How We Use AI and Data
Status: July 2026 · aligned to canon generation 8.
Doctrine: AI compresses the work. Source systems prove the facts. Humans make every credentialing, privileging, hiring, and clinical decision.
Rōvn's AI is the labor layer, not the judge. Agents are not the moat; agentic recruiting and agentic workflow are not unique claims. The moat is what the agents can governably touch: consented cross-employer evidence, each organization's versioned policy, resolution history, and receipts. That framing is deliberate, because it is the one that survives both a diligence read and a hospital General Counsel read.
1. What the agents operate
Governed agents run the repetitive work between opportunity and approved work: detecting risk, opening Resolution Cases, requesting or reusing evidence, extracting documents, performing allowed source checks, chasing renewals, computing impact, preparing packets and recommendations. They pause at every regulated gate. The loop terminates in a Work Activation, approved to start, with proof: a named human signs, and a receipt binds the evidence, the policy version, and the decision. AI never approves, hires, rejects, credentials, privileges, or screens out anyone.
2. We do not build a foundation model
The compounding asset is the Rōvn Workforce Model: the proprietary exception-resolution and outcome-labeled dataset that operating the workflow produces, plus the specialist models trained on it, not a foundation model Rōvn built. The fight is not with frontier labs. A general model can reason broadly; it cannot know how a specific facility clears workers, which documents stall which role, or which exceptions repeat, because that data only exists inside operated workflows.
3. The data that compounds
The valuable data is generated by operating, not scraped:
- identity conflicts and source disagreements;
- missing or ambiguous evidence and policy exceptions;
- worker corrections and disputes;
- named-human decisions and their rationale;
- clearance trajectories, monitoring events, and assignment impact;
- operational outcomes per facility and per requirement version.
4. Two intelligence layers, one hard boundary
- Shared layer: de-identified source behavior, reusable patterns, evaluation sets, and specialist models.
- Private organization layer: local policy, reviewer patterns, mapping templates, exceptions, and operating state.
Raw tenant data never becomes cross-tenant training data by default. Training eligibility requires a contractual basis, de-identification status, an approved dataset version, lineage, access controls, and an evaluation plan. Protected characteristics and hidden proxies are excluded at the schema layer.
5. The proof gate for any model claim
Rōvn claims no proprietary model accuracy advantage today. The specialist-model moat becomes real only when Rōvn can show:
- a held-out evaluation set;
- a raw frontier-model baseline;
- a measured Rōvn specialist result;
- accuracy, abstention, and error analysis;
- improvement over time;
- no hidden protected-class proxy use.
Until that exists, the AI moat is prospective, and this room labels it that way.
6. How this is different, stated defensibly
Every layer of this market has a real company on it, and the full concession-first competitive read lives in 10.3 Competitive Landscape. The short version: employer systems administer one organization's workforce, marketplaces reuse credentials inside their own economics, and credentialing platforms run facility-owned workflows. Rōvn's center of gravity is repeated, governed Work Activation across employers outside any single marketplace: worker-controlled evidence, organization-specific policy, one Resolution Case, named-human approval, signed proof, no placement or shift take rate, and continuous post-hire operation. The falsifiable test is measured, not asserted: the second Work Activation must be materially faster and cheaper than the first, or the thesis is wrong.