Trust
Most AI platforms ask you to trust a policy document. NOW OS is built so the trust properties are structural — enforced by the architecture, checked by machines, and re-verified while you watch.
The five structural guarantees
Every consequential thing an agent discovers or recommends is filed as a proposal to your Strategic Advisory Board. A named human approves, denies, or defers — the engine cannot promote its own output into your operation. This isn't a setting; there is no other code path.
Governance decisions and execution events land in a hash-chained audit log. When anyone views it, the chain is recomputed live — a broken link is visible immediately. The "verified" badge is earned on each page load, never cached.
Work touching FERPA, HIPAA, PCI, export-control, and similar regimes carries control floors the scoring engine cannot cross — it can never be silently over-automated, no matter how favorable the score. Non-compliant work is refused at ingest, fail-closed.
Every fact entering the live knowledge graph is validated against schema and contract before it is written, queued for a named reviewer, and re-validated on promotion. There is no side door, and destructive operations carry pre-state snapshots.
Role walls, ownership checks, and persona boundaries are covered by regression suites and adversarial review passes — the product's own release gates fail if a wall regresses. What a requestor can see is decided server-side, fail-closed.
When an answer draws on reference or shared knowledge, the borrowed source is labeled in the response and in the sealed audit payload — you always know which facts are yours and which are contextual.
At production
A production tenant runs with enterprise single sign-on (Microsoft Entra ID), secrets held in Azure Key Vault (fail-closed — no baked credentials), tenant isolation at the runtime, and live-system connectors that ingest your data under your governance. Evaluation instances are the same runtime with the production machinery dormant — cutover swaps identity, secrets, and data sources; the engine doesn't change.
Sealed intellectual property: the per-dimension scoring weights, readiness formulas, and policy-floor values are never disclosed — the platform always explains what each dimension means and why an action landed where it did. Aspects of NOW OS are the subject of pending U.S. provisional patent applications.
Common questions
No. The conversational surfaces exist, but the core is a computed model of your organization's work — decomposed to atomic actions, scored across 26 dimensions, and operated with human gates. Retrieval is one layer of it; the work model is the product.
It computes an honest answer per action: some work automates, some gets augmented or assisted, and some stays Human-Only by design — accountable approvals can't be automated away, and restricted work carries enforced floors. The output is a governed division of labor, not a headcount plan.
Nothing, to evaluate — the demo instance runs on simulated data. A pilot needs your policy corpus and organizational data for one process; the work model is discovered from your real documents, not hand-authored from templates.
Leading frontier models (Anthropic's Claude family) for reasoning, orchestrated over your knowledge graph with tool use, plus a library of 49+ deterministic quantitative capabilities. Model reasoning is grounded in retrieved facts with provenance, and audit-grade outputs are bit-reproducible.
Not until they're earned. Everything runs today on a high-fidelity reference dataset; at cutover your live data recalibrates the models, and stronger claims are earned through non-circular backtests under your board's sign-off. We say "built" and "validated" precisely, and never swap one for the other.
You do. The engine recommends; your Strategic Advisory Board approves, denies, or defers; the ledger records. Every decision is replayable.
A briefing covers the architecture, the governance model, and the audit trail live on the running system.