UX Roundup: UX Benchmark | Impact of AI Mentions | Character Design Workflow | AI Maturity | UX Hero: Stu Card | Microsoft FDE

Summary: UX benchmark for AI | Being mentioned in AI answers drives subsequent user behavior | Cheaper character design workflow for new video | AI maturity model | My hero: Dr. Stuart K. Card of Xerox PARC | Microsoft FDE

UX Roundup for July 6, 2026 (GPT-Images-2)

USA 250

Happy birthday, USA!

With apologies to Emanuel Leutze and John Trumbull for tigerizing their paintings, “Washington Crossing the Delaware” and “Declaration of Independence,” respectively. (GPT-Images-2)

UX Benchmark for AI

The tech industry measures AI progress with capability benchmarks: pass the bar exam, solve logic puzzles, or write sophisticated code. But as commercial AI assistants scale to millions of daily users, a glaring disconnect has emerged. A model’s raw cognitive intelligence doesn’t translate into usability. A chatbot that can code an entire application in seconds can still be infuriating to converse with.

The standard AI benchmarks, or even AI achievements in winning Math Olympiad Gold medals, are irrelevant for AI’s usefulness in solving real problems for users and delivering solutions with high usability. We need new measures of AI usefulness and usability. (GPT-Images-2)

A new research report, UXBench: Benchmarking User Experience in AI Assistants, from the Hong Kong Polytechnic University and Tencent (June 2026), confronts this reality. The study evaluates 26 frontier language models, including GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro, and DeepSeek V4 Pro, on a metric the leaderboards ignore: human-perceived interaction quality.

UXBench is important because it tries to evaluate AI assistants from the user’s point of view rather than from abstract model accuracy alone. The benchmark is grounded in real user feedback from AI assistant interactions and tests whether models can predict good and bad user experience across scenarios, domains, and failure patterns. The study finds that even frontier models struggle to identify bad experiences reliably, with some models mapping many bad cases into neutral categories. This matters because product teams increasingly use AI judges and reward models to optimize assistants, but those judges may systematically under-detect user dissatisfaction.

The Thumbs-Up Illusion: Behavior Is the Real Ground Truth

Historically, evaluating AI dialogue has relied on synthetic preferences or sterile, rubric-based judgments by LLMs themselves. The most persistent bottleneck in evaluating genuine AI user experience is the reliance on explicit feedback mechanisms. As the UXBench report highlights, users rarely click the “thumbs up” or “thumbs down” buttons. Out of 400,000 randomly sampled dialogue turns, users provided explicit feedback a mere 0.40% of the time.

Users rarely give AI explicit feedback on the quality of its work. Instead, they either abandon the AI or try again with a different prompt. (GPT-Images-2)

If users don’t explicitly tell the AI they’re frustrated, how do developers know they are? The researchers solved this by extracting implicit behavioral signals. When a user receives a poor response, they don’t usually click “dislike”; instead, they skim the text, rapidly submit a new prompt, rephrase their question in frustration, or explicitly complain. By calculating the “gap ratio” (the temporal discrepancy between a generated response and the user’s next input) and tracking re-prompting behaviors, the researchers built a dataset of 7,400 test cases that captures real human friction from over 70,000 interaction logs.

Users aren’t abandoning AI because of factual hallucinations; they’re leaving because of interaction friction. The top failure dimension, accounting for over 34% of bad experiences, is simply “verbosity and redundancy.”

AI often overwhelms users with too much information. (GPT-Images-2)

The 3 UXBench Tests: Judge, Generate, Recover

To evaluate how well models handle the human element, UXBench subjects the 26 models to three distinct tasks:

1. UX Judge (Predicting Satisfaction): Can an AI look at a dialogue history and correctly guess if the user is happy? The results here are concerning. The study reveals a strong positive bias across all models. Because LLMs are inherently trained to favor AI-generated text, they act as sycophants, assuming almost all responses are excellent. Even the best zero-shot model, Claude Opus 4.7, missed 38.5% of bad interactions. They systematically fail to detect subtle experiential failures like emotional insensitivity or conversational redundancy.

It’s common to judge the quality of AI responses by having another AI judge. This doesn’t work because AI has a strong bias toward liking AI output.  (GPT-Images-2)

2. UX Eval (Generating Satisfying Responses): When given a prompt that previously resulted in a bad user experience, can the model generate a better one? The highest-performing model, Gemini 3.1 Pro, achieved only a 57.1% success rate. Interestingly, the report highlights a weak scaling trend. In traditional AI benchmarks, making a model larger yields massive performance improvements. In UXBench, newer models are only slightly better than their predecessors, possibly because AI labs have recently been chasing performance gains in software development rather than UX skills.    

AI models are getting better at improving their user experience as they scale, but the improvement curve for each family is less impressive than their progress in other areas prioritized higher by AI labs, such as software coding. That said, the “bitter lesson” holds even for UX, which is mostly being ignored: bigger models do better.

3. UX Recovery (Bouncing Back from Complaints): This is the ultimate stress test. When a user explicitly complains (“You're not answering my question!”), can the AI fix the relationship? Almost universally, the models failed, with success rates hovering around a dismal 12%. The researchers found that models use wildly different strategies. Weaker models defensively diagnose the error to explain why they were wrong. Better-performing models lean into social repair, using genuine apologies, humor, and immediate corrections to rebuild trust.

The Generative Reward Model Solution

Because off-the-shelf LLMs exhibit strong self-preference bias (GPT models prefer GPT outputs; Gemini prefers Gemini outputs) and fail at UX judgment, the researchers trained a specialized Generative Reward Model (GRM) purely on real user feedback. This point-wise model far outperformed frontier LLMs in predicting actual human satisfaction, proving that UX evaluation is a learnable capability that requires training on behavioral data.

Takeaways for UX Professionals Designing AI Products

For UX leaders, researchers, and product designers working on AI-native applications, the UXBench report is a mandate to put human-centered design back at the center of AI product development.

• Stop Relying on “LLM-as-a-Judge” for UX: The report proves that frontier models are highly biased judges. In the long term, we need specialized Generative Reward Models (GRMs) trained purely on actual behavioral telemetry and human frustration signals, rather than relying on generic LLM API calls to evaluate product quality. Evaluating UX with an out-of-the-box model will yield artificially inflated scores.

• Redefine Your Success Metrics: Stop relying solely on explicit user ratings, because you won’t get enough. Implement behavioral telemetry. Track the skimming pattern, the rate of prompt-rephrasing, and session abandonment. These implicit signals are better measures of your product’s usability. If users are rapidly firing follow-up questions, your AI is likely failing to satisfy their intent, regardless of how accurate the factual content might be.

• Tone, Formatting, and Verbosity are Core UX Features: Users make split-second judgments based on the first sentence of an AI’s response. Engineer your AI product to eliminate meta-commentary (e.g., “Certainly! I can help you with that.”). Responses must immediately deliver value in the opening line. Verbosity is the leading cause of user dissatisfaction; UX writers must tune system prompts to prioritize directness and appropriate brevity.

• Design for Social Repair, Not Error Correction: When an AI fails, it shouldn’t act like a debugging tool. UX professionals must design conversational recovery protocols. If a user expresses frustration, the AI should prioritize social repair (validation, direct fixes, humor) before attempting to pedantically explain its reasoning errors. Users want their problems solved, not an essay on why the AI hallucinated.

4 Ways UXBench Should Grow

 

UXBench is a good first step, but human–AI interaction changes fast. Here’s how the benchmark should grow:

Expand to Multimodal and Voice Interactions

The current UXBench evaluates text-based dialogue only. However, the future of AI assistants is voice-first and multimodal. A future iteration of UXBench must measure the UX of voice latency, interruption handling (how well an AI deals with conversational “barge-ins”), and the emotional prosody of synthesized speech. A text response that reads well might sound condescending or unnatural when spoken aloud. Verbosity in text is annoying; verbosity in voice is excruciating.

Speak too much, and you will flatten your audience, whether humans or foxes. (GPT-Images-2)

Measure the Evolving Baseline of Expectations

The researchers touch upon a vital psychological phenomenon: as AI models become more capable, user expectations scale accordingly. What was considered a “mind-blowing” interaction in 2024 is considered slow and clunky in 2026. Future versions of UXBench should incorporate longitudinal tracking to establish how the threshold for satisfaction shifts over time as humans adapt to better AI, ensuring the benchmark never becomes stagnant.

Introduce Relational and Contextual Baselines

Currently, UXBench judges responses in a vacuum. But a great user experience is personalized and built over time. UXBench should next benchmark an AI’s ability to maintain its relationship with each individual user. This involves evaluating an AI’s ability to remember a user’s preferences across weeks of interaction and adapt its persona on the fly based on user history, inferred cognitive level, and emotional state.

Cross-Cultural and Demographic Localization

A “good” user experience is highly subjective. A concise, blunt response might be viewed as efficient in one culture but downright rude in another. Future expansions should introduce cultural localization metrics to benchmark how well models adapt their conversational strategies to different regional interaction norms and demographic profiles.

The same experience can be received very differently by users from different cultures and backgrounds. (GPT-Images-2)

AI labs optimize what benchmarks measure. Make the benchmarks measure UX, and the industry will finally close the gap between impressive intelligence and everyday usefulness.

The Hidden Funnel: How AI Mentions Drive Open-Web Commerce

If you ask ChatGPT, “What running watch should I buy?” and it answers "Garmin, Coros, or Polar," what happens next? Most likely, you open a new tab, search for Garmin on Google, and browse a product page. On your company’s analytics dashboard, this looks like the sale was driven by organic search, even though the AI recommendation was the true catalyst.

A new study by Michael Iannelli and Alan Ai documents this hidden consumer journey. In their paper, “From Prompt to Purchase: How AI Brand Recommendations Move Consumers on the Open Web,” the researchers linked an opt-in clickstream panel with the same users’ ChatGPT, Claude, and Gemini chat logs. Their findings quantify how being mentioned in AI answers influences downstream open-web behavior, revealing that conversational AI is a powerful, unlogged engine for top-of-funnel customer acquisition.

We must observe the full funnel to understand the value of AI mentions. The new study did exactly that by benefiting from users who had agreed to have their entire clickstream captured. (GPT-Images-2)

The Clean Acquisition Effect

To isolate the true impact of AI brand mentions, researchers filtered out existing customers whose buying journey was already underway and incidental name-drops (like casually mentioning “my Netflix download”). They focused entirely on observably-unengaged users: people with no recent search, own-site, or retail activity for a brand.

The researchers limited their study magnet to only consider users who were not already engaged with a brand. People who had recently purchased or otherwise engaged with the brand were excluded. (GPT-Images-2)

When an AI assistant explicitly recommends a brand to these users, their behavior shifts over the following 7 days. Search (Googling the brand) increases by 4.3 percentage points. Visiting the brand’s website rises by 2.4 percentage points, and retail consideration (visiting a brand-specific page on a third-party retailer) jumps by 1.0 percentage point.

These lifts represent pure acquisition-like reach. The AI’s attitude toward a brand matters: an explicit recommendation moves consumer behavior 2–3 times more than a neutral, incidental mention.

It is worth much more when an AI actively recommends you than when it simply mentions you. (GPT-Images-2)

A Search-Mediated Blind Spot

The most useful finding: how users navigate after receiving an AI recommendation. Direct click-through rates from chatbot answers are negligible. (That’s why we rarely see referrals from AI in our website analytics.) Instead, the journey is strongly search-anchored. Users read a brand name they didn’t initially type, open a search engine, and navigate to the brand’s site or a retailer. Because this process involves an intermediate search, traditional last-click attribution models miss the AI’s role, misallocating credit to Google or direct traffic.

We must consider the full journey that brought a prospective buyer to our website, not just the last click, which is provided by our site analytics. Increasingly, the journey starts with AI answers. (GPT-Images-2)

The study also dispels several assumptions. Unlike traditional SEO ranking algorithms, the position of the brand in the AI’s response doesn’t matter. Being mentioned first, second, or third yields the exact same behavioral lift. This is a revolutionary change in user behavior, compared to the tyranny of needing to rank at the top of the SERP (search engine results page).

AI mentions are like the pies in this cartoon: it doesn’t matter whether they come first, second, or third. This user behavior reverses 30 years of prioritizing top search engine results above all else, making it a surprising finding. (GPT-Images-2)

Takeaways for Internet Strategy

As AI assistants absorb more information-seeking traffic, businesses must adapt to this hidden funnel.

• Rethink Web Attribution: Last-click attribution leaves you blind to a major acquisition channel. Strategists must accept that an increasing volume of “organic search” traffic is actually AI-prompted. Marketing teams should incorporate GEO metrics and incrementality testing to evaluate AI’s true impact on the funnel.

• Prioritize AI Visibility: Because AI recommendations drive measurable sales pathways for non-customers, brand presence in AI training data and Retrieval-Augmented Generation (RAG) pipelines is becoming as critical as traditional search engine ranking. GEO (Generative Engine Optimization) is essential.

• Design for the Tab-Switch Journey: Stop obsessing over getting chatbots to display clickable links. Users largely ignore them. Assume the user will read your brand name and Google it. UX professionals must ensure branded search results and landing pages align with the specific phrasing and context AI uses to describe products.

• Omnichannel Consistency: The study demonstrates that users view brand sites and third-party retailers as parallel destinations. Your UX must be equally good whether the user lands on your flagship site to research or a retail partner's page to buy.

Landing pages often employ overblown design that doesn’t echo the plainspoken benefits-oriented language in the AI mentions that drove people to visit. The jarring disconnect may make them leave. (GPT-Images-2)

Cheaper Character Design Workflow for New Video

I made a new music video about Dark Design (YouTube, 3 min.), this time as a country song, with visuals driven by B-roll animation to illustrate various dark design patterns. However, Suno (still the best music model) likes to insert long instrumental segments in the songs, so I needed video material to cover these parts of the music that weren’t about specific design patterns.

New country music song about Dark Design. (GPT-Images-2)

For videos that feature a singing avatar (like my blues rock song about Dark Design), I treat the instrumentals as dance breaks and animate the avatar dancing, which usually makes for entertaining clips. But I have been disappointed with the recent progress in avatar animation, which is why my newest music video doesn’t feature an on-screen singer. Instead, I’m showing the band playing the instrumental segments.

Thus, the need for a band design.

I first designed 34 different robot bands with GPT-Images-2. Here are 4 of them:

Alternate country music robot band designs. (GPT-Images-2)

I then selected the 5 designs I liked best and used Seedance 2 Mini to make short animations of each. (You can watch this 1-minute test reel on my Instagram.) Seedance Mini is a fast and cheap video model that approximates the full Seedance 2.0 I used for the final video, at the cost of slightly lower quality in resolution, rendering, and cinematography. While I wouldn’t use Mini for a published video, it was great for this stage of my workflow: seeing the bands actually play music and move on stage helped me decide which design to use in my music video.

Multi-step workflow for video design: create many visual ideas with an image model (I use GPT-Images-2, including for this comic strip), then animate the best with a cheap video model to see how they move, and finally render your best choice in 4K with the expensive video model.

As mentioned, this new music video makes extensive use of B-roll animations. For several of these, I also used Seedance Mini to make draft versions when I had multiple competing ideas. It’s much easier to choose the concept that works best with a piece of music when you see it animated and hear it with the correct soundtrack (which I easily dubbed in CapCut).

After selecting my preferred B-roll idea, I then made it in 4K resolution with the full Seedance 2.0 model. Even a short 6-second clip takes about 15 minutes to render in Seedance 2.0, so it would disrupt my creative workflow too much to wait to see several speculative ideas rendered. The full model is also 9x as expensive as the Mini model, making it prohibitively costly to create large amounts of throwaway footage. I spent about $200 in AI credits to make my 3-minute music video in 4K resolution, and the cost would probably have been closer to $500 if I had not employed the Mini shortcut for my review footage. (The benefit of only using the full model would have been that if I reviewed a draft and liked it, then I could have used that clip immediately instead of having to wait for it to be redone in high quality.)

Frontier AI models are slow and expensive, especially for video creation. Smaller models allow you to explore more ideas faster and cheaper. (GPT-Images-2)

Seedance 2.0 recently added 4K native video generation, which increases costs and response time, but removes the need for an extra upscaling step at the end of video editing. Also, the video quality is higher when generated natively rather than upscaled from lower-quality footage. Topaz is brilliant at upscaling (see, for example, the fur of a gorilla struggling with fat-finger problems when using a smartphone in my video about the 10 worst UI annoyances). But it always has to guess at what details to add when the original content isn’t sharp enough. Native 4K simply looks better. In my new country song, notice how the strings vibrate as the robot plays the upright bass.

Despite all my praise for Seedance 2.0 as the best current video model (especially when used to generate native 4K output), it’s very far from perfect, which you can easily see in this latest video. While the upright bass is animated nicely, the piano is animated poorly: the keys don’t depress as the pianist plays.

The two-step workflow of using a cheap AI model for review drafts, followed by an expensive AI model to generate production output, saves time and money, so I recommend it for now.

The main downside is that even the “fast” Mini model is too slow for discovery-based creation. I can’t freely navigate the latent design space when each attempt takes several minutes. Response time drives usability, and 10 seconds is the longest acceptable delay between expressing initial intent and seeing the result, in order to redirect my intent according to my assessment of that result.

The pursuit of changing intent in the latent space of AI creation requires AI to be much faster than it is now, especially for video creation. (GPT-Images-2)

AI Maturity Model

The Software Engineering Institute (SEI) at Carnegie Mellon University (CMU) has released a detailed AI adoption maturity model (63-page PDF). SEI was the originator of the most widely used maturity model for software development, so it’s interesting to see how it treats AI maturity.

The framework shifts the focus from deploying technology fast to building trustworthy, well-governed organizational capabilities in a disciplined way. Understanding it will help you shape the corporate ecosystem in which digital products get built.

The 5 Levels of AI Adoption Maturity

The SEI model treats maturity as a strategic progression aligned with business outcomes and risk tolerance. Climbing the levels takes disciplined, institutionalized execution.

The 5 levels of AI maturity in the new SEI model. (GPT-Images-2)

Level 1. Exploratory AI: The organization focuses on learning and discovery: ad-hoc experimentation in controlled sandboxes to assess technical feasibility and identify viable use cases. The primary goal is to validate concepts, scope pilot projects, and begin building foundational AI literacy within the workforce.

Level 2. Implemented AI: Moving beyond the sandbox, organizations at this level have deployed select AI-enabled systems or workflows into production. To manage the associated risks, foundational practices are established, including basic data lifecycle management, AI model security, and initial risk protocols. A formal AI strategy begins to take shape.

Level 3. Aligned AI: This is the turning point: AI initiatives map directly to core business objectives, and the focus shifts to workflow re-engineering. Organizations implement rigorous measurement practices to track ROI, while enforcing human-in-the-lead automation and system transparency to ensure AI applications are managed safely and consistently.

Level 4. Scaled AI: At this stage, AI is no longer siloed; it’s integrated across the enterprise with reliable, predictable performance. The organization can rapidly apply AI to new use cases and scale solutions efficiently. This is supported by shared infrastructure, robust supply chain management, and standardized deployment practices that optimize resources and minimize operational friction.

Level 5. Future Ready AI: The pinnacle of the model describes an AI-native enterprise. A Future Ready organization utilizes predictive analytics and causal modeling to anticipate technological shifts. It can rapidly absorb cutting-edge innovations—such as autonomous agentic systems—without compromising governance or stability. Innovation is continuous, and the workforce keeps pace with AI as it changes.

The 8 Core Capability Dimensions

To progress through these maturity levels, the SEI model defines 8 core dimensions, structurally divided into two overarching categories: Organizational Change and AI Lifecycle Engineering.

The 8 dimensions for assessing your AI maturity. (GPT-Images-2)

Organizational Change Dimensions:

1. Organizational Strategy: Focuses on positioning the enterprise to deliver business value through AI. This includes developing a formal AI strategy, cultivating strategic third-party partnerships, calibrating organizational structures, and engaging in proactive future-ready planning.

2. Workforce and Culture: Acknowledges that AI transformation is a human endeavor as much as a technical one. It emphasizes building an AI-literate workforce, reskilling talent, and fostering an open, collaborative culture that embraces continuous learning and AI integration.

3. Workflow Re-Engineering: The engine of true business innovation. It moves organizations beyond simply layering AI over old processes. Key capabilities include structured experimentation, business workflow innovation, human-in-the-lead automation, and rigorous measurement and analysis.

4. Risk and Governance: Ensures that AI is operationalized safely and ethically. This dimension mandates robust risk management frameworks, strict policy compliance, and the enforcement of Responsible AI principles to mitigate issues like algorithmic bias and privacy violations.

AI Lifecycle Engineering Dimensions:

5. Data: The bedrock of reliable AI. This dimension requires disciplined data lifecycle management and data quality assurance, ensuring both structured and unstructured data are secure, relevant, tracked for provenance, and properly governed.

6. Engineering: Translates AI concepts into production-ready architecture. It encompasses AI architecting, specialized test and evaluation (including bias and hallucination testing), legacy system integration, and practices for system transparency and explainability.

7. Operations: Ensures the long-term sustainability of deployed AI. It demands AI model management (tracking version control and data drift), model and agent security, and continuous telemetry monitoring to catch performance degradation.

8. Technology Ecosystem: Manages the foundational infrastructure required to scale. It covers technology lifecycle management, the complex supply chain of third-party AI vendors, and the architectural capacity for enterprise-wide deployment.

Comparing the SEI Model with the Capability Maturity Model for AI in Design

Compare the SEI AI Adoption Maturity Model with my Capability Maturity Model for AI in Design, and you see two complementary frameworks operating at different altitudes.

The SEI model is a macro-level, top-down systems engineering framework. It approaches AI as a complex organizational ecosystem, heavily focused on cybersecurity, data provenance, risk governance, and enterprise architecture. By contrast, the design maturity model is a micro-level, practitioner-focused framework. It charts the evolution of a design team’s relationship with AI, progressing from Stage 1 (No AI), to Stage 2 (Individual/Ad-hoc use as a brainstorming tool), to Stage 3 (Formal/Team-wide integration), and ultimately to Stage 4 (Fully AI-driven design processes).

Despite the difference in scope, their evolutionary arcs are similar. SEI’s Level 1 (“Exploratory AI”) maps directly to my Stage 2, where isolated, ad-hoc experimentation reigns. As organizations mature, SEI’s emphasis on “Workflow Re-Engineering” and “Aligned AI” mirrors the design model Stage 3 goal of formally standardizing AI tools across the design department and redefining how the actual work gets done.

Most importantly, the SEI model explicitly validates the UX discipline’s core mission. It mandates “Human-in-the-Lead Automation” and “Transparency and Explainability” as essential engineering capabilities. Both models agree that enterprise AI can only succeed at scale if it’s usable, understandable, and operating under human oversight. The SEI model builds the backend governance and data pipelines to make this possible, while the UX Tigers maturity model ensures the frontend application actually delivers usability, creativity, and human value.

From Assessment to Actionable Roadmap

The point of the SEI model is to generate an actionable roadmap through a 6-step process, not to hang a static maturity score on the wall. The SEI report emphasizes a core philosophy echoed in my infographic: advance intentionally, not automatically. Organizations shouldn’t blindly race toward the highest maturity tier across all capabilities. Instead, leaders must define future target states tailored to their unique business goals, risk appetite, and regulatory constraints (Steps 2 and 3).

The 6-step process for using the maturity model to create a roadmap for your improved use of AI. (GPT-Images-2)

The assessment process (Step 1) is rigorous and evidence-based. To prevent enterprises from masking vulnerabilities behind isolated successes, the SEI model employs a strict scoring mechanism where a dimension is only considered as mature as its lowest-ranking capability area. Furthermore, data is triangulated through expert-led stakeholder workshops, artifact analysis, and system telemetry rather than relying on automated self-assessments.

Once the baseline and target states are established, organizations measure capability gaps and uncover critical inter-capability dependencies (Step 4). Finally, by plotting these initiatives on an “Effort vs. Value” matrix in Step 5, teams can prioritize what to do now vs. later (Step 6). This sequencing secures high-ROI quick wins while properly resourcing the foundational work needed for heavier AI investments.

Takeaways for UX Leaders

The SEI AI Adoption Maturity Model offers a clear view of enterprise leadership priorities. UX teams must align with this trajectory to gain influence. Here are 3 takeaways for the UX community:

1. Own the “Human-in-the-Lead” Experience

The SEI model explicitly calls out human-in-the-lead automation as a necessary capability for achieving Aligned AI. UX professionals are uniquely equipped to define what this looks like in practice. You must step up to design the interfaces, handoffs, and feedback loops that empower humans to monitor, override, and collaborate with autonomous AI agents, especially at critical and irreversible decision points.

2. Translate Explainability into User Trust

Engineers approach the “Transparency and Explainability” dimension through system logs and algorithmic traceability. However, backend data doesn’t automatically equal user trust: if users don’t understand the supposed “explanation,” then AI hasn’t been explained. UX leaders must bridge this gap by translating technical explainability into usable front-end experiences. Designing and testing patterns that clearly communicate an AI’s confidence level, data sources, and reasoning is critical for widespread user adoption and psychological safety.

3. Re-engineer the Design Workflow

Heed the call for workflow re-engineering within your own discipline. Transition your team from ad-hoc AI usage to the structured integration of AI tools for user research, synthesis, and prototyping. By systematizing how your design team uses AI and demonstrating ROI through improved design quality and efficiency, you lead by example and set a precedent for the rest of the enterprise.

In conclusion, the SEI model provides the operational and engineering bedrock for AI at scale, but it’s up to UX professionals to build the human-centered structures on top of it. Align design maturity with enterprise maturity, and you keep the AI revolution centered on the humans it serves.

My comic strip about a product team proceeding through the SEI AI maturity levels:

I reused the characters from my comic strip about predictions for 2026, but transformed them from 2-D cartoon characters into 3-D CGI characters. I think I prefer the 2-D version, which is also what I used when having these characters explain the new book, Sentient Design. (GPT-Images-2)

UX Hero: Stuart K. Card of Xerox PARC

Probably the leading theoretician of user experience is Dr. Stuart K. Card, who led the user interface at Xerox PARC during the many years it was the world’s leading HCI research laboratory. At the very least, Xerox PARC was always one of the top three labs: there were a few years when I would have ranked Bell Communications Research number one in the world, but, of course, I am biased since that was my lab.

In any case, leading user interface research while his lab was number one in the world most years and in the top three every year for three decades is an impressive record.

(GPT-Images-2)

Stu Card is probably best known as one of the co-authors of the pioneering book The Psychology of Human–Computer Interaction and as the originator of the GOMS method for quantifying usability, described in the book. However, information foraging is more important in my book, since it’s the theory that describes why web users behave the way they do, flitting between websites like so many animals browsing for food.

The information foraging theory developed by Stuart Card and Peter Pirolli posits that people search for information online in much the same way that wild animals hunt for food in their natural habitat. Drawing on optimal foraging theory from ecology and evolutionary biology, the model suggests that just as a predator instinctively weighs the energy it will gain from catching prey against the effort required to pursue it, a person browsing the web continually estimates how much useful information a given source is likely to yield relative to the time and effort needed to obtain it. (GPT-Images-2).

Central to information foraging is the concept of “information scent”: the collection of cues that hint at the value and relevance of a source before the user actually engages with it. Just as an animal follows the trail or smell of its prey before committing to the chase, a web user reads link labels, headings, snippets, icons, and other visual signals to judge whether a page is worth visiting and reading. Strong scent, meaning clear and relevant cues that closely match the user's goal, encourages people to follow a path, whereas weak or misleading scent prompts them to abandon it and look elsewhere.

The theory also introduces the idea of information “patches,” such as a website, a set of search results, or a single document, where related information tends to cluster. Foragers must constantly decide whether to keep exploiting their current patch or move on to a more promising one, a judgment that mirrors how animals choose when to leave a depleted feeding ground. When the perceived cost of extracting further value from a patch outweighs the expected gain, users tend to move on in search of richer territory. Understanding these behaviors pays off in web design: interfaces with strong, accurate information scent help users find what they need quickly, reducing frustration and the likelihood that they abandon a site altogether.

I never had the privilege of working directly with Stu Card, but I enjoyed many fine dinners with him while we served together on various boards and committees. A swell guy, in addition to being the number one theoretical thinker in UX.

Microsoft Frontier Company = MS FDE

Microsoft has announced the formation of “Microsoft Frontier Company” with a $2.5 billion investment. The new organization will have about 6,000 forward-deployed engineers (FDE) to help customers innovate in AI.

Microsoft claims a competitive advantage by promising that its FDEs will use the best AI models for the customers’ needs, whether Microsoft’s own or those from leading vendors such as OpenAI or Anthropic. Of course, it remains to be seen how vendor-independent the advice from an MS FDE will be in practice.

I would advice customers to get the mug shown in this cartoon: “Trust but Verify.” Hopefully, Microsoft’s new FDE consultants will, in fact, offer vendor-independent advice. But do verify. (GTP-Images-2)

This announcement is part of a broader trend of using the FDE model to advance AI in companies. The old model of installing a chatbot and avoiding fundamental integration with the company is dead. (For an entertaining overview of what an FDE does, watch my FDE music video.)

Just get a chatbot. Nice and easy way to install AI, but it rarely works. Follow the band instead. (GPT-Images-2)

One caveat: we now have extensive experience that substantial corporate profits from AI require deep workflow redesign, and not just installing technology. This is more of a service design problem than an implementation problem, which is why I recommend supplementing the FDE with an FDD (forward-deployed designer). I hope the new Microsoft Frontier Company will take this advice to heart.

You need a forward-deployed designer to point the FDE in the right direction. Workflow redesign before implementation. If Microsoft won’t offer this extra service, hire your own FDD, though there are not a lot of qualified candidates on the market yet, because most UX professionals have been late in specializing in enterprise workflow redesign for AI. (GPT-Images-2)

Amazon is investing a comparatively piddling $1 billion in its own FDE service to help AWS customers improve their use of AI. I can only give them the same advice as I gave Microsoft: spend some of that money on a strong FDD team, even if you have to train them yourself.

Conclusion: Intelligence Is Abundant, Usability Is Scarce

6 stories, 1 picture. UXBench proves that models that win Math Olympiad gold still bungle a basic conversation: when a user complains, the AI recovers only 12% of the time. The clickstream study proves that AI already moves markets, but invisibly: your analytics dashboard credits Google for the customers ChatGPT sent you. My music video shows a frontier model that renders vibrating bass strings in native 4K while the piano keys sit frozen. And SEI needed a 63-page maturity model (and Microsoft a $2.5 billion bet on 6,000 forward-deployed engineers) because installing AI is easy, whereas profiting from it requires deep workflow redesign.

The common thread: raw AI capability has sprinted years ahead of delivered user value. Closing that gap is the defining business problem of 2026.

The way to close it comes from the oldest story in this issue. Stu Card taught us 30 years ago to treat users as information foragers: watch what they do, not what they say. UXBench found its ground truth in behavioral signals because users click feedback buttons on a mere 0.40% of turns. The commerce researchers found the hidden funnel by logging clickstreams because last-click attribution lies. The method that built HCI in the 1980s is the method that will fix AI in the 2020s: observe real behavior, then design for it.

Intelligence became abundant; usability stayed scarce. Scarcity commands a premium, so the next wave of AI profits will flow to the teams that close the capability–usability gap. UX professionals: that gap is your job description for the rest of the decade.

A fitting thought for America’s 250th birthday: the founders shipped a stirring Declaration in 1776, then spent 13 years debugging the implementation before the Constitution shipped in 1789. AI in 2026 is at the Declaration stage: the manifesto dazzles, but the implementation frustrates. Get debugging, but let’s not spend 13 years this time.

I’m experimenting with reskinning my recurring narrator characters, Alice and Zimo, in alternative cartooning styles. Something AI does beautifully. Here’s my conclusion retold as a comic strip in Mid-Century Flat Modernist style. (GPT-Images-2)