Workers whose AI agents can autonomously retrieve personal context gain a sharply amplified advantage: when users must manually attach documents, task-success probability collapses as knowledge corpora grow, creating a structural divide in AI usefulness and workplace outcomes.

The Context Access Divide: Interaction-Level Architecture as a Complementary Dimension of Agentic Inequality

Masahiro Fujita · July 09, 2026 · arXiv (Cornell University)

openalex theoretical n/a evidence 7/10 relevance Full text usable extracted full text Source PDF

The paper defines the Context Access Divide (CAD) — whether AI systems autonomously retrieve a user's knowledge versus requiring manual attachments — and shows via a probabilistic model that manual attachment induces a combinatorial collapse in task-success probability as personal corpora and task conjunctivity grow, producing a structural source of AI-mediated stratification.

Sharp et al. (2025) introduce "agentic inequality" as a framework for analyzing disparities in access to AI agents across three dimensions: availability, quality, and quantity. These person- and organization-level dimensions characterize who can access agents and at what capability, but do not address a structurally important divide operating at a finer level: the individual interaction. Two users with nominally equivalent agent access may experience qualitatively different AI utility depending on whether the system can autonomously retrieve context from the user's knowledge corpus (Dynamic Context Retrieval) or requires the user to manually identify and attach relevant documents at each query (Manual Attachment). We term this the Context Access Divide (CAD). For knowledge-intensive workers whose intellectual capital spans tens of thousands of files, the CAD constitutes a qualitative threshold in AI usefulness: below it, the cognitive burden of context curation falls on the human, reproducing the inefficiencies AI is meant to eliminate. We propose contextuality -- the degree to which an AI system autonomously accesses a user's accumulated knowledge capital -- as a dimension of AI-mediated inequality that complements, but is not reducible to, the Sharp et al. framework. We formalize the CAD with a probabilistic model grounded in the fan effect literature in cognitive psychology, demonstrating that manual context attachment leads to a combinatorial collapse in task-success probability as corpus size and task conjunctivity grow, while dynamic retrieval architectures are structurally insulated from this collapse. We analyze the technical basis of this divide in the Model Context Protocol (MCP) and retrieval-augmented generation (RAG) architectures, and examine its implications for knowledge-work stratification and AI platform governance.

Summary

Main Finding

The paper identifies the Context Access Divide (CAD): an interaction-level architectural inequality that determines whether an AI system autonomously retrieves a user’s personal/organizational knowledge corpus (dynamic context retrieval) or requires the user to manually attach context each query. CAD produces threshold, non‑linear effects for knowledge‑intensive tasks: under Manual Attachment Model (MAM) human memory and search limits (modeled using the fan effect) cause a combinatorial collapse in task‑success probability as corpus size (N) and conjunctive context requirements (k) grow. Dynamic retrieval architectures (Walled or Open DCRM) are structurally insulated from this collapse. The CAD therefore complements Sharp et al.’s availability/quality/quantity framework by adding a micro‑architectural dimension—contextuality—that aggregates into macro distributional consequences for labor, organizations, and platform power.

Key Points

Definition: Contextuality = degree to which an AI system autonomously accesses a user’s accumulated knowledge capital. The CAD is the distributional gap along this axis.
Three context‑supply architectures:
- Manual Attachment Model (MAM): user must find & attach context each turn.
- Walled Dynamic Context Retrieval (Walled DCRM): system retrieves across a single provider’s ecosystem.
- Open Dynamic Context Retrieval (Open DCRM): system retrieves across heterogeneous sources via open protocols (e.g., MCP + RAG).
Cognitive mechanism: human recall/identification limits (fan effect) mean per‑document recall probability falls with corpus size; when tasks require k conjunctively necessary documents, success probability under MAM approximates (p(N))^k and rapidly collapses as N and k increase.
Threshold/phase dynamics: partial context often yields qualitatively inferior outputs for conjunctive tasks (not linear interpolation); hence moving from MAM to any DCRM can produce discontinuous productivity jumps for certain task classes.
Platform implications:
- Walled DCRM relieves users inside ecosystem but creates platform lock‑in; users with cross‑ecosystem corpora may remain effectively in MAM.
- Open DCRM (enabled by MCP + agentic RAG) offers cross‑ecosystem relief but requires configuration and governance to be widely available.
Empirical grounding and adoption signals: MCP adoption grew rapidly after launch (examples in paper: server downloads ~100k→8M Nov 2024–Apr 2025; by Q1 2026 ~97M monthly SDK downloads; >10k MCP servers; independent census 17,468 MCP servers), showing rapid move toward architectures that can enable Open DCRM—though user experience depends on deployment/configuration.
Distributional effects: CAD systematically advantages organizations/individuals with resources to deploy DCRM (technical staff, integrated systems), amplifying existing stratification among knowledge workers and firms.

Data & Methods

Methods:
- Conceptual and theoretical analysis situating CAD within agentic inequality and digital‑divide literatures.
- Formal probabilistic model of human recall/attachment under MAM, grounded in the cognitive‑psychology fan effect (Anderson et al.). Model shows task success probability decays combinatorially with corpus size N and conjunctivity k.
- Technical architecture analysis comparing MAM, Walled DCRM, and Open DCRM; review of retrieval‑augmented generation (RAG) and Model Context Protocol (MCP) as enabling technologies.
- Use of published adoption/usage statistics for MCP and citations to empirical studies on heterogeneous AI productivity gains to motivate distributional claims.
Data cited (from paper): MCP server/download adoption trajectory (100k→8M from Nov 2024–Apr 2025; ~97M monthly SDK downloads by Q1 2026; >10k MCP servers; independent census 17,468 MCP servers). Prior empirical findings referenced: personal information management, fan effect literature, and heterogeneous productivity papers (e.g., Brynjolfsson et al., Dell’Acqua et al.).
Limitations noted in paper: primarily theoretical/formal; large‑scale causal empirical validation of CAD’s labor‑market effects remains to be done; real‑world configurations and privacy constraints complicate binary classification of user experience.

Implications for AI Economics

Productivity & nonlinearity: CAD implies non‑linear productivity gains from agent adoption. Firms/individuals crossing the DCRM threshold can realize outsized, discontinuous improvements on conjunctive knowledge tasks—producing tipping points in competitive advantage.
Complementarities and returns:
- Strong complementarity between organizational technical capacity (integration, engineering, governance) and AI productivity. Returns accrue more to firms that can implement Open/Walled DCRM at scale.
- Human capital effects: technical skill in configuring/curating connections and governance becomes a scarcifying complement.
Inequality & stratification:
- CAD amplifies within‑occupation heterogeneity (senior vs junior, well‑resourced vs solo practitioners) and between‑firm divergence (large incumbents vs small firms/startups).
- Potential for winner‑take‑most dynamics as integrated contextual AI raises switching costs and increases lock‑in.
Market structure & platform power:
- Walled DCRM increases incentives for ecosystem consolidation and vertical integration; Open DCRM mitigates this but its real‑world benefits depend on adoption, governance, and interoperability.
- Network effects: better contextualized agents improve firm productivity, attracting more users/data, reinforcing platform dominance.
Policy and measurement implications:
- New metrics: measure “contextuality” (e.g., availability of DCRM for a user, fraction of corpus reachable, corpus size distribution, task conjunctivity distributions).
- Interventions: promote open protocols/interoperability (standards like MCP), lower technical adoption barriers (subsidies, tooling), antitrust scrutiny of walled DCRM lock‑in, privacy frameworks that balance utility and data protection.
- Empirical needs: field experiments/randomized rollouts to quantify CAD productivity effects; mapping of corpus distributions across occupations; measurement of how often tasks are conjunctive and sensitive to omitted context.
Research and business strategy:
- Firms should assess whether key tasks are conjunctively context‑dependent; investments in DCRM integration can yield outsized returns for such tasks.
- Economists and policymakers should treat contextuality as a distinct dimension when modeling AI’s distributional and market effects, alongside availability, quality, and quantity.

Overall, the paper argues that interaction‑level architecture (who fetches the context) is a critical economic variable: it creates thresholds in AI usefulness that reshape productivity, inequality, and platform dynamics in ways that person‑level access measures alone cannot capture.

Assessment

Paper Typetheoretical Evidence Strengthn/a — This is a conceptual and formal/theoretical paper that develops a probabilistic model and technical analysis rather than providing empirical or experimental causal evidence. Methods Rigormedium — The paper provides a formal probabilistic model grounded in established cognitive psychology (fan effect) and a technical analysis of system architectures (MCP, RAG), which is methodologically careful for a theory paper; however, it relies on simplifying assumptions and lacks empirical calibration or robustness checks against real-world retrieval and user-behavior data. SampleNo empirical sample; the paper uses a probabilistic formal model (inspired by the fan effect literature) and analytic comparisons of two interaction regimes (Manual Attachment vs Dynamic Context Retrieval) and of system protocols (Model Context Protocol, retrieval-augmented generation); discussion and normative implications are supported by theoretical derivations and architectural analysis rather than data. Themesinequality productivity human_ai_collab GeneralizabilityNo empirical validation — conclusions depend on model assumptions and parameter choices that are not calibrated to real-world corpora or user behavior, Focus on knowledge-intensive workers with large personal corpora; may not apply to routine or externally-sourced tasks, Architecture-specific — results depend on the defined properties of MCP and RAG implementations and may not hold for hybrid or emerging designs, Ignores institutional constraints like privacy, legal access, or IT policies that affect practical context access, Assumes particular combinatorial structure of tasks (task conjunctivity) and document relevance that may vary across domains

Claims (9)

Claim	Direction	Outcome	Confidence & Evidence	Details
Sharp et al. (2025) introduce "agentic inequality" as a framework for analyzing disparities in access to AI agents across three dimensions: availability, quality, and quantity. Governance And Regulation	mixed	existence/definition of a conceptual framework (agentic inequality with three dimensions)	Reading fidelity high Study strength medium	not reported 0.12
There exists a finer-grained divide at the level of individual interaction — the Context Access Divide (CAD) — whereby two users with nominally equivalent agent access may experience qualitatively different AI utility depending on whether the system can autonomously retrieve context (Dynamic Context Retrieval) or requires manual document attachment (Manual Attachment). Output Quality	negative	AI utility experienced by the user (qualitative usefulness)	Reading fidelity high Study strength speculative	not reported 0.02
For knowledge-intensive workers whose intellectual capital spans tens of thousands of files, the CAD constitutes a qualitative threshold in AI usefulness: below it, the cognitive burden of context curation falls on the human, reproducing the inefficiencies AI is meant to eliminate. Task Completion Time	negative	cognitive burden / effective AI usefulness for knowledge work	Reading fidelity high Study strength medium	not reported 0.12
The CAD is formalized with a probabilistic model grounded in the fan effect literature in cognitive psychology. Research Productivity	mixed	formal modeling of context-access effects (theoretical task-success dynamics)	Reading fidelity high Study strength medium	not reported 0.12
The probabilistic model demonstrates that manual context attachment leads to a combinatorial collapse in task-success probability as corpus size and task conjunctivity grow. Output Quality	negative	task-success probability	Reading fidelity high Study strength medium	not reported 0.12
Dynamic retrieval architectures are structurally insulated from the combinatorial collapse in task-success probability that afflicts manual attachment approaches. Output Quality	positive	task-success probability under dynamic retrieval	Reading fidelity high Study strength medium	not reported 0.12
The paper analyzes the technical basis of the Context Access Divide in Model Context Protocol (MCP) and retrieval-augmented generation (RAG) architectures. Adoption Rate	mixed	architectural sources of context-access differences	Reading fidelity high Study strength medium	not reported 0.12
The CAD has implications for knowledge-work stratification and AI platform governance. Inequality	negative	knowledge-work stratification / governance outcomes	Reading fidelity high Study strength speculative	not reported 0.02
We propose 'contextuality' — the degree to which an AI system autonomously accesses a user's accumulated knowledge capital — as a dimension of AI-mediated inequality that complements, but is not reducible to, the Sharp et al. framework. Governance And Regulation	mixed	conceptual dimension (contextuality) as explanatory variable for inequality in AI utility	Reading fidelity high Study strength speculative	not reported 0.02