Jan 16, 2026
In the 2023-2024 cycle, SaaS vendors prioritized “generative” features—tools focused on drafting content, generating smart replies, and basic copilot assistance. While useful, these features remain fundamentally passive. They require the user to initiate, refine, and finalize every output. The 2025 trend signals a definitive move toward Agentic AI—systems designed to take actions rather than just generate text.
Agentic systems differ from generative ones in their ability to orchestrate multi-step tasks across various tools. An agent within a CRM, for example, does not just draft an email to a lead; it analyzes the lead’s behavior, triggers a workflow in the marketing platform, updates the record in the ERP, and schedules a follow-up, all within set constraints. This shift necessitates a move from “human-in-the-loop” to “human-on-the-loop” or “human-at-the-helm” configurations, where the designer’s job is no longer to guide clicks and taps but to shape systems that fulfill user intent.
| Feature Category | Generative AI Focus (2024) | Agentic AI Focus (2025) |
| Primary Output |
Text, images, code snippets |
Actions, workflow triggers, record updates |
| User Role |
Prompter and Editor |
Director and Auditor |
| System Capability |
Content drafting and summaries |
Multi-step task orchestration |
| Integration Level |
Standalone or sidebar copilot |
Deeply embedded, action-taking agents |
| Value Driver |
Time saved on creative tasks |
Automation of entire workflows (e.g., invoice matching) |
The implications for UX are profound. Traditional design focuses on navigation—helping users find the right button in a static menu. Agentic design focuses on Intent Architecture—ensuring the system correctly parses what the user wants to achieve and why. As AI moves from a novelty to “Industrial-Grade MLOps,” the focus on reliability and scalability becomes the new benchmark for design excellence.
As the AI-as-a-Service (AIaaS) market is projected to reach approximately $\$176 \text{ billion}$ by 2032, the way SaaS companies package and price these capabilities is undergoing a radical change. Founders must decide how to monetize AI without alienating their core user base or cannibalizing their existing seat-based revenue.
The emerging strategy is the “unbundling and rebundling” of AI features into specific tiers. This approach allows companies to offer a “core” product while providing “pro AI add-ons” and “enterprise AI automation” tiers for power users.
Pricing is moving away from traditional “flat seat” models toward usage-based models. Customers are increasingly charged based on:
Tokens: The amount of data processed by the underlying Large Language Model (LLM).
API Calls: The number of times the system interacts with external or internal services.
Messages or Documents: Specific units of work completed by the agent.
SaaS providers are effectively becoming “verticalized, UX-friendly layers” that resell AI infrastructure from major providers like OpenAI, Anthropic, or open-weight model hosts. The value they add is not the model itself, but the design of the interaction layer that makes that model usable for specific business tasks. This necessitates a “FinOps” approach to AI UX, where designers must be aware of the cost implications of model selection, token usage, and redundant API calls from the very first day of development.
To move beyond the limitations of the chatbot, the Applied Innovation Exchange (AIE) has distilled six key design patterns that define the modern rules of interaction. These patterns provide a framework for integrating AI into the user experience without overwhelming the user.
This is the most basic pattern, where AI operates as a standalone tool separate from the primary interface. While it requires the least amount of engineering effort, it forces users to switch contexts, which is often a source of friction. An example would be having a separate window for ChatGPT or Claude open while working in a different application.
In this pattern, the AI leverages real-time data to adapt to the user’s context but does not directly alter the primary interface. It provides suggestions or summaries in a sidebar or side panel, such as Replit’s code suggestions or Slack AI’s search summaries. The core experience remains untouched, but it is “influenced” by AI insights.
Here, AI is embedded directly into the native interface as a permanent part of the workflow. It assists without requiring the user to navigate away. A prime example is Notion AI, which allows users to generate or edit content directly within the document they are writing. Crucially, this pattern still requires user approval for changes.
This is a more advanced pattern where the AI actively generates significant components of the user experience, such as dynamic content or personalized layouts. Canva’s Magic Design, which generates entire layouts based on a simple prompt, illustrates this. While highly personalized, it can reduce predictability, making it harder for users to develop a stable mental model of where things are.
Point patterns involve “micro-agents” focused on specific, localized tasks. They provide instant feedback at a granular level. A familiar example is the AI in Google Sheets that suggests autofill formulas or identifies errors in specific cells. These micro-agents enhance the existing workflow without modifying the broader system.
The conversational pattern uses natural language for open-ended, exploratory interactions. This is ideal for tasks requiring iterative refinement or language practice, such as Duolingo’s AI chatbot. While flexible, it is often slower than direct assistance patterns because it requires the user to formulate structured queries.
| Pattern | Best For | Risk |
| Full Separation | General-purpose brainstorming |
Context-switching fatigue |
| Influenced | Surfacing insights without disruption |
Users may ignore the sidebar |
| Integrated Feature | Native content creation/editing |
Can clutter the primary UI |
| Generated Features | Rapid prototyping and personalization |
Loss of user control and predictability |
| Point | Error correction and micro-tasks |
Can be annoying if too frequent |
| Conversational | Exploratory and clarification tasks |
High interaction cost/slow speed |
Traditional UX patterns rely on fixed layouts and clear, pre-defined user paths. However, as AI personalization evolves, interfaces are becoming dynamic, adapting to user behavior, click patterns, and usage history. This is the transition from a “one-size-fits-all” UI to “Smart Interfaces.”
Salesforce’s use of predictive layouts has led to a 3.2x higher lead conversion rate by prioritizing commonly used features and hiding less relevant ones based on the specific user’s behavior. This approach simplifies the workspace, making it cleaner and more efficient. Similarly, HubSpot’s “Adaptive Workspaces,” launched in late 2023, improved user productivity by 32% and cut support tickets related to navigation by 22%.
Microsoft 365 takes this a step further by customizing layouts based on roles. A marketer using the platform might be presented with content creation tools, while an IT admin sees system controls, all driven by ongoing behavioral analysis. This role-based adaptation ensures that the most relevant tools are always within reach, reducing the cognitive load of searching through complex menus.
Predictive UX is not just about aesthetics; it is a driver of efficiency.
Adobe Creative Cloud: Powered by Sensei AI, it offers users up to 30% time savings on routine tasks through smart feature suggestions.
Dropbox: Machine learning-powered suggestions have increased engagement with shared content by 30% by predicting which files a user is most likely to need.
Netflix: Known for its contextual AI, it introduced dynamic interface adjustments to tailor the viewing experience to individual user habits as early as late 2022.
To successfully implement these adaptive systems, product leaders should build strong data collection systems that prioritize privacy while enabling real-time processing for instant interface updates. It is critical to introduce advanced features gradually as users become more skilled, a process known as “progressive feature rollout”.
Designing for agentic AI is fundamentally different from designing for a standard application; it is the design of a partnership. This relationship must be built on clear communication, mutual understanding, and established boundaries. As systems move from suggesting actions to executing them, designers must utilize patterns that follow the functional lifecycle of an agentic interaction.
The rise of agentic AI marks a shift from “empathy by proxy” (designing for generic personas) to “empathy by pattern” (designing for real-time individual behavior). Designers are no longer just creating layouts; they are becoming “Intent Architects” who craft behavior models and trust protocols. This evolution fosters a more personalized user journey, where the system responds to real-time behaviors to create trust in the interaction.
Explainable AI (XAI) is no longer a “nice-to-have” feature; it is a set of processes and methods that allow humans to comprehend and trust the results of machine learning algorithms. As AI models become more complex, they often become “black boxes” that even their creators cannot fully interpret. XAI aims to make these outputs understandable and transparent.
Google’s rubric highlights 22 key pieces of information that should be shared with users to ensure transparency and fairness. It is divided into three levels:
General Level: Provides a high-level overview of the role of AI in the product, the business model, and steps taken to ensure safety and address bias.
Feature Level: Details specific AI-powered features, including when they are active, user control options, system limitations, and how user data is processed.
Decision Level: Clarifies how specific decisions are made, the system’s confidence in its outputs, and how to contest a result or report an error.
IBM’s model focuses on creating purposeful, context-aware solutions by breaking down the experience into several critical considerations:
Understanding Intent: Prioritizing human-centric goals and emotions.
Data and Policy: Governing the handling of raw data with ethical and regulatory standards.
Machine Understanding and Reasoning: The ability of the system to apply logic and decide the best course of action while updating its knowledge dynamically.
Human Reactions and System Improvement Loop: Designing systems that work with humans, ensuring a balance between automation and human agency.
Implementing XAI allows organizations to operationalize AI with trust and confidence, speeding the time to results and mitigating the risks associated with model governance and unintended bias. It is particularly crucial in high-stakes sectors like healthcare, where an AI early warning system might predict the likelihood of critical illness, or banking, where it explains why a loan was approved or denied.
| XAI Method | Description | Example |
| Intrinsically Interpretable |
Algorithms that are transparent by design |
Decision trees or linear regression |
| Feature Importance |
Highlights which input variables had the most impact on a decision |
Showing that “Income” was the main reason for loan denial |
| Local Interpretability (LIME) |
Explains the prediction of a specific classifier |
Explaining why a specific image was flagged as fraudulent |
| Traceability (DeepLIFT) |
Compares the activation of each neuron to a reference |
Showing dependencies between different data points in a neural network |
| Visualization Tools |
Uses heat maps or dashboards to illustrate outcomes |
A saliency map showing which parts of an X-ray the AI focused on |
Product leaders operate in the realm of outcomes: revenue, retention, and activation. While many teams are skilled at measuring clicks and page views, these are often “vanity metrics” that do not reflect actual success in an AI context. For example, high “time on site” could mean a user is deeply engaged or that they are utterly confused by a vague AI response.
To prove the ROI of AI UX, designers must bridge the gap between usability and business outcomes.
Task Success Rate (TSR): This is the most fundamental metric. Can users do the thing the product exists for? A TSR below 85% is a signal that the product is working against its users.
Time on Task: Faster is not always better, but in most SaaS contexts, it is. A meaningful AI-driven redesign typically cuts task completion time by 30% or more by reducing friction and confusion.
Error Rate: Below 5% is the zone where users can operate with confidence. Every percentage point above this costs trust and completion rates.
System Usability Scale (SUS): This 10-question questionnaire produces a single score from 0 to 100. An average product scores 68; anything above 80 is considered excellent and gives UX credibility in the boardroom.
Conversion Rate: Even a 1% improvement in a high-volume flow (like Amazon’s checkout or Airbnb’s search) can represent millions in revenue.
Customer Lifetime Value (CLV): Great UX extends the relationship. Users churn quietly without feedback when the experience is poor; high CLV proves the UX is keeping people.
Microsoft’s research suggests a “tipping point” for AI retention: users need to interact with an AI tool for a minimum of 11 minutes across 11 weeks to transform sporadic usage into a regular, habit-forming behavior. The most successful companies often focus on a narrow user base initially, iterating until they achieve at least 50% retention before scaling broadly.
As shipping speed outpaces coherence, recurring “anti-patterns” are emerging in AI-generated software and UI. Researchers increasingly describe unmonitored AI as “an army of talented juniors without oversight”.
A 2025 analysis of over 300 repositories identified recurring anti-patterns in 80-100% of AI-generated code, including incomplete error handling and inconsistent architecture. While this code works initially, it creates a “18-month wall” of technical debt that accumulates faster than traditional review processes can manage. In multi-tenant SaaS, AI-generated logic often omits strict tenant isolation, creating risks for cross-tenant data leakage.
Bernard Marr identifies several “fatal” mistakes that could damage a business in 2025:
Omitting Human Oversight: Factual errors can be found in up to 46% of AI-generated texts. Without “human-in-the-loop” verification, businesses risk looking foolish or facing legal liability.
Substituting GenAI for Creativity: Using AI to churn out high volumes of blogs or social posts often leads to “AI slop”—generic, uninspiring content that disconnects the audience. Video game publisher Activision Blizzard was recently criticized for using AI artwork in place of human-created pieces.
Creating “Black Boxes”: If a medical diagnostics app doesn’t clarify whether a result came from an algorithm or a doctor, users won’t know if they should trust it. Uncertainty leads to mistrust.
Mistaking GenAI for a magic bullet is a common tactical blunder. Some teams attempt to replace real user empathy with “synthetic data” generated by AI. However, AI struggles with the nuances of human emotion and behavior; it cannot tell you why a user found an interaction obtrusive or why they didn’t understand an element. Designers must use AI to augment research (e.g., analyzing transcripts), but always validate insights with real human interaction.
SaaS onboarding is the series of moments that move a sign-up from a “curious clicker” to an “activated user”—someone who has experienced core value at least once. In 2025, every extra minute of onboarding time lowers conversion by 3%.
The 5-Second Wow Moment: Use a short, looping video above the fold to show the end result of the product (e.g., Calendly showing a booked meeting).
Progressive Profiling: Ask only for an email and password initially. Role and company size come later via micro-surveys that unlock specific features.
Quick-Win Setup (2 minutes max): Pre-populate sample projects so the UI never looks “naked.” Offer one-click imports from tools like Trello or Notion.
Contextual Checklist Widget: Gamify the process with a progress bar. Starting the bar at 20% utilizes the Zeigarnik effect (the tendency to remember uncompleted tasks) to drive completion.
Human Touch at Minute 23: Trigger a chat from a real success manager—not a bot—once the user completes initial setup steps. Using 15-second selfie videos can jump response rates to 34%.
Social Proof Email: Send a short email showing what another user achieved on their first day.
Expansion Upsell: Only offer upsells after the user has invited at least one teammate. Timing is more important than the offer itself.
Founders must ruthlessly cut “time-to-value inflation” and replace 2016-style 12-step popup tours with 10-second silent videos. The goal is to reach one “Aha” metric per persona as quickly as possible.
At Redbaton, the agency recognizes that businesses and their customers are evolving at rapid speed. As a digital design studio based in Bangalore, Redbaton serves as a turnkey project partner for brands seeking innovative solutions in UI/UX, branding, and design strategy. The agency is characterized by a “rabid obsession” to create work that people care about, captured in the motto: “Make Good to Get Better”.
Redbaton’s approach to AI design is rooted in the science of intuition. As documented in their study on “The Intuitive Design,” the agency leverages psychological principles—including Ivan Pavlov’s conditioning experiments—to reduce cognitive overload. By accessing previous learnings and user familiarity, Redbaton creates interfaces that feel natural and require less “active” learning from the user.
In practice, this means moving away from static wireframes to “Behavior Models.” Instead of designing a fixed path, Redbaton designs “Trust Protocols” that ensure an AI system’s actions are legible and its decisions are contestable. Whether redesigning a suite of 9 WordPress/PHP websites for a business management company or building a UI for a music learning platform, the agency’s methodical, structured workflow ensures significant performance improvements and stakeholder approval.
How can organizations avoid the GenAI “productivity trap”?
The trap occurs when general-purpose tools are deployed without clear governance or specific workflows. Break free by focusing on high-value use cases where AI augments a human in a defined workflow, such as customer service triage or code assistance.
What is the “11 by 11” tipping point?
It is a retention metric from Microsoft research indicating that users who use an AI tool for 11 minutes over 11 weeks are much more likely to become long-term, activated customers.
What are the key elements of a trustworthy AI?
Trustworthy AI doesn’t have to be perfect; it has to manage expectations well. This involves being upfront about limitations, clearly labeling “AI-generated” content, and allowing for easy human correction or “undo” actions.
How do you measure Task Success Rate for AI?
Define what “success” looks like for a core workflow, track where users drop off, and break that data down by user role to see if specific personas are struggling more than others.
