Designing human-centered AI products can be arduous.Keeping up with the overall pace of change isn’t easy. But here’s a bigger challenge:The wildly different paces of change attached to the key elements of AI product strategy, design, and development can make managing those elements — and even thinking about them — overwhelming.Yesterday’s design processes and frameworks offer priceless guidance that still holds. But in many spots, they just don’t fit today’s environment.For instance, designers used to map out and user-test precise, predictable end-to-end screen flows. But flows are no longer precisely predictable. AI generates dynamic dialogues and custom-tailored flows on the fly, rendering much of the old practice unhelpful and infeasible.It’s easy for product teams to feel adrift nowadays — we can hoist the sails, but we’re missing a map and a rudder. We need frameworks tailored to the traits that fundamentally set AI apart from traditional software, including:its capabilities for autonomy and collaboration,its probabilistic nature,its early need for quality data, andits predictable unpredictability. Humans tend to be perpetually surprised by its abilities — and its inabilities.AI pace layers: design for resilienceHere’s a framework to address these challenges.Building on Stewart Brand’s “Shearing Layers” framework, AI Pace Layers helps teams grow thriving AI products by framing them as layered systems with components that function and evolve at different timescales.It helps anticipate points of friction and create resilient and humane products.Each layer represents a specific domain of activity and responsibility, with a distinct pace of change.* Unlike the other layers, Services cuts across multiple layers rather than sitting between them, and its pace of change fluctuates erratically.Boundaries between layers call for special attention and care — friction at these points can produce destructive shearing and constructive turbulence.I’ll dive deeper into this framework with some practical examples showing how it works. But first, a brief review of the precursors that inspired this framework will help you put it to good use.The foundationsThis model builds on the insights of several influential design frameworks from the professions of building architecture and traditional software design.Shearing layers (Duffy and Brand)In his 1994 book How Buildings Learn, Stewart Brand expanded on architect Frank Duffy’s concept of shearing layers. The core insight: buildings consist of components that change at different rates.Shell, Services, Scenery, and Sets. (Frank Duffy, 1992).“…there isn’t any such thing as a building. A building properly conceived is several layers of longevity of built components.” — Frank DuffyShearing Layers of Change, from How Buildings Learn: What Happens after they’re built (Stewart Brand, 1994).Expanding on Duffy’s work, Brand identified six layers, from the slow-changing “Site” to the rapidly evolving “Stuff.”As the layers move at different speeds, friction forms where they meet. Buildings designed without mindful consideration of these different velocities tear themselves apart at these “shearing” points. Before long, they tend to be demolished and replaced.Buildings designed for resiliency allow for “slippage” between the moving layers — flexibility for the different rates of change to unfold with minimal conflict. Such buildings can thrive and remain useful for hundreds of years.Pace layers (Brand)In 1999, Brand drew insights from ecologists to expand this concept beyond buildings and encompass human society. In The Clock Of The Long Now: Time And Responsibility, he proposed “Pace Layers” — six levels ranging from rapid fashion to glacially-slow nature.Brand’s Pace Layers (1999) as sketched by Jono Hey.Brand again pointed out the boundaries, where the most intriguing and consequential changes emerge. Friction at the tension points can tear a building apart — or spur a civilization’s collapse–when we try to bind the layers too tightly together. But with mindful design and planning for slippage, activity along these boundary zones can also generate “constructive turbulence” that keeps systems balanced and resilient.The most successful systems survive and thrive through times of change through resiliency, by absorbing and incorporating shocks.“…a few scientists (such as R. V. O’Neill and C. S. Holling) have been probing the same issue in ecological systems: how do they manage change, how do they absorb and incorporate shocks? The answer appears to lie in the relationship between components in a system that have different change-rates and different scales of size. Instead of breaking under stress like something brittle, these systems yield as if they were soft. Some parts respond quickly to the shock, allowing slower parts to ignore the shock and maintain their steady duties of system continuity.” — Stewart BrandRoles and tendencies of the fast (upper) and slow (lower) layers. (Brand).Slower layers provide constraints and underpinnings for the faster layers, while faster layers induce adaptations in the slower layers that evolve the system.Elements of UX (Garrett)Jesse James Garrett’s classic The Elements of User Experience (2002) presents a five-layer model for digital design:Surface (visual design)Skeleton (interface design, navigation design, information design)Structure (interaction design, information architecture)Scope (functional specs, content requirements)Strategy (user needs, site objectives)Structure, Scope, and Strategy. Each layer answers a different set of questions, with the questions answered at each level setting constraints for the levels above. Lower layers set boundaries and underpinnings that help define the more concrete layers.Jesse James Garrett’s 5 layers from The Elements of User Experience Design (2002)This framework doesn’t focus on time, or on tension points resulting from conflicting velocities. But it provides a comprehensive structure for shaping different aspects of digital product design, from abstract strategy to concrete surface elements.AI Pace Layers: diving deeperBuilding on these foundations, the AI Pace Layers framework adapts these concepts specifically for AI systems design.Let’s explore each layer and understand how design expertise contributes across the framework.SessionsPace of change: Very fast (milliseconds to minutes)Focus: Performance of real-time interactions.This layer encompasses real-time dialogue, reasoning, and processing. These interplays happen between the user and AI, and between AI agents and other services and people, on behalf of the user. Sessions draw on lower-layer capabilities and components to deliver the “moments of truth” where product experiences succeed or fail. Feedback from the Sessions layer is crucial for improving and evolving the lower layers.Key contributors: Users and AI agents — usually with zero direct human involvement backstage.Example actions/decisions/artifacts: User/AI dialogue. Audio, video, text, images, and widgets are rendered on the fly (using building blocks provided by lower levels). Real-time adaptations to context.SkinPace of change: Moderately fast (days to months)Focus: Design patterns, guidelines, and assetsSkin encompasses visual, interaction, and content design.Key contributors: Designers, content strategists, front-end developers, and user researchers.Design’s role: This is where designers’ traditional expertise shines. They craft the interface elements, establish visual language, define interaction patterns, and create the design systems that represent the product’s capabilities to users.Example actions/decisions/artifacts: UI component libraries, brand guidelines, prompt templates, tone of voice guidelines, navigation systems, visual design systems, patterns (UI, interaction, and conversation), content style guides.ServicesPace of change: Wildly variable (slow to moderately fast)Focus: AI computation capabilities, data systems orchestration, and operational intelligenceThe Services layer provides probabilistic AI capabilities that sometimes feel like superpowers — and like superpowers, they can be difficult to manage. It encompasses foundation models, algorithms, data pipelines, evaluation frameworks, business logic, and computing resources.Services is an outlier that behaves differently from the other layers:• It’s more prone to “shocks” and surprises that can ripple across the rest of the system.• It varies wildly in pace of change. (But its components rarely change faster than Skin, or slower than Skeleton.)• It cuts across multiple layers rather than sitting between two of them. That produces more cross-layer boundaries, more tension points, more risks of destructive friction, and more opportunities for constructive turbulence.Key contributors: Data scientists, engineers, service designers, ethicists, product teamsDesign’s role: Designers partner with technical teams on evaluation frameworks, helping define what “good” looks like from a human experience perspective. They contribute to guardrails, monitoring systems, and multi-agent collaboration patterns, ensuring technical capabilities translate to meaningful human experiences. Service design expertise helps orchestrate complex, multi-touchpoint AI capabilities.Example actions/decisions/artifacts: Foundation model selection, changes, and fine-tuning. Evals, monitoring systems, guardrails, performance metrics. Business rules, workflow orchestration. Multiagent collaboration and use of external tools (APIs, A2A, MCP, etc.) Continual appraisal and adoption of new tools, protocols, and capabilities.SkeletonPace of change: Moderately slow (months) Focus: Fundamental structure and organizationThis layer establishes the foundational architecture — the core interaction models, information architecture and organizing principles.Key contributors: Information architects, information designers, user researchers, system architects, engineersDesign’s role: Designers with information architecture expertise are important in this layer. They design taxonomies, knowledge graphs, and classification systems that make complex AI capabilities comprehensible and usable. UX researchers help ensure these structures fit the audience’s mental models, contexts, and expectations.Example actions/decisions/artifacts: Taxonomies, knowledge graphs, data models, system architecture, classification systems.ScopePace of change: Slow (months to years)Focus: Product requirementsThis layer defines core functional, content, and data requirements, accounting for the probabilistic nature of AI and defining acceptable levels of performance and variance.Key contributors: Product managers, design strategists, design researchers, business stakeholders, data scientists, trust & safety specialistsDesign’s role: Design researchers and strategists contribute to requirements through generative and exploratory research. They help define error taxonomies and acceptable failure modes from a user perspective, informing metrics that capture technical performance and human experience quality. Design strategists balance technical possibilities with human needs and ethical considerations.Example actions/decisions/artifacts: Product requirements documents specifying reliability thresholds, data requirements (volume, diversity, quality standards), error taxonomies and acceptable failure modes, performance metrics frameworks, responsible AI requirements, risk assessment, core user stories and journeys, documentation of expected model variance and handling approaches.StrategyPace of change: Very slow (years)Focus: Long-term vision and business goalsThis foundation layer defines audience needs, core problems to solve, and business goals. In AI products, data strategy is central.Key contributors: Executive leadership, design leaders, product leadership, business strategists, ethics boardsDesign’s role: Design leaders define problem spaces, identify opportunities, and plan roadmaps. They deliver a balance of business needs with human values in strategy development. Designers with expertise in responsible AI help establish ethical frameworks and guiding principles that shape all other layers.Example actions/decisions/artifacts: Problem space and opportunity assessments, market positioning documents, long-term product roadmaps, comprehensive data strategy planning, user research findings on core needs, ethical frameworks and guiding principles, business model documentation, competitive/cooperative AI ecosystem mapping.Practical examples: tension points between layersTension point example 1: Bookmuse’s timeline troubles(Friction between Sessions and Skin)Bookmuse is a promising new AI tool for novelists. Samantha, a writer, tries it out while hashing out the underpinnings of her latest time-travel historical fiction thriller. The Bookmuse team planned for plenty of Samantha’s needs. At first, she considers Bookmuse a handy assistant. It supplements chats with tailored interactive visualizations that efficiently track character personalities, histories, relationships, and dramatic arcs.But Samantha is writing a story about time travelers interfering with World War I events, so she’s constantly juggling dates and timelines. Bookmuse falls short. It’s a tiny startup, and Luke, the harried cofounder who serves as a combination designer/researcher/product manager, hasn’t carved out any date-specific timeline tools or date calculators. He forgot to provide even a basic date picker in the design system.Problem: Bookmuse does its best to help Samantha with her story timeline (Sessions layer). But it lacks effective tools for the job (Skin layer). Its date and time interactions feel confusing, clumsy, and out of step with the rest of its tone, look, and feel. Whenever Samantha consults the timeline, it breaks her out of her creative flow.Constructive turbulence opportunities:a) Present feedback mechanisms that ensure this sort of “missing piece” event results in the product team learning about the type of interaction pothole that appeared — without revealing details or content that compromise Samantha’ privacy and her work. (For instance, a session tagging system can flag all interaction dead-ends during date choice interactions.)b) Improve timeline/date UI and interaction patterns. Table stakes: Standard industry-best-practice date picker components that suit Bookmuse’s style, tone, and voice. Game changers: Widgets, visualizations, and patterns tailored to the special time-tracking/exploration challenges that fiction writers often wrestle with.c) Update the core usability heuristics and universal interaction design patterns baked into the evaluation frameworks (in the Services layer), as part of regular eval reviews and updates. Result: When the team learns about a friction moment like this, they can prevent a host of future similar issues before they emerge.These improvements will make Bookmuse more resilient and useful.Tension point example 2: MedicalMind’s diagnostic dilemma(Friction between Services and Skeleton)Thousands of healthcare providers use MedicalMind, an AI-powered clinical decision support tool. Dr. Rina Patel, an internal medicine physician at a busy community hospital, relies on it to stay current with rapidly evolving medical research while managing her patient load.Thanks to a groundbreaking update, a MedicalMind AI model (Services layer) is familiar with new medical research data and can recognize newly discovered connections between previously unrelated symptoms across different medical specialties. For example, it identified patterns linking certain dermatological symptoms to early indicators of cardiovascular issues — connections not yet widely recognized in standard medical taxonomies.But MedicalMind’s information architecture (Skeleton layer) was tailored to traditional medical classification systems, so it’s organized by body system, conditions by specialty, and treatments by mechanism of action. The MedicalMind team constructed this structure based on how doctors were traditionally trained to approach medical knowledge.Problem: When Dr. Patel enters a patient’s constellation of symptoms (Sessions layer), MedicalMind’s AI can recognize potentially valuable cross-specialty patterns (Services layer). But these insights can’t be optimally organized and presented because the underlying information architecture (Skeleton layer) doesn’t easily accommodate the new findings and relationships. The AI either forces the insights into ill-fitting categories or presents them as disconnected “additional notes” that tend to be overlooked. That reduces their clinical utility and Dr. Patel’s trust in the system.Constructive turbulence opportunities:a) Create an “emerging patterns” framework within the information architecture (Skeleton layer) that can accommodate new AI-identified patterns in ways that augment, rather than disrupt, the familiar classification systems that doctors rely on.b) Design flexible visualization components and interaction patterns and styles (in the Skin layer) specifically for exploring, discussing, and documenting cross-category relationships. Let doctors toggle between traditional taxonomies and newer, AI-generated knowledge maps depending on their needs and comfort level.c) Implement a clinician feedback loop where specialists can validate and discuss new AI-surfaced relationships, gradually promoting validated patterns into the main classification system.These improvements will make MedicalMind more adaptive to emerging medical knowledge while maintaining the structural integrity that healthcare professionals rely on for critical decisions. This provides more efficient assistants for clinicians and better health for patients.Tension point example 3: ScienceSeeker’s hypothesis bottleneck(Friction between Skin and Services)ScienceSeeker is an AI research assistant used by scientists worldwide. Dr. Elena Rodriguez, a molecular biologist, uses it to investigate protein interactions for targeted cancer drug delivery.The AI engine (Services layer) recently gained the ability to generate sophisticated hypothesis trees with multiple competing explanations, track confidence levels for each branch, and identify which experiments would most efficiently disambiguate between theories. It can reason across scientific domains, connecting molecular biology with physics, chemistry, and computational modeling.But the interface (Skin layer) remains locked in a traditional chatbot paradigm — a single-threaded exchange with responses appearing sequentially in a scrolling window.Problem: The AI engine and the problem space are natively multithreaded and multimodal, but the UI is limited to single-threaded conversation. When Dr. Rodriguez inputs her experimental results (Sessions layer), the AI generates a rich, multidimensional analysis (Services layer), but must flatten this complex reasoning into linear text (Skin layer). Critical relationships between hypotheses become buried in paragraphs, probability comparisons are difficult, and the holistic picture of how variables influence multiple hypotheses is lost. Dr. Rodriguez resorts to taking screenshots and manually drawing diagrams to reconstruct the reasoning that the AI possesses but cannot visually express.Constructive turbulence opportunities:a) Develop an expandable, interactive, infinite-canvas “hypothesis tree” visualization (Skin) that helps the AI dynamically represent multiple competing explanations and their relationships. Scientists can interact with this to explore different branches spatially rather than sequentially.b) Create a dual-pane interface that maintains the chat for simple queries but provides the infinite canvas for complex reasoning, transitioning seamlessly based on response complexity.c) Implement collaborative, interactive node-based diagrams for multi-contributor experiment planning, where potential experiments appear as nodes showing how they would affect confidence in different hypothesis branches.This would transform ScienceSeeker’s limited text assistant into a scientific reasoning partner. It would help researchers visualize and interact with complex possibilities in ways that better fit how they tackle multidimensional problems.Navigating the future with AI Pace LayersAI Pace Layers offers product teams a new framework for seeing and shaping the bewildering structures and dynamics that power AI products.By recognizing the evolving layers and heeding and designing for their boundaries, AI design teams can:Transform tension points into constructive innovationAnticipate friction before it damages the product experienceGrow resilient and humane AI systems that absorb and integrate rapid technological change without losing sight of human needs.The framework’s value isn’t in rigid categorization, but in recognizing how components interact across timescales. For AI product teams, this awareness enables more thoughtful design choices that prevent destructive shearing that can tear apart an AI system.This framework is a work in progress, evolving alongside the AI landscape it describes.I’d love to hear from you, especially if you’ve built successful AI products and have insights on how this model could better reflect your experience. Please drop me a line or add a comment. Let’s develop more effective approaches to creating AI systems that enhance human potential while respecting human agency.Part of the Mindful AI Design series. Also see:The effort paradox in AI design: Why making things too easy can backfireBlack Mirror: “Override”. Dystopian storytelling for humane AI designStay updatedSubscribe to be notified when new articles in the series are published. Join our community of designers, product managers, founders and ethicists as we shape the future of mindful AI design.AI Pace Layers: a framework for resilient product design was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.