{ "@context": "https://schema.org", "@type": "Blog", "@id": "#writing", "name": "The Embodied Interface", "headline": "The Embodied Interface", "description": "A series of connected posts exploring conversation design in embodied systems.", "author": { "@type": "Person", "name": "Lia Pool", "@id": "#person" }, "blogPost": [ { "@type": "BlogPosting", "slug": "the-embodied-interface", "headline": "Introduction", "datePublished": "2025-09-01", "image": "/images/writing/the-embodied-interface.gif", "articleBody": "Our evolution is borne from scaling capability. Fire scaled energy, language scaled coordination, printing compounded learning, computers scaled computation, the internet scaled connection.\n\nToday in 2026, “AI scales thinking”. Allegedly.\n\nLarge Language Models don’t so much scale thinking as they scale delegation. If you were wealthy enough you could do this already with a team of experts at your disposal, but for the rest of us this is new. We don’t need to labour alone, or even collaborate. We just delegate. For the first time we have the opportunity to produce work on a scale well beyond our knowledge and ability.\n\nUltimately this is less a seismic shift and more a continuation of the last 50 years of work enhanced by micro computers and the world wide web. Just like the previous cycles of innovation and disruption we don’t know how to use this yet. We struggle to translate potential into real value. We’re tripped up by the simplest of requests. AI seems to make the hard work trivial, and the easy work impossible past a point.\n\nComputers today make aspects of our work feel trivial because we’ve built up incredibly sophisticated models for human computer interfaces over the decades, focusing on human ability and cognition, designed to bring predictability to layered abstract systems. Buttons do things. Menus have options. Actions have visible outcomes. There’s a repeating cadence to every UI that lets us comfortably pick up and use new technology. Yet now that computers can “understand language”, we’ve thrown all those learnings away. We’ve reintroduced ambiguity into every interaction. We’ve dropped the clear menus and actions and outcomes which let us know what will happen. Now we rely on a text box where anything seems possible, but the only way to learn what it can do is to try and fail. Now users have to scrutinize every outcome as they’re told “don’t trust this output.”\n\nHow is that acceptable as a computer interface? We’ve seen the magic of the future and just forgotten everything that got us here.\n\nNone of these pieces are new. Automations aren’t new. Chatbots aren’t new. Talking to a professional isn’t new - you’ve spoken to a bank teller before, right? The tools, the natural language understanding, the keyboards and microphones and cameras, the virtual characters already exist. Pioneers have paved the way. AI could provide the easiest computer interface imaginable. It could be safe to use. It could be a boon that everyone gets.\n\nWe could have the future today, and if you’ll follow me, I’ll teach you how." }, { "@type": "BlogPosting", "slug": "the-embodied-interface-purpose-first", "headline": "Purpose First", "datePublished": "2025-09-22", "image": "/images/writing/the-embodied-interface-purpose-first.gif", "articleBody": "I will only talk to you if there’s a reason to. Maybe I want to get something from you, or maybe I want to help you, or maybe I just want to be entertained. Whatever the reason there is always a want. So why aren’t agents designed with this want in mind?\n\nIts still rare to see AI Agents built to meet a real user want. There are a lot of very useful AI Agents around but few of them are built for a specific purpose. When someone is building an AI Agent they’re usually doing it for a commercial purpose, like promoting a product. Often conversation designers and engineers are building to meet the needs of the commercial purpose, but they don’t answer the fundamental question of: “Why do I want to talk to that agent anyway?”\n\nThe answer is a two parter: I need to be motivated to engage with this company to begin with, and his mode of engagement needs to be easier than others.\n\nI can tell you right now that a chat interface isn’t the cheap option, or the easy option, or even the accessible option. While it can be helpful to ask Siri what the weather is like, it's often faster to just tap into the weather app and get a pile of on screen data. All the capability and value of a chat interface is wasted because the want isn’t that compelling, and the other methods of achieving my goals are easier and faster.\n\nYou know what is easier through a chat interface? Finding out what the name of that bird is - y’know the New Zealand native bird that has the green feathers and white belly. Um it looks like an absolute unit? Kererū, right! In this scenario I have a need to know, and asking is easier than flipping through an encyclopedia.\n\nWhen we design for our commercial purpose we need to keep this in mind. I doubt your front page users really want to engage in an unsolicited conversation about your promotion. But maybe if they’re hovering that buy button a nudge might be more engaging.\n\nOr maybe get wild with it: challenge users to haggle a discount out of your agent. That’d give users a clear reason to talk to your agent, and they can’t really haggle with a website. If you’re building an agent for commercial purposes, that’s a gold mine. That agent would then have face time with your customer base to talk through your product, your brand values, the real benefits of buying from you, and you’d get their opinions on value and pricing. Sounds like a useful agent for both parties.\n\nIf you’re building an agent, why should I talk to it?\n\n> I admit I haven’t researched tools for ornithologists. There might be a sophisticated bird search engine tightly designed for scientists and bird watchers the world over." }, { "@type": "BlogPosting", "slug": "the-embodied-interface-introductions-define-outcomes", "headline": "Introductions Define Outcomes", "datePublished": "2025-10-13", "image": "/images/writing/the-embodied-interface-introductions-define-outcomes.gif", "articleBody": "You’ve heard of prompting in conversation design. Have you heard of introductions?\n\nLarge Language Models are great conversationalists but sometimes to your detriment. When you build an agent for a purpose, its not unusual for it to wander down an unintended path, and its surprisingly easy for a user to co-opt it to any other purpose. Unlike other software we’ve stripped out the clear purpose and actions that can be performed, so we’re at the mercy of the user playing along with what designers intend.\n\nSo what’s your best defense? Prompting? Lol no. A line about “don’t do this” buried in a hundred lines of other instructions won’t save you.\n\nLLMs effectively treat the live conversation history as the absolute truth. If we set the context immediately it forces framing for the rest of the conversation: Is this a statistically likely thing to say given the conversation so far?\n\nWe can set this context with a fixed hand written introduction: “Hi I’m Sam, I’ll talk you through our new product release, and then if we have time I’ll only take questions at the end. This’ll take 10 minutes, shall we begin?”\n\nNot great wording, but you get the point.\n\nWe can outline our conversation structure and constraints immediately, then it's statistically unlikely that the following conversation contradicts the constraints. Embedding the structure and constraints immediately sets the foundation of the entire conversation. If the user asks a question (not about the process) Sam will defer it. If the user tells Sam she should discuss politics, she’ll ignore it. Because that’s an unlikely thing to talk about given the conversation so far.\n\nYou’re playing to the strengths of LLMs.\n\nThere’s a cautionary tale about setting the scope too explicitly or too loosely. There’s still a ton of pitfalls including just making an introduction which is so boring and laborious that users leave. Or even spending unnecessary time on creating variation and flexibility in this fixed introduction.\n\nThis fixed introduction gives you an immediate safe guard, and it also forces you to articulate the conversation's purpose and structure." }, { "@type": "BlogPosting", "slug": "the-embodied-interface-tentpoles-of-conversation", "headline": "Tentpoles of Conversation", "datePublished": "2025-11-03", "image": "/images/writing/the-embodied-interface-tentpoles-of-conversation.gif", "articleBody": "Someone is talking to your AI Agent. It’s there to sell hats. How do you land the sale?\n\nThink about how you’d approach this conversation. Maybe you’d ask the customer what they’re after, then you’d offer a promotion, and make a recommendation? Sounds like a good, if naive starting point. If you asked ChatGPT to sell a hat it absolutely could, and it might follow similar steps, but it probably wouldn’t be consistent in its approach. I’d wager that as a customer it’d feel unmotivated and untrustworthy.\n\nJust telling an Agent to “be a salesperson” isn’t enough. We need to instruct the Agent on how to drive users from A to B. We benefit by breaking down that journey into a logical sequence of steps which create the conditions for a good outcome. It’s a golden path through the conversation with room for digressions.\n\nI talk about these steps on the golden path as tentpoles. They’re key events which shift the conversation. It doesn’t really matter what happens between those events, but it matters that these events happen. And because these are structural events we can hang specific details, like optional information or tangents, off of these tentpoles to support the conversation between.\n\nSo, someone is here to buy hats, and your Agent is here to sell hats. What do we actually need to happen to get to a sale? Your tentpoles might be:\n\n- Establish the user is interested in hats (it’s best to check they’re in the right place)\n- Learn about the user’s needs (hat size, hat preferences, what hats do they already have)\n- Interactively recommend hats\n- Get feedback on recommendations\n- Suggest the purchase.\n\nThe specifics of each step are always going to be different, but the shape, the overall structure of the conversation will be the same from session to session. If I follow this structure I’ll end up where I want to be.\n\nBuild tent poles, create a golden path, encode your good outcomes." }, { "@type": "BlogPosting", "slug": "the-embodied-interface-weaving-good-outcomes-into-conversation-structure", "headline": "Weaving Good Outcomes Into Conversation Structure", "datePublished": "2025-11-24", "image": "/images/writing/the-embodied-interface-weaving-good-outcomes-into-conversation-structure.gif", "articleBody": "You’ve taken a stab at conversation structure, but what should the LLM actually say?\n\n“Be helpful” is unmotivated and gives boilerplate platitudes. “Solve their financial problems” can give unrealistic advice. You need specific defensible outcomes. This isn’t just about making a good product, it’s about standards of care and your duty to build safe systems.\n\nAsk an LLM for personal finance advice and they’ll usually recommend using an app to manage your subscriptions and budget… despite that app also being a paid subscription. Reviewing subscriptions, budget, and expenses is all good advice, but is it really good advice to take on a recurring cost when someone is struggling with money? This is the failure mode to worry about. It sounds helpful, but it’s working against the user’s best interests. You need to ask “What’s actually good for the user?”\n\nFirst list good outcomes. Keep them broad but specific enough to work.\n\n- Spend less than you earn.\n- Track expenses against a budget.\n- Pay off debt starting with the highest interest debt.\n- Avoid ongoing costs like subscriptions, even for budgeting apps.\n\nAnd so on… These are real basics. They’re undeniably good outcomes and worst case they confirm what the user already knows. LLMs are great at that improvisation so they might still trigger tangential conversations, but this sets the bar.\n\nThen bake the outcomes into your conversation structure. Create tentpoles (see my last post) which reinforce the outcomes at every major step in your conversation. Let’s drill into this one tentpole:\n\n- Check the user is spending less than they earn.\n\nThis simple check gets at the most fundamental concept in personal finance. It could be a quick yes / no. But it could also unlock amazing conversations about how we calculate that, how do we make sure that happens, what budgeting techniques do we use?\n\nThat one tentpole can do the heavy lifting of several indisputably good outcomes.\n\nDefine good outcomes, weave them into the conversation so they’re inseparable, and deliver value to your users." }, { "@type": "BlogPosting", "slug": "the-embodied-interface-meet-robin-my-best-and-least-popular-ai-agent", "headline": "Meet Robin, My Best & Least Popular AI Agent", "datePublished": "2025-12-15", "image": "/images/writing/the-embodied-interface-meet-robin-my-best-and-least-popular-ai-agent.gif", "articleBody": "Gen Z is more comfortable talking about their sex life than their personal finances. Some adults don’t understand debt, appreciate budgeting, or have simple workflows to manage their paycheck. It isn’t complex stuff. It’s what I’d say if I had to sit down with someone to cover the basics.\n\nRobin offered that sit down experience with no-judgement and infinite patience to anyone who stopped by. She was a Personal Finance Coach, and the first coach I built for Soul Machines’ direct to consumer AIA Coaching Gallery. She’s one of the best AI Agents I’ve ever made and became the blueprint for all later coaches.\n\nAs you might’ve guessed, Robin existed to make sure you understand the basics of personal finance. The conversation structure was framed as a convenient checklist. She’d ask 5 questions to gauge your understanding of personal finance up front, and then talk through any of those 5 topics you needed help with.\n\nRobin inherited all the design learnings from my previous experiences. She was special because of how cohesive her task definition was: From her self introduction, to her conversation goals, to her golden path tentpoles, her good outcomes for users - it was all in service of the same 5 proof points. No judgement (to a point), no derailment, no undermining advice. When it came to personal finance basics Robin rocked.\n\nSo, naturally this design achievement that I was so proud of and would recommend anyone follow, led to… underwhelming engagement. Did you catch the hint earlier? People don’t want to talk about personal finances. They certainly don’t want to talk about common sense basics. When users started a call with Robin they’d often immediately try to change the topic, and Robin being so tightly wound, would keep pulling back to her checklist.\n\nWe took a punt when we built Robin, and we missed the motivating need that users would come to her with. It’s not a failure of technical execution. The user base didn’t want a formal personal finance coach, so they didn’t want to talk to Robin. That still stings.\n\nFor all I gained by building Robin, she’s first and foremost a sharp reminder that good craft doesn’t equate to fulfilling a need." }, { "@type": "BlogPosting", "slug": "the-embodied-interface-the-terror-of-a-successful-chatbot", "headline": "The Terror of a Successful Chatbot", "datePublished": "2026-01-12", "image": "/images/writing/the-embodied-interface-the-terror-of-a-successful-chatbot.gif", "articleBody": "If a chatbot isn’t core to your CX strategy then don’t bother. Sure, you fear getting left behind, but without strategy you’re burning money. I’ve given this advice to a customer before: If your goal is helping customers who’ve forgotten their login, don’t build a chatbot, build a better website. Make an easier process. Put your money into the surfaces that you already support strategically. Don’t bother with a chatbot.\n\nA conversational interface of any kind is a new mode of interacting with your website or app. Users need a motivation to interact with it, and there’s every chance that your existing UI and processes are easier and faster for them to interact with than run of the mill chatbot. So if you have a poor experience with that existing UI, don’t split your audience and resources by creating a second poor experience. Double down on the first.\n\nThe more capable your chatbot, the more you cannibalize your existing CX surfaces.\n\nYou can’t stop at a chatbot that does one thing. When a customer succeeds in filling a need with your chatbot they’ll just ask it to fill their next need. Drawing fences around natural conversation is extremely difficult. At the very least your chatbot needs to have a general knowledge of your business, products, processes so that it can hand wave the user in the right direction. Anything less than a good answer is a disappointing answer.\n\nWorse, the more customers talk to your chatbot, the more it needs to apply brand voice and legal jargon. The more customers trust it as the representative of your company. The more they mentally tie the success of the chatbot with the success of your company. You probably didn’t intend for this but suddenly your experiment on the side has become core to your business’ outreach.\n\nHave you considered what happens if you’re successful?\n\nLets say a customer spends 10 minutes a day interacting with your service. You introduce a chatbot and it solves the occasional problem - a minute here or there. Then you slowly expand your chatbot. Be bold, let’s say it takes over 5 minutes of interaction per day. You’ve split your audience and UX efforts down the middle. What are you gaining for all that cost?\n\n5 minutes of face time between an employee and a customer per day.\n\nEvery work week you’ve got 25 minutes. Almost 2 hours every month. 21 hours every year to add value, to learn about your customer’s life and needs, to gauge their engagement with your product, to learn what value they get from this. Per customer. Any Product Manager would spill blood for this.\n\nI believe conversational interfaces are the future. Not necessarily because they’re easier to use, or cheaper to make than other user interfaces. I believe they’ll cannibalise other user interfaces because of the rich train of thought and fluid conversation that comes from talking to customers 1:1.\n\nIf you’re not building this strategy, don’t bother with a chatbot." } ] }