Administrative Burden in Safety Work

• Administrative burden in safety work is the cumulative cost of learning, psychological strain, and compliance required to maintain auditable safety records across fragmented systems.
• It arises from dynamic change propagation, manual traceability maintenance, and coordination challenges across teams, suppliers, and regulatory boundaries.
• Integrated approaches using formalization, automation, and incremental safety methods offer actionable strategies to mitigate overhead and prevent administrative bloat.

Searching arXiv for the cited papers and closely related work on administrative burden in safety-critical and regulated systems. Administrative burden of safety work is the aggregate learning, psychological, and compliance cost required to keep safety-relevant artifacts, decisions, and evidence synchronized under changing technical and regulatory conditions. In software-intensive safety engineering, it appears as repeated manual effort to maintain consistency between disconnected “content islands”; in public-administration research, it is classically framed as “learning, psychological, and compliance costs”; and in organizational theory it can escalate into administrative bloat when process creation outpaces pruning of obsolete controls (Cârlan et al., 10 Sep 2025, Jo et al., 28 May 2025, Yang et al., 2024). Across automotive development, nuclear licensing, healthcare prior authorization, and welfare administration, the recurrent pattern is not merely paperwork, but the maintenance of a defensible, auditable, and current safety record across tools, teams, organizations, and time.

1. Domain scope and conceptual framing

The literature treats administrative burden in safety work as a cross-domain phenomenon rather than a sector-specific inconvenience. In safety engineering, the burden is generated by fragmented tools, document-centric workflows, manual traceability maintenance, recurrent impact analysis, and the need to re-establish the validity of a safety case after change (Cârlan et al., 10 Sep 2025). In scaled agile safety-critical development, the same burden is described through “living traceability,” “continuous compliance,” and “organizational flexibility,” all of which require ongoing coordination across teams, suppliers, baselines, and variants (Steghöfer et al., 2019). In nuclear licensing, the burden is located in the formal human-to-human pipeline across an organizational trust boundary, where even routine clarification can trigger a request-for-additional-information cycle with drafting, routing, response preparation, docketing, and recordkeeping (Dave et al., 5 Jun 2026).

Related literatures extend the concept beyond conventional engineering safety. Prior authorization is described as one of the most burdensome administrative processes in U.S. healthcare, with safety-relevant consequences because delays in authorization can delay treatment, contribute to burnout, and increase downstream harms such as emergency department utilization and preventable hospitalizations (Awan et al., 31 Mar 2026). SNAP administration is analyzed as a high-stakes welfare environment in which small failures can produce delayed, denied, or lost assistance, making administrative burden part of social safety infrastructure rather than a mere information problem (Jo et al., 28 May 2025).

Domain	Primary burden mechanism	Representative paper
Software and system safety	Manual consistency maintenance between fragmented artifacts	(Cârlan et al., 10 Sep 2025)
Scaled agile safety-critical development	Living traceability and continuous compliance across teams	(Steghöfer et al., 2019)
Nuclear licensing	Human-to-human regulatory exchange across a trust boundary	(Dave et al., 5 Jun 2026)
Prior authorization	Weak administrative scaffolding around clinically strong letters	(Awan et al., 31 Mar 2026)
SNAP benefits administration	Trust-shaped learning, compliance, and psychological costs	(Jo et al., 28 May 2025)

A common misconception is that administrative burden is synonymous with forms and filing. The surveyed work instead depicts it as “manual glue work”: the labor of preserving semantic, procedural, and legal coherence across heterogeneous artifacts and institutions.

2. Structural sources of burden

A central source of burden is fragmentation. Safety tools are described as “isolated from each other and from software engineering tools,” each with its own persistence mechanism, producing “islands of (partially overlapping) content.” In that setting, automated checking of traceability and consistency becomes difficult, and safety experts invest manual effort to maintain consistency between content islands (Cârlan et al., 10 Sep 2025). Documentation is often produced “just” to satisfy process and legal requirements, frequently in plain text, and information sent to assessors remains document-based rather than available in the development environment.

The same structural pattern appears in large-scale agile practice. Requirements, safety requirements, design decisions, code, tests, reviews, safety cases, and baselines live in different systems, evolve at different rates, and are owned by different teams. Traceability therefore becomes an ongoing coordination problem rather than a one-time documentation step. Baseline creation is difficult because artifacts have different lifecycles and tool locations, and supplier interaction is often document-mediated, forcing manual re-entry of requirements and manual trace-link updates (Steghöfer et al., 2019).

Nuclear regulation introduces an additional structural factor: the trust boundary between applicant and regulator. The expensive part is not only technical analysis but the maintenance of a shared technical record under strict auditability and confidentiality constraints. The paper on the Regulatory Context Protocol reports review cycles of 60–90 days per cycle in NRC practice and shows, from an analysis of 1,236 documents, that 69% were RAI responses while 95% had no blocking issues. The implication drawn there is that a large fraction of the burden resides in routine cross-boundary exchange rather than substantive disagreement (Dave et al., 5 Jun 2026).

This convergence across domains suggests that administrative burden is best understood as a property of socio-technical architecture. When authoritative content is duplicated, stored in incompatible formats, or transmitted through document-heavy channels, the burden emerges from reconciliation itself.

3. Change propagation, traceability, and continuous compliance

Safety work is unusually sensitive to change. One paper states explicitly that whenever a change occurs, safety standards require performing an impact analysis describing the change, affected work products, and impact on functional safety, followed by actions to “re-establish the validity of the system’s safety case” (Cârlan et al., 10 Sep 2025). Administrative burden therefore recurs with every modification rather than only at release time.

Scaled agile amplifies this burden because the system evolves continuously while safety evidence must remain current. The industrial challenges identified in automotive and medical-device settings are concentrated in three areas: living traceability, continuous compliance, and organizational flexibility. Traceability must support both top-down and bottom-up use; it must also cover review status, changes, decisions, and rationale, not only product artifacts. Continuous compliance requires updating the safety case on demand, supporting delta analysis instead of full reassessment, and covering product-line variants rather than a single configuration. Organizational flexibility adds supplier coordination, mixed-criticality handling, and change-request triage (Steghöfer et al., 2019).

The burden becomes especially visible where change granularity and compliance granularity are misaligned. If delta analysis is unavailable, a small modification can trigger work resembling re-certification. If the traceability information model is underspecified, teams may be forced into large amounts of manual link maintenance. If variant structure is not reflected in the safety case, organizations may either oversample and overdocument or miss relevant combinations. These are administrative burdens because they consume effort in artifact maintenance and coordination rather than in hazard reduction itself.

The literature’s recurring remedy is incrementality. In the “safety factories” formulation, effort should become proportional to the size of the change rather than the size of the system (Cârlan et al., 10 Sep 2025). In the scaled-agile literature, the analogous targets are incremental safety analysis, modular safety cases, and localized but consistent safety ownership (Steghöfer et al., 2019).

4. Dynamic accumulation and administrative bloat

The most explicit mechanism model appears in the system-dynamics account of administrative bloat. There, the organization is represented by two process stocks, useful processes $R_u$ and obsolete processes $R_o$, with dynamics

$\frac{dR_u}{dt} = f_c - f_d$

and

$\frac{dR_o}{dt} = f_d - f_p$

where $f_c$ is process creation, $f_d$ is decay from useful to obsolete, and $f_p$ is pruning (Yang et al., 2024). Obsolescence is modeled by

$f_d = \frac{R_u}{T_d},$

so faster environmental change corresponds to smaller $T_d$. Administrative effort is proportional to total process count,

$w_A = k_A(R_o + R_u),$

which means that as processes accumulate, administration consumes a growing share of fixed work resources.

The paper identifies two qualitative regimes: sustainable equilibrium and runaway administrative bloat. The transition depends on management propensities for creation and pruning, denoted $R_o$0 and $R_o$1. High creation propensity is especially dangerous because created processes enter as useful, later decay into obsolete stock, and cannot be undone immediately by pruning. Administrative cost is measured as

$R_o$2

and administrative waste as

$R_o$3

The paper further reports that rapid environmental change worsens the threshold, and that a temporary 12-month shock can push the system from equilibrium into runaway bloat if it exceeds critical magnitude (Yang et al., 2024).

The safety-work implications are direct. Safety organizations routinely respond to incidents, audit findings, or regulatory changes by adding controls, forms, approvals, training refreshers, and reporting steps. The paper’s interpretation of safety contexts is that these processes are initially useful but can become obsolete when technology, operations, regulations, hazards, or staffing change. Once obsolete, they still require sign-offs, audits, exception handling, and compliance verification. The result is “safety bureaucracy” or “compliance clutter”: administrative burden becomes self-reinforcing because the effort needed to prune outdated controls is crowded out by the burden those controls already impose (Yang et al., 2024).

A notable counterintuitive result is that prioritizing direct production can increase administrative waste. The model’s explanation is that raising the desired priority of direct production reduces effort available for both creation and pruning; existing useful processes continue to decay, obsolete share rises, and waste grows even if total administration does not immediately explode (Yang et al., 2024). This is relevant for safety programs that attempt to “go faster” without corresponding investment in cleanup.

5. Formalization, automation, and protocolization

The most explicit engineering response to administrative burden is the “safety factories” program. Its central proposal is to invest more in formality upfront by capturing safety work products in semantically rich, machine-processable models, defining automatic consistency checks, and automating the generation of documentation (Cârlan et al., 10 Sep 2025). The paper organizes this around nine principles, of which several directly target administrative overhead: “Treat Safety Work Products as Code,” “Single Source of Truth,” “Automated Impact Analysis,” “Safety Builds,” “The System Safety Case Drives the Safety Builds,” “Safety Case is Eventually Consistent,” “Live Documentation Close to the Content,” “Company-wide Safety Mindset,” and “Accountability/Ownership as a Given.”

The intended mechanism is straightforward. If safety artifacts are machine-processable and stored with code and tests in the same repository, safety builds can run alongside system builds, continuously checking consistency of safety work products and corresponding software and system artifacts. The safety case becomes machine-readable and functions as a quality gate; build success is associated with a closed safety case and claims supported by sufficient evidence. Duplication is reduced because each information item is saved exactly once, with different views generated automatically. Plain-text documents are replaced by machine-interpretable work products that can be queried, and assessors interact with live models rather than PDFs (Cârlan et al., 10 Sep 2025).

The scaled-agile literature points in the same direction, although with less integrated machinery. Its proposed burden-reduction measures include modular safety cases, incremental or delta safety analysis, better guidance for traceability information models, requirements exchange standards such as ReqIF, safety contracts for black-box component integration, and automation in CI/CD for proof of compliance, link maintenance, change impact detection, and variant handling (Steghöfer et al., 2019).

At the inter-organizational level, the nuclear case study proposes protocolization rather than only model formalization. The Regulatory Context Protocol is a domain profile of an agent-to-agent standard specialized for regulated, safety-critical workflows. It uses four primitives—Agent Cards, Messages, Tasks, and Artifacts—and writes every cross-boundary exchange into a shared append-only Context Stream $R_o$4. Safety-significant or legally binding actions must pass through an auth-required checkpoint where a named human signs cryptographically; without the signature the task cannot complete (Dave et al., 5 Jun 2026). This preserves statutory accountability while replacing routine human-to-human routing with typed, logged, auditable agent communication.

The paper’s quantitative claim is that, against a reconstructed baseline of \$R_o$554M–\$R_o$621M–\$44M. The paper argues that the residual gap is structural, not algorithmic, because it arises from the inter-organizational pipeline that only an agent-to-agent standard can compress (Dave et al., 5 Jun 2026).

These proposals share a common logic: move effort from repetitive manual reconciliation to upfront formalization, executable checks, structured exchange, and incremental validation. They do not eliminate safety work; they redistribute it.

6. AI assistance, verification burden, and trust

The recent literature on AI-assisted administrative work shows that automation can reduce some burdens while creating others. In prior authorization, three commercially available LLMs—GPT-4o, Claude Sonnet 4.5, and Gemini 2.5 Pro—were evaluated on 45 physician-validated synthetic scenarios across five specialties, producing 135 letters in total. Overall rubric performance was very high: mean scores were 11.69/12 for GPT-4o, 11.98/12 for Claude, and 11.82/12 for Gemini, and the paper reports “zero detected clinical hallucinations across 135 letters” (Awan et al., 31 Mar 2026). Clinical narratives were therefore strong.

The administrative side was much weaker. The secondary analysis checked for eight practical prior-authorization elements, including CPT/HCPCS billing codes, authorization duration requests, place-of-service specifications, clinical trial citations, urgency language, peer-to-peer review offers, attachment references, and FDA or NCCN guideline citations. Billing codes were present in 0 of 45 GPT-4o letters, 22 of 45 Claude letters, and 23 of 45 Gemini letters; authorization duration requests were present in 0 of 45 GPT-4o letters, 0 of 45 Gemini letters, and 12 of 45 Claude letters. The paper repeatedly emphasizes that “billing codes and authorization duration represent the largest universal gaps” (Awan et al., 31 Mar 2026). Its implementation recommendation is therefore hybrid: the LLM should generate the persuasive clinical narrative, while a deterministic layer supplies administrative precision from structured sources such as EHRs, formulary databases, payer policy catalogs, and billing-code repositories, with HL7 FHIR and the Da Vinci Prior Authorization Support Implementation Guide as enabling standards.

In welfare administration, the burden transformation is analyzed through trust. The SNAP study used 10 semi-structured interviews with applicants and a GPT-4o-based prototype called SNAP-LLM. It retains the classic administrative-burden taxonomy—learning costs, psychological costs, and compliance costs—but shows that LLMs introduce new burdens within each category (Jo et al., 28 May 2025). Learning costs now include learning how to interact with the AI, what information is safe to share, how the model interprets inputs, and what institutional affiliation and data practices govern the system. Compliance costs now include verifying whether AI responses are correct, cross-checking answers with official sources, and deciding whether to trust the output. Psychological costs include fear of dehumanization, punitive automation, lack of empathy, and concern about whether one is being observed or judged by a machine.

The same paper maps these effects onto three trust dimensions: competence, integrity, and benevolence. Low competence trust increases verification work; low integrity trust increases concern about bias, data misuse, and source legitimacy; low benevolence trust increases fear that the system will not act in the applicant’s best interests. The proposed mitigations are source attribution, recency indicators, uncertainty expressions, clear institutional disclosure, hybrid human-AI arrangements, and task-based user choice between AI and human caseworkers (Jo et al., 28 May 2025).

Taken together, these studies undermine the assumption that AI simply removes administrative overhead. A more precise reading is that AI can automate narrative generation and routine explanation, but may shift burden toward structured data integration, verification, trust calibration, and governance.

7. Persistent controversies and open design commitments

Several controversies recur across the literature. One concerns the role of human review. The “safety factories” paper states that manual reviews should remain “complementary verification methods,” not the primary mechanism of consistency maintenance (Cârlan et al., 10 Sep 2025). The RCP paper preserves human oversight at safety-significant and legally binding decision points through mandatory signature checkpoints (Dave et al., 5 Jun 2026). The prior-authorization paper likewise insists that physician review remains essential because PA letters are safety-critical documents affecting access to treatment (Awan et al., 31 Mar 2026). The common pattern is selective retention of human authority rather than full automation.

A second controversy concerns front-loaded formalization. The safety-factories argument is explicit that later burden reduction requires more formality upfront: semantically rich models, automatic consistency checks, and automated documentation generation all increase initial specification effort (Cârlan et al., 10 Sep 2025). The scaled-agile literature shows why organizations hesitate: defining a traceability information model is costly, changing it later is expensive, and insufficient guidance exists on which traceability decisions best support impact analysis, program comprehension, safety reporting, and trace consistency (Steghöfer et al., 2019). This suggests that burden reduction is partly a capital-allocation problem: organizations must invest in representations and tooling before they can realize gains in recurring work.

A third issue is whether more process always improves safety. The dynamic model argues the opposite can occur: if organizations codify responses to problems faster than they prune obsolete controls, they can cross into runaway bloat, and rapid environmental change makes this threshold easier to cross (Yang et al., 2024). In safety contexts, the implication is that new checklists, approvals, and reporting steps are not costless protections; without retirement mechanisms they can become self-defeating overhead.

The research directions named across the papers are concrete. They include machine-processable safety work products, automatic consistency checks, automated impact analysis, live documentation, modular safety cases, delta-based safety analysis, requirements exchange standards such as ReqIF, safety contracts, typed and auditable inter-organizational protocols, evidence-based information disclosure, and explicit trust calibration (Cârlan et al., 10 Sep 2025, Steghöfer et al., 2019, Dave et al., 5 Jun 2026, Jo et al., 28 May 2025). A plausible implication is that the future of safety administration lies less in producing more documents and more in maintaining executable, queryable, and selectively human-governed safety knowledge.

The administrative burden of safety work is therefore not a residual inefficiency at the edge of technical practice. It is a central systems problem produced by fragmentation, change propagation, trust boundaries, and process accumulation. The literature converges on a common conclusion: sustainable reduction of this burden requires integrated representations, incremental reasoning, deliberate pruning of obsolete controls, and human oversight concentrated at genuinely safety-significant points.