AI-Native Workforce Governance

The Inquiry: How must traditional workforce governance — the twelve HR functions that manage organizational participation through policy — transform when AI actors join the workforce, given that AI breaks the epistemic contract that makes policy-based management work?

Falsifiable formulation: If policy-based approaches can reliably enforce the human/AI output distinction, then governance infrastructure is unnecessary overhead. The empirical evidence (Dell'Acqua et al. 2023, Kellogg et al. 2026) suggests that policy-based approaches fail precisely at this boundary — organizations that add AI governance policies to existing HR stacks find that 80%+ of deployment effort goes to governance challenges that policy cannot solve.

The epistemic break is the foundational problem. The traditional HR stack works because of an implicit epistemic contract among human actors: when someone asserts something, the assertion carries the weight of personal understanding, professional judgment, and individual accountability. AI breaks this contract. An AI system can produce output that is syntactically indistinguishable from expert human work but is epistemically different in kind — it is pattern inference, not grounded assertion. Dell'Acqua et al. (F4) showed this is not a theoretical concern: BCG consultants using AI performed worse on tasks outside the jagged frontier because they could not distinguish AI capability from AI limitation. The epistemia literature (F5) identifies this as a structural condition — not a user error but an architectural property of how AI-generated output enters knowledge workflows.

The epistemic break argument is stronger than a "policy gap" argument. It is not that organizations lack AI governance policies. It is that policy cannot enforce a distinction it cannot see. A policy that says "human review required before publishing AI output" is unenforceable when the reviewer cannot reliably identify which parts of a document were AI-generated, AI-edited, or human-authored. Structural enforcement — typing the distinction into the record itself — is the only reliable mechanism.

Traditional HR theory does not accommodate AI actors. The foundational frameworks — Mintzberg's archetypes (F1), Lepak & Snell's HR architecture (F2), Wright & McMahan's SHRM perspectives (F3) — all assume exclusively human actors. This is not a criticism of these frameworks; they were developed before AI actors were organizationally relevant. But it means that extending traditional HR to AI actors is not a matter of adding a "fifth employment mode" or a "seventh coordination mechanism." The foundational assumptions (human capital, employment relationships, social coordination) break. The VRIO framework (Barney 1991) breaks: AI actors are valuable but not rare, capable but not inimitable. Agency theory breaks: AI actors have no interests to misalign, but also no judgment to apply. Institutional theory (DiMaggio & Powell) explains the current mimetic isomorphism — organizations copying each other's ad-hoc approaches — but not what a theoretically grounded alternative looks like.

Four transformation patterns. The twelve traditional HR functions transform along four patterns (F8): some are replaced by structurally superior governance mechanisms (performance management by evidence-based graduation; compliance by structural enforcement; records by typed decision lineage). Some are split into multiple AI-native categories (recruitment into human recruitment + AI evaluation + hybrid team composition; training into human development + AI configuration + world model orientation). Some gain new triggers and modalities (offboarding adds model migration; compensation adds token-based cost accounting). Some face unprecedented problems with no traditional analog (model version succession, cross-type capacity planning, epistemic provenance tracking).


The parallel-stack problem. NIST AI RMF (F17) and similar AI governance frameworks create governance structures for AI systems. Traditional HR creates governance structures for human workers. When a human and an AI collaborate on the same deliverable, both stacks apply — but they don't integrate. The parallel-stack problem is the organizational manifestation of the epistemic break: two governance systems operating on different assumptions, with no integration point, for actors producing indistinguishable output. The solution is not to merge the two stacks (they govern genuinely different things) but to create a unified governance infrastructure that coordinates them — a typed record system where every piece of output carries its epistemic provenance (who/what produced it, under what authority, with what evidence).


The 80% finding as validation. Kellogg et al.'s finding (F14) that 80%+ of AI deployment effort goes to governance — not model development — validates the thesis that governance infrastructure is the primary workforce challenge. The five heavy lifts (data integration, model validation, economic value, drift monitoring, governance) are all workforce governance functions wearing technical disguises. "Data integration" is the workforce question of what information each actor can access. "Model validation" is the workforce question of performance assessment. "Drift monitoring" is the workforce question of ongoing capability verification. The hardest work "isn't in deploying the model or writing smarter algorithms, but transforming the organization to support these things" (Kellogg) — this is the workforce governance thesis stated by a practitioner.