Congrats everyone, we did it. AI prototypes are incredibly easy to build.
This brings us to our next, less shiny, phase: making AI implementations lasting and impactful to your product and business.
At this year’s SXSW conference, even if you weren’t in the AI track, you were in the AI track. Across sessions led by founders, researchers, and developers, one theme came up again and again: flashy demos aren’t enough. If you want to create long-term value with AI, you need to move beyond proof of concept and build systems that are accurate, reliable, and integrated with the real world.
Below I’ve pulled from the insights I gleaned at three SXSW 2025 sessions:
- “Beyond proof of concept: Making AI useful for business” by Richard Socher, founder of You.com
- “AI essentials: Practical skills to get started” by Tom Hewiston of General Purpose
- “How to be a smarter AI user” by Maarten Sap and Sherry Tongshuang Wu, researchers from Carnegie Mellon University
Each session focused not on the next big model, but on how to thoughtfully build AI into your workflows, products, and strategy.
For full transparency and to get meta with it, I created an outline of this blog post by loading all of my AI session notes into ChatGPT4o and working through some prompts to land on the overarching themes and sections for two recap posts. I’m using AI to enhance my content creation abilities and speed the outlining and drafting process, not replace me as a writer. The session speakers would be so proud.
The myth of the AI prototype
Richard Socher of You.com kicked off his session by saying: “It’s never been easier to build a prototype using AI.” With tools like ChatGPT, Claude, and Replit, nontechnical founders can go from idea to mockup in an hour - and that’s exactly what he proceeded to do with his allotted time. (Stay tuned, because I am dying to attempt to recreate his steps for a future episode of AI in Focus.)
The ability to brainstorm product ideas, generate specs, and build a landing page or prototype using agents and assistants is real and oh boy, dare I say, transformative. But there’s a catch.
While the barrier to entry is lower than ever, most prototypes are just that: surface-level demos that break down when scaled, require too much oversight, or fail to solve a real need. The leap from prototype to product requires grounding your AI in factual data, aligning it with human goals, and integrating it into core workflows.
In other words, to make AI useful, you need to treat it like more than a tool. You need to treat it like a team member.
Think like an AI manager, not an AI magician
Tom Hewiston, whose company trains nontechnical teams to use AI tools, offered a practical mental model. He advised thinking of your AI assistant like a “bright, enthusiastic intern you can only talk to over Slack.” It’s fast, capable, and has boundless energy for tasks. But to be useful, it needs structure, clarity, and oversight.
Tom argues that being effective with AI isn’t about technical skill. It’s about being a good manager.
For example:
- Want help writing marketing copy? Feed the model your last 50 headlines and tell it what’s worked.
- Want help coding? Provide a spec, test cases, and feedback as if you were assigning the task to a junior dev.
- Want a competitive landscape analysis? Ask for one, but verify the data. AI can hallucinate confidently.
The heavy users of AI aren’t just better at prompting. They’re better at defining what they want, iterating, and managing output.
The AI “last-mile” problem is the real opportunity
Many SXSW speakers used the same metaphor: AI today is great at getting you 80 percent of the way there. But that last 20 percent - contextual relevance, personal preferences, business logic - is where the category royalty will be crowned.
Richard described this as the “last mile” problem, and it’s the gap that thoughtful teams need to close. For example:
- Booking a trip with an AI agent sounds simple, but actually requires dozens of preferences: budget, flight time, walking tolerance, hotel style, activities, and more.
- Writing an email with an AI assistant is easy, but tailoring it to your brand voice, customer segment, and current campaign takes effort.
- Using an LLM to summarize internal documentation only works well if the structure of the data is right and the model can cite accurately.
The products that can solve these last-mile problems, either through smart interfaces, better data design, or hybrid workflows, will be the market winners.
Use cases are the new AI differentiator
One striking insight came from the Carnegie Mellon researchers Maarten and Sherry: LLMs aren’t as smart as they seem. They’re literal, not great at inferring intent, and often overconfident. They don’t know when they’re wrong, and they’ll rarely tell you.
In this context, AI isn’t a magic solution. It’s a high-powered generalist tool that gets its value from the use case you give it (ie I can keep my job safe from our robot overlords and you probably can too).
That’s good news for founders and teams who know their customers well. Your advantage isn’t the model you use. It’s the problem you solve.
That means your real differentiator comes from:
- The quality and structure of your data
- The clarity of the user need you’re meeting
- The workflows you design around the AI
- The way you measure and improve performance over time
The question isn’t “What can this model do?” It’s “What should it do for my product, and how do I help it do that well?” Our designers and developers conduct AI discovery sprints with client teams to help them hone in on just this.
From flashy to foundational
The sessions at SXSW 2025 made it clear: we’ve moved past the phase of “wow” moments that took center stage at SXSW 2024. The real opportunity now is in thoughtful, well-scoped, and useful AI integration. Is that boring? To some, maybe. Honestly, I did hear many comments about the tech sessions this year being “boring” and “not revolutionary”.
I think I’m joined by many of my colleagues in thinking the AI innovations are delightfully boring. At long last the AI hype machine is grounded in reality and focusing on real-life, incremental applications.
Whether you’re a founder validating your first idea or a team expanding your product’s capabilities, the game has shifted. Use AI like you’d manage a junior team. Build workflows that enhance, not just automate. Focus on real needs, real users, and real outcomes. That’s how you turn a prototype into something powerful.
If you want help figuring out your best AI use cases and a plan forward, let’s chat! We’d love to help you make your AI initiative a success.
