01 · Working principle
An operating teammate is not a generic chatbot
A generic chatbot begins with a prompt and returns language. An operating teammate begins with a bounded leadership question: which cohort, which time window, which definition, which evidence vintage, and which decision is actually in scope? The output should preserve those boundaries rather than smoothing them away.
The useful unit is not an impressive answer. It is a reviewable chain from question to evidence to human decision. Hospital leaders should be able to see what is measured, what is inferred, what remains unknown, and who owns the next step.
02 · Working principle
Ask → Brief → Decide → Act → Learn
sanalytics uses a five-stage operating framework. Ask scopes the question. Brief assembles definitions, signals, relevant comparisons, and limits. Decide records the human judgment. Act connects approved work to an owner and review date. Learn returns the observed outcome to the next review without claiming causation that the evidence cannot support.
- Ask: name the decision, population, time window, and decision owner.
- Brief: separate observed facts, possible explanations, and missing evidence.
- Decide: preserve options, assumptions, dissent, and human approval.
- Act: record the owner, operating measure, and next review point.
- Learn: compare what followed with the original expectation and limitations.
03 · Working principle
A practical evaluation standard
Before adopting an executive AI workflow, ask whether it can cite the definition and source behind a number, distinguish deterministic calculations from generated explanation, and make uncertainty visible. Also ask how approvals are recorded, how access is bounded, and whether a future reviewer can reconstruct why a decision was made.
Those questions are more useful than a long feature list. They reveal whether the system supports management judgment or merely produces fluent text around disconnected data.
Executive checklist
What leaders should ask next
- Can every number be traced to a definition, source, cohort, and vintage?
- Which parts are deterministic, which are generated, and which require human judgment?
- How are uncertainty, conflicting interpretations, and missing data shown?
- Who can approve a decision, and how is that approval recorded?
- What measure and review date will show what happened next?
Interpretation boundary
Limits to keep visible
- AI should not make patient-level clinical recommendations or silently change metric definitions.
- A generated explanation is not causal evidence and does not prove that an intervention produced an outcome.
- The public sanalytics experience is scenario-based and uses synthetic views; deployment scope is established separately.
Authoritative references
Sources and method notes
- National Institute of Standards and Technology: AI Risk Management Framework
Voluntary framework for managing AI risk, roles, and accountability. Accessed August 19, 2026.