Facts, evidence and governed state
Source-backed information and persisted state that can be traced to the underlying basis.
How CAI works
From system truth to controlled action.
Across decision preparation, professional assurance and transformation, CAI separates deterministic truth, AI reasoning and human authority instead of collapsing them into one opaque workflow.
The common architecture
Six layers that recur across the CAI product portfolio.
The domain changes. The control philosophy stays consistent.
- 01
System truth
Assemble authoritative evidence, exact source state and the operating context that matters.
- 02
Deterministic rules
Apply policy, thresholds, hard blockers and state transitions explicitly where the answer should not be probabilistic.
- 03
Bounded AI reasoning
Use model assistance for interpretation, extraction, challenge and explanation within an explicit scope.
- 04
Human authority
Keep material review, approval, reliance, sign-off and override attributable to authorised people.
- 05
Controlled action
Route the approved next state or action through defined controls rather than treating a recommendation as execution.
- 06
Provenance & evaluation
Preserve the basis, changes and outcome so the governed state can be inspected, re-evaluated or reperformed.
What stays explicit
CAI does not turn every input into one opaque score or model answer.
The operating design preserves the distinctions that matter when the outcome is consequential.
Rules, thresholds and state transitions
Logic that should be reproducible and testable remains explicit.
Interpretation with bounded uncertainty
AI can extract, classify, challenge and explain without inventing missing facts or silently becoming authority.
Accountable judgement and authority
Consequential decisions remain attributable to the people authorised to make or sign them.
Governed automation
Automation is earned, not assumed.
CAI increases automation only where the consequence, confidence and control environment justify it. Higher-consequence work keeps stronger review, approval and override boundaries.
Deterministic first
If a fact, rule or blocker can be established deterministically, the model does not replace that logic.
AI within a boundary
Model reasoning is used where interpretation adds value and its output can remain distinguishable from authoritative state.
Human where consequential
Review, approval, reliance and sign-off stay explicit when the consequence warrants professional or business authority.
Provenance and evaluation
The basis is part of the product.
A governed result should preserve enough context to understand what evidence and rules applied, what AI proposed, what a human decided and what changed afterwards.
Depending on the product, that may support decision lineage, professional reperformance, historical equivalence or forensic before-and-after verification.
Where CAI applies
Start where the consequence justifies a governed operating system.
The strongest fit is work where evidence is fragmented, deterministic logic matters, judgement remains material and the resulting state must be inspectable.