<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Jenks Miller on Medium]]></title>
        <description><![CDATA[Stories by Jenks Miller on Medium]]></description>
        <link>https://medium.com/@jenks.miller.ux?source=rss-d13bb1eaa404------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*7iV661lSEuQ0NGwy</url>
            <title>Stories by Jenks Miller on Medium</title>
            <link>https://medium.com/@jenks.miller.ux?source=rss-d13bb1eaa404------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 20:06:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jenks.miller.ux/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Conversational interfaces in agentic systems: UX strategy in the development of AI agents]]></title>
            <link>https://medium.com/@jenks.miller.ux/conversational-interfaces-in-agentic-systems-ux-strategy-in-the-development-of-ai-agents-80235e7dcd2f?source=rss-d13bb1eaa404------2</link>
            <guid isPermaLink="false">https://medium.com/p/80235e7dcd2f</guid>
            <category><![CDATA[ai-agents-in-action]]></category>
            <category><![CDATA[ux-design]]></category>
            <category><![CDATA[conversation-design]]></category>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jenks Miller]]></dc:creator>
            <pubDate>Fri, 16 Jan 2026 15:46:37 GMT</pubDate>
            <atom:updated>2026-01-21T13:33:35.465Z</atom:updated>
            <content:encoded><![CDATA[<p>Every company is racing to add AI chat to everything. But what if conversational interfaces are actually the <em>wrong</em> default?</p><p>Before you add a chat interface to your product, ask: <strong>Are my users actually unclear about what they want? Or are they crystal clear, and my system is just getting in their way?</strong></p><h3>The problem</h3><p>Here’s the problem: We’ve conflated “AI agents” with “chat interfaces.” This conflation limits strategic value. Not every agentic system needs a conversation.</p><p>Successful deployment of agentic systems requires a more nuanced and strategic approach to interface design. Importantly, it requires understanding when users need a GUI and when they need a conversation.</p><p>My team successfully developed an AI service agent for a multi-billion-dollar global company. In our work, we leveraged a UX strategy that effectively handles this nuance.</p><p>What we learned may help other teams working with agentic systems understand <strong>where</strong> and <strong>how</strong> GenAI tools provide value/ROI.</p><h3>Why GUIs win</h3><p>UX principles tell us that given the choice between a well-designed Graphical User Interface and a chat interface, most users choose the GUI.</p><p>Why? <a href="https://blog.cognifit.com/feeling-tired-after-socializing-brain-explanation/#:~:text=Social%20Interaction%20Is%20a%20Cognitive,language%20processing%2C%20and%20executive%20function.">Conversation requires cognitive effort</a>. You must think about what you want, translate your goal to language, and articulate it coherently. This is true whether your audience is a human or a Large Language Model. Verbalizing intent takes time and increases mental load.</p><p>The GUI’s original value was exactly this: visualize intent, minimize cognitive load, make tools accessible. That value still holds.</p><figure><img alt="A human interacting with a graphical user interface" src="https://cdn-images-1.medium.com/max/1024/1*JxFqNTNwG51PBw4iPN5qmg.png" /></figure><h3>Jobs to Be Done &amp; intent clarity</h3><p>Specialized tools and platforms enjoy a great competitive advantage: In the use of a specialized tool, both the user and the system already know what the user wants to accomplish.</p><p>In designing specialized tools, we can and should anticipate the outcomes our users want to achieve, or their “<a href="https://strategyn.com/jobs-to-be-done/">Jobs to Be Done</a>.”</p><p>If we can effectively anticipate the user’s Jobs to Be Done, we can design a GUI that <a href="https://en.wikipedia.org/wiki/Don%27t_Make_Me_Think">reduces the need for users to think</a>.</p><p>In a specialized tool, a conversational interface is not a replacement for a well-designed GUI. AI can help optimize the path to the user’s goal without requiring a conversation from the user.</p><h3>When conversation matters</h3><h4>The value of conversational interfaces</h4><p>So when should we use a conversational interface rather than a GUI?</p><p>Simply put, a conversational interface is <strong>a tool for navigating ambiguity</strong>. By ambiguity, I mean a gap between the user’s Job to Be Done and the system state. When a user and system aren’t aligned around the user’s intent, we find friction. The system can no longer anticipate the user’s needs with precision. Often, the user cannot articulate their own goals because signals from the system are misaligned.</p><h4>Where GUIs excel:</h4><ul><li><strong>Known workflows:</strong> “I know what I want to do (the system should know this, too!)”</li></ul><h4>Where conversational interfaces excel:</h4><ul><li><strong>Intent ambiguity:</strong> “What am I trying to do and why do I have to do it this way?”</li><li><strong>Diagnostic ambiguity:</strong> “Why isn’t this working as expected?”</li></ul><p>While ambiguity should be minimal in your user’s primary user journey, it is the central aspect of the <em>service</em> journey. Service journeys are rich in ambiguity because users arrive after something has gone wrong or confused them in the primary journey<em>.</em> <em>By definition, the service journey involves a mismatch between user expectations and system state</em>. Ambiguity is the fundamental consideration in all service design.</p><p>Navigating ambiguity is what human service agents do all day, every day. Naturally, it’s a central part of the strategy that should drive agentic customer service systems:</p><p>1) Address ambiguity for the user as quickly as possible<br>2) Return the user to their primary journey in the interface/GUI</p><h4>A heuristic</h4><p>We can turn the value of different interfaces into a basic heuristic when strategizing for agentic system development:</p><ul><li><strong>Intent clarity/known workflow</strong> → GUI (with AI and automation support in the background)</li><li><strong>Intent ambiguity</strong> → Conversational interface</li><li><strong>Diagnostic ambiguity</strong> → Conversational interface (some exceptions apply)</li></ul><p>Primary journey in a specialized tool = alignment on Jobs to Be Done = optimize for GUI (“Don’t make me think!”)</p><p>Secondary/service journey = intent ambiguity = optimize for conversation (user has to think to navigate ambiguity)</p><h3>Service agent performance: Validation of the strategy</h3><p>The right interface choice directly impacts operational costs. Every conversation that could have been a click is money left on the table.</p><p>Our hybrid approach (conversation for service journey ambiguity → return to a GUI for primary journey) has delivered measurable results since we deployed the most recent iteration of our service agent in April 2025:</p><p><strong>Containment Rate (“True” Agentic Resolution Rate):</strong> 22.57%, up from 8.84% baseline, representing thousands of successfully resolved customer interactions monthly.<br><strong>CSAT:</strong> 4.1/5, up from 2.4/5 baseline</p><p>These metrics validate that conversational interfaces, <em>when strategically deployed to resolve ambiguity</em>, deliver both efficiency and satisfaction.</p><h3>Understanding user behavior: When users build shortcuts</h3><p>In the initial push to adopt AI tools, organizations may experience a proliferation of user-created, conversational AI agents. Enablement efforts may result in a sudden explosion of Glean agents, Claude agents, ChatGPT agents, etc. But the value and adoption of these agents may be unclear.</p><ul><li><strong>What is the long-term value of each of these agents? Is that value worth the rising cost of agent run fees?</strong></li><li><strong>Are users actually using these agents? <em>Should</em> they be using them, or are better alternatives available?</strong></li></ul><p>A UX perspective may be relevant here.</p><p>In any UX analysis, unanticipated user behavior is one of the best indicators of a system gap. This is because users own their desired outcomes … the system is just a means to an end.</p><p>Where the system fails or makes tasks unnecessarily complicated, users invent shortcuts to achieve their desired outcomes. UX designers often find that systems can be improved by building user-invented shortcuts into the system (with appropriate adjustments to optimize integration and scalability).</p><p>In this way, a user’s shortcut signals an unmet need. It tells us <strong>where</strong> a gap exists between a user need and system function. And it often illuminates a path to improvement, i.e. <strong>how</strong> that gap can be closed.</p><p><strong>An AI agent may be a shortcut.</strong> A user may build a conversational agent because the system is not optimized to perform an important job. However, this does not mean that a conversational interface is the most efficient and cost-effective tool for that job. In this way, a conversational agent may have value not because it is itself an optimal solution, but <em>because it can show us the way to an optimal solution</em>.</p><p>For example: As a sales rep, I need to gather data points A, B, and C across my accounts so that I can create a selling strategy. I have to do this several times a day, but my GUI does not display this data in the same place, requiring me to traverse several platforms to accomplish my goal. So I build an AI agent to gather data points A, B, and C. Now each time I need to accomplish this task, I can open my chat interface and type “Find A, B, and C for accounts W, X, Y, and Z.” This requires less effort than traversing multiple tools. But I don’t actually want to type this stuff all day! I just need the data. That data could be more readily accessible in a GUI dashboard. The agent’s conversational interface is not optimal … it’s a temporary shortcut that reveals a gap.</p><p><strong>The conversational workaround</strong></p><ul><li>Open chat interface</li><li>Type: “Find A, B, and C for accounts W, X, Y, Z”</li><li>Wait for response</li><li>Copy data to wherever you actually need it</li></ul><p><strong>The optimal solution</strong></p><ul><li>Dashboard widget displays A, B, C automatically</li><li>One glance, no typing, no waiting</li></ul><p>The agent revealed the need. A GUI should fulfill it.</p><p>Efforts aimed at operational efficiency might consider systematically <strong>moving the functionality of high-use conversational agents into a GUI</strong> where possible, thereby reducing both the expense of agent run fees and the cognitive load required in conversational interfaces.</p><h3>Design principles for hybrid systems</h3><ol><li>Detect intent ambiguity early through user behavior and explicit signals</li><li>Make the mode transition (GUI ↔ conversation) seamless</li><li>Always provide escape hatches (human escalation, return to GUI)</li><li>All handoffs must preserve context and momentum</li><li>Design for graceful degradation: when an AI agent can’t resolve ambiguity, provide clear paths to human escalation without making users feel like they’ve failed</li><li>Measures of success differ by interface:</li></ol><ul><li>GUI = Task completion rate &amp; CSAT</li><li>Conversation = Containment rate, CSAT, and resolution time</li></ul><h3>The ChatGPT trap</h3><p>From a UX perspective, the push to replicate the ChatGPT experience in every tool risks thinking with a technology rather than optimizing for user value.</p><p>Nevertheless, the ChatGPT example is instructive: ChatGPT is the ultimate example of ambiguity. It is <em>the opposite</em> of a specialized platform. Because the range of user intent is so broad, ChatGPT and similar “Swiss army knives” must accommodate everything, which means they can optimize for nothing. This is likely why the Claude experience opens with “Coffee and Claude?” or “Welcome back!” rather than providing something clickable.</p><p>Specialized tools and platforms do not share this level of ambiguity, and therefore should not seek to emulate the experience of ChatGPT in the primary user journey. Specialized tools and platforms have the advantage of intent clarity. Agentic system designers should capitalize on that clarity wherever possible to ensure usability and maintain operational efficiency.</p><p>So before you add a chat interface to your product, ask: <strong>Are my users actually unclear about what they want? Or are they crystal clear, and my system is just getting in their way?</strong></p><p>In my next article, I’ll discuss aspects of ambiguity in agentic system design.</p><p><em>Jenks Miller is a UX Designer specializing in human-centered agentic system design and service design.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=80235e7dcd2f" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>