Notes · updated 2026-06-27
AX and Design: A Cross-Comparison of Academic and Industry Perspectives
This note provides a cross-sectional comparison of the intersection between AX (Agentic Experience) and design, drawing on 36 academic sources (agentic-experience-literature) and 73 industry sources (agentic-experience-industry), totaling 109 items.
Both sides address the same problems, but differ in vocabulary and focus. Industry articulates design principles and market opportunities under the new term “AX,” while academia accumulates empirical research within the framework of “human-agent interaction.” This note identifies the points of convergence between the two, as well as the points visible to only one side.
Related: design-agent-tools-landscape-2026 (tool landscape) / mcp-design-agent-integration (MCP integration) / the roundtable on delivering multi-agent value (value of multi-agent systems) / vibe-coding-design-production (UI generation agents).
Three Points of Convergence Between Industry and Academia
1. Agency Distribution as the Central Problem
Industry: Every major vendor has adopted “agents execute, humans decide” as a design principle. Microsoft’s “Built for Intent (augmentation, not replacement),” Anthropic’s five principles, OpenAI’s “always under control,” and Figma’s “focus on goals, not on managing AI” all express the same structure in different words. Harrison Chase’s concept of “leash length” (v0 = short, Cursor = middle, Devin = long) offers a practical framework for describing intervention granularity along a continuum.
Academia: Zhang et al. (2025) systematized agency distribution patterns from 134 papers across CHI/CSCW/UIST. Adam et al. (2024) experimentally demonstrated that AI-driven task assignment increases users’ sense of “self-threat” and reduces delegation acceptance. Houtti et al.’s (2025) OAI framework (Observe, Ask, Intervene) empirically showed that users prefer “question-first” approaches.
Cross-comparison: Industry declares “humans decide” as a principle, but it is academia that provides the specifics of how to embed decision-making into design. Zhou et al.’s (2026) finding that “intermediate checkpoints are optimal” offers quantitative guidance for Chase’s leash length concept, addressing the question of which length is actually best.
2. Transparency Alone Does Not Resolve Trust
Industry: Anthropic found that 93% of users approve permission prompts, leading them to conclude that per-action approval is not functioning as effective oversight. They subsequently shifted their design to a bounded-autonomy model. Prithvi Rajasekaran reported a tendency for agents to praise their own outputs even when quality is low.
Academia: He et al. (2025) demonstrated that presenting high-quality plans improves trust and performance, but also documented the risk of overreliance on plausible-looking plans. Grunde-McLaughlin et al. (2026) reported that improving oversight UIs yielded only limited improvements in error detection accuracy. Yu et al. (2025) showed that process transparency (exposing chain-of-thought reasoning) significantly improves trust and satisfaction, though this is a separate issue from whether errors are actually caught.
Cross-comparison: Industry’s “93% approval rate” and academia’s “oversight UI improvements yield limited accuracy gains” illuminate different facets of the same problem. Increasing transparency is a necessary condition for trust but not a sufficient one. Industry has responded by shifting to “bounded autonomy” in design, but academia does not yet have an answer to how those boundaries should be designed.
3. Context Sets the Ceiling for Accuracy
Industry: Harrison Chase stated definitively that “both success and failure come down to context.” Figma observed that “asking AI to generate code without design context is like having a new hire write production code before onboarding.” NNg defined “context architecture” as a new design discipline. Google’s DESIGN.md and Apple’s App Intent Domains represent implementations of a declarative approach to structuring context.
Academia: In Naik et al.’s (2025) developer interviews on multi-agent GenAI systems, “complexity management” was extracted as the top challenge. Goyal et al.’s (2024) six dimensions of human-agent alignment made the multidimensionality of context explicit.
Cross-comparison: It is notable that industry itself acknowledges the structural constraint that “the maturity of a design system sets the ceiling for agent output accuracy,” a position consistent with academic findings (design systems as the substrate layer for AI). However, academia still lacks quantitative evidence on which types of context produce what degree of accuracy improvement.
Three Points That Industry Omits but Academia Addresses
1. Junior Designers’ Learning Pathways
Industry promotes “lowering the floor” (reducing barriers to entry) but does not address how juniors learn design judgment once that floor has been lowered.
Academia has documented this gap. Li et al. (2024), in a 20-participant interview study, reported that experienced designers view GenAI as an assistant, while junior designers express concern about skill atrophy. Luo et al.’s (2025) systematic literature review of 83 papers identified the structural risk that juniors plateau at prompt competency without developing deeper design skills.
The academic grounding for the “the better it teaches, the sooner they churn” dilemma discussed in the roundtable on delivering multi-agent value can be found here. It is reasonable to infer that industry avoids this topic because junior learning pathways may conflict with the value proposition of their own services.
3. Structural Limitations of Oversight UIs
Industry design principles invoke “transparency,” but the structural limitations of oversight UIs, specifically the finding that the problem of users missing errors cannot be resolved through UI improvements alone, have not been deeply examined on the industry side.
Grunde-McLaughlin et al.’s (2026) finding (that oversight UI improvements reduce error detection time but yield only limited accuracy gains) and Shome et al.’s (2026) finding (misalignment between the mental models of 102 commercial agents and their users) cast doubt on the implicit industry assumption that “improving transparency will resolve trust problems.”
Three Points That Academia Omits but Industry Observes
1. Cost Structure
Academia discusses agent design and UX but does not treat inference costs and the sustainability of business models as research subjects.
Anthropic’s multi-agent 15x token consumption, harness 20x cost, and Figma’s gross margin decline from 92% to 86% demonstrate that AX design takes place within economic constraints. These constraints directly affect design decisions, as cost-prohibitive designs will not be adopted.
2. Agent-Washing
Gartner’s estimate that “of thousands of vendors, only roughly 130 are substantive” indicates a phenomenon in which existing products are being rebranded under the “AX” label. Academia has not directly studied the structural dynamics of this buzzword inflation.
3. Declarative Primitive Architecture
Apple’s App Intents / Dynamic Profiles, Google’s A2UI / DESIGN.md, and Anthropic’s workflows/agents distinction represent architectural choices that implement AX design as “defined by data (declarations) rather than executable code.” While academia proposes abstract frameworks, the impact of the declarative versus imperative architectural choice on UX remains unstudied.
Academic Standing of AX
As of June 2026, virtually no peer-reviewed papers use the term “Agentic Experience” (agentic-experience-literature). The academic community studies the same problems under the headings of human-agent interaction, AI delegation, and agentic AI UX.
The CHI 2026 workshop “Agentic Automation Experiences” (DOI: 10.1145/3772363.3778732) represents an academic germination of AX, but whether Maeda’s terminology gains academic traction remains to be seen. On the industry side, Salesforce has officially adopted “AX” and published a definition. NNg does not use the term “AX,” instead employing its own conceptual apparatus (“treat AI agents as users,” “context architecture”).
Whether the industry term (AX) and the academic framework (human-agent interaction) will converge or diverge going forward can be assessed from trends emerging at CHI 2027 and beyond.
Comparison Table
| Issue | Industry Position | Academic Position | Divergence / Convergence |
|---|---|---|---|
| Agency distribution | Declared as a design principle | Experimentally exploring optimal parameters | Convergence (academia provides specifics) |
| Transparency | Espoused as a principle | Limitations empirically demonstrated (overreliance, accuracy ceiling) | Partial convergence (industry does not discuss limitations) |
| Context dependence | Acknowledged as a structural constraint | Conceptually identified (quantification lacking) | Convergence |
| Definition of AX | Fragmented across 4 lineages | Studied as HAI, not using “AX” | Divergence (at the terminological level) |
| Junior learning | Not addressed | Skill atrophy risk empirically documented | Industry blind spot |
| Exploration vs. efficiency | Juxtaposed (contradiction unacknowledged) | Inverse correlation identified | Industry blind spot |
| Structural limits of oversight | 93% approval rate prompted design change | UI improvements have limited impact on accuracy | Partial convergence (industry responds through implementation) |
| Cost structure | Empirical data disclosed | Not studied | Academic blind spot |
| Agent-washing | Gartner issues warnings | Not studied | Academic blind spot |
| Declarative architecture | Implemented by Apple/Google | Not studied | Academic blind spot |
References
For academic references, see agentic-experience-literature; for industry references, see agentic-experience-industry.