# UX of AI systems Topic radar · 1,243 documents · 2021-01-01–2026-09-23 · updated 2026-09-24 Canonical: https://slate.greyb.com/radar/ux-of-ai-systems/innovations ## In one minute ### The interface is shifting from a visual layer to a "context engine" that mediates non-deterministic outcomes Across the portfolio, R&D is pivoting away from static component libraries toward dynamic architectures where the interface is synthesized in real-time. This is evidenced by a heavy concentration in "Co-Agency" frameworks, which account for 69.7% of the total corpus (866 documents), signaling a move toward trust-mediated, multimodal interfaces [1]. This shift is supported by specialized funding rounds in 2025 and 2026, including $4 million for Thesys and £3.7 million for Heywa Labs, specifically targeting adaptive interface layers that prioritize psychological outcomes over traditional aesthetics [2]. **Why this is worth reading** The record shows a massive technical commitment to human-computer interaction (755 inventions published between 2021 and 2026) that is not yet matched by public market visibility. This suggests the company is building a quiet infrastructure for "interface empathy" that relies on real-time context engines rather than raw model accuracy to drive professional adoption. 1. [Co-Agency & Intent Discovery](/radar/ux-of-ai-systems/innovations?bucket=collaborative_hci_and_intent) 2. [medial.app](https://medial.app/news/thesys-bags-dollar4-mn-to-provide-platform-for-adaptive-ai-interfaces-c7c836963b591;web,https://www.startupmag.co.uk/funding/heywa-labs-2026-seed-funding/) ### The interface has become the primary site of technical defensibility, outstripping model performance A sharp divergence has emerged between technical invention and public discourse: while the portfolio contains 755 inventions published between 2021 and 2026 focused on human-computer interaction frameworks, this volume is met by only 111 items of public coverage. This suggests the company is building a quiet, proprietary infrastructure for "co-agency" that the market has not yet parsed. The concentration of work in trust-mediated multimodal sensing—accounting for 69.7% of the total corpus—indicates that pricing power is shifting from raw model accuracy to the interface's ability to manage professional skepticism and ensure safety in high-stakes fields like cardiology. ### The portfolio is pivoting from automation to trust-mediated "co-agency" frameworks The company has committed 69.7% of its total corpus (866 documents) to human-computer interaction frameworks that prioritize collaborative tooling and trust-mediated co-agency. This massive technical concentration, featuring 755 inventions published between 2021 and 2026, signals a strategic decision to solve for "interface empathy" through multimodal sensing rather than raw model accuracy. While the technical foundations are deep, a sharp disconnect exists between invention volume and public visibility—755 inventions are met by only 111 items of coverage—suggesting the company is building a quiet infrastructure that the market has not yet fully parsed. ### A "trust deficit" is triggering a move toward formalized UX certification and safety auditing There is a measurable divergence between technical model performance and human-centered design, leading to what the record describes as a usability crisis. To address this, institutional partnerships, such as those involving the University of Cambridge, are codifying design principles into formal certifications like the Certified Human-AI Interaction Designer (CHAID) to ensure consistency in non-deterministic systems. This professionalization is a direct response to "AI Anxiety" and the "IKEA Effect," where early-stage AI products fail due to organizational misalignment rather than poor code. ### Generative UI architectures are shifting R&D from visual design to context engineering The portfolio is moving away from manual prototyping toward "vibe-based" orchestration, where interfaces are synthesized in real-time based on user intent. This shift is evidenced by specialized seed funding—including $4 million for Thesys and £3.7 million for Heywa Labs—targeting adaptive interface layers. However, there is a conspicuous absence of inventions supporting "Invisible IoT" or ambient intelligence in the current window, despite market reports suggesting a pivot toward voice-less intelligence. ### Generative UI architectures are replacing static libraries despite a technical execution lag There is a transition underway toward dynamic "Generative UI" where interfaces are synthesized in real-time based on user intent. However, the evidence reveals a significant lag: market coverage outruns inventions by a factor of seven, and while reports suggest a pivot toward ambient "Invisible IoT", there are zero inventions supporting this in the current window. The real technical momentum is concentrated in generative construction, where interfaces are increasingly validated by synthetic user models rather than human templates. ### Autonomous "computer use" agents are marginalizing traditional GUI navigation A strategic pivot is underway from chat-based interfaces to autonomous agents that manipulate operating systems directly, effectively replacing application menus with Large User Interface Models (LUIMs). While 88% of AI budgets are increasing, only 42% of organizations have initiated the workflow changes needed to support these agents, creating a bottleneck that new tools like UXAgent—designed to simulate usability testing with LLM agents—are attempting to bridge. ### Autonomous "computer use" agents are positioned to bypass traditional application menus The portfolio is betting on a transition where Large User Interface Models (LUIMs) replace static menus, with agents navigating operating systems directly. This is a distinct strategic pivot from the "human-in-the-loop" collaboration seen in larger clusters, committing instead to a future where the agent is the primary navigator. Despite this vision, a critical gap exists: there is a lack of protected intellectual property for the underlying LUIMs, and while 88% of AI budgets are increasing, only 42% of organizations have initiated the necessary workflow redesigns. ## Overview The UX of AI landscape is undergoing a structural pivot from chat-based interfaces toward trust-mediated "co-agency" frameworks that prioritize multimodal sensing over raw model accuracy. While the industry is vocal about full autonomy, the technical record reveals a massive concentration in collaborative tooling designed to bridge the gap between machine logic and professional domain needs. This suggests a strategic bet that the primary competitive differentiator is no longer the model itself, but the "interface empathy" required to manage professional skepticism in high-stakes sectors like medicine and finance. 1. Technical R&D is concentrated in collaborative co-agency rather than full autonomy or intent discovery. While "Agentic & Autonomous Experience" generates significant market discourse, the physical invention record is dominated by human-computer interaction frameworks, which account for 866 documents—nearly 70% of the total corpus. The weight of this work is anchored by 755 inventions published between 2021 and 2026, signaling that the industry is building a quiet infrastructure for trust-mediated collaboration that the market has not yet fully parsed. This technical density in co-agency stands in sharp contrast to "Intent Discovery" and "Sensemaking," which remain largely conceptual and thin on implementation. 2. A significant "implementation gap" exists between public discourse and technical filings in trust and ethics. There is a sharp contradiction between the volume of ethical discourse and actual technical mechanisms to solve trust failures, with coverage items outnumbering inventions 61 to 11 in the Trust, Ethics & Responsible AI area. While academic and policy groups are developing taxonomies for "dark patterns" and "AI Anxiety," the lack of patented interaction mechanisms suggests that the industry is currently better at diagnosing trust deficits than engineering the solutions. This gap indicates that the first players to codify "Safety UX" into auditable technical standards will likely capture significant pricing power in regulated sectors. 3. The competitive moat is migrating from visual design to proprietary context engines and synthetic validation models. As layout generation becomes automated, R&D momentum is shifting toward the engineering of context engines that govern how AI interprets user needs before rendering a response. The reliance on synthetic user models for interface validation suggests a future where software usability is optimized for the mathematical averages of AI agents rather than human idiosyncrasies, potentially creating a new market for "Human-Certified UX." 4. Enterprise implementation is prioritizing "agentic oversight" layers over functional tool execution. Major enterprise players like Salesforce and ServiceNow are driving a shift toward interfaces that function as thin monitoring layers for autonomous background processes rather than tools for manual labor. Despite 19 items of commercial coverage regarding the elimination of the "toggle tax," the enterprise workflow segment shows 0 new inventions, suggesting that the architecture for unified interaction is currently being built on existing models rather than fundamental breakthroughs. This transition favors dominant "agentic hubs" that can guarantee deterministic outcomes from non-deterministic natural language inputs. ## Research ### Synthetic user models are replacing human testing in the design feedback loop The portfolio shows a transition from human-validated templates to interfaces validated by synthetic user models to accelerate the feedback loop. This is particularly visible in the automation of UI layouts, where R&D is moving toward "vibe-based" orchestration and real-time synthesis. ## Strategy ### A "trust deficit" is triggering a move toward formalized Safety UX certifications A sharp contradiction exists between the volume of discourse and technical implementation in responsible AI: coverage items outnumber inventions 61 to 11. The industry is currently better at diagnosing failures like "AI Anxiety" than it is at filing patents for technical solutions. In response, institutional partnerships are beginning to codify design principles into formal certifications—such as the Certified Human-AI Interaction Designer (CHAID) —to ensure consistency in non-deterministic systems. ## Areas - Co-Agency & Intent Discovery · 866 documents - Foundational Usability & Market Trends · 78 documents - Trust, Ethics & Responsible AI · 72 documents - Generative & Adaptive UI Architecture · 66 documents - Agentic & Autonomous Experience · 36 documents - Enterprise & Vertical Implementation · 74 documents