someone shows you an AI tool: what happens on the day it’s wrong, and how would you find out?
With these three in place, the answer is concrete. The rules are wrong in a way you can read and correct. The blast radius was limited before it happened. And you found out because the system wrote down what it did.
That’s not the most exciting thing to put in a demo. It’s the reason the software is still trustworthy in month eighteen.
I’m not a lawyer and I’m not a compliance consultant. I build software, and I design it against the rules that apply to the business it’s for. What your obligations are is between you and your own counsel — I can only tell you what I designed for and why.