<?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 Elvia Perez on Medium]]></title>
        <description><![CDATA[Stories by Elvia Perez on Medium]]></description>
        <link>https://medium.com/@elvia_perez?source=rss-5d512ed77981------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Na-Han3AVQAB3F4wsPHULA.jpeg</url>
            <title>Stories by Elvia Perez on Medium</title>
            <link>https://medium.com/@elvia_perez?source=rss-5d512ed77981------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 00:25:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@elvia_perez/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[Why We’re Bullish on Vertical AI in 2025 and How These Companies Create Long-Term Value]]></title>
            <link>https://medium.com/@elvia_perez/why-were-bullish-on-vertical-ai-in-2025-and-how-these-companies-create-long-term-value-c38fbe8055a6?source=rss-5d512ed77981------2</link>
            <guid isPermaLink="false">https://medium.com/p/c38fbe8055a6</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[vertical-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <dc:creator><![CDATA[Elvia Perez]]></dc:creator>
            <pubDate>Sat, 11 Oct 2025 21:29:30 GMT</pubDate>
            <atom:updated>2025-10-11T21:29:30.380Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Irl6n-noXB9Vt00H" /><figcaption>Photo by <a href="https://unsplash.com/@kilgrim?utm_source=medium&amp;utm_medium=referral">fabio Spano</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction: From Systems of Record to Systems of Action</h3><p>Enterprise software is entering a new era. For years, horizontal SaaS platforms like Salesforce and Workday defined the landscape by organizing workflows and storing information. But in 2025, the frontier is no longer software that records activity; it’s software that acts.</p><p>This evolution, known as Vertical AI, represents the next generation of intelligent enterprise systems. These tools don’t just assist workers; they understand industry context, make decisions, and increasingly perform tasks once reserved for human experts. As Andrii Brodetskyi notes in <em>The Next Trillion-Dollar Opportunity: Why Vertical AI is the Future of SaaS</em>, software is evolving “<em>from a system of record to a system of actions.</em>”</p><p>The shift is both technological and economic in nature. Vertical AI companies are building specialized intelligence for specific sectors.</p><h3>The Depth Advantage: How Vertical AI Redefines Enterprise Software.</h3><p>We’re drawn to Vertical AI because it solves a problem that general-purpose AI and horizontal SaaS never fully addressed: depth. These new systems combine large language models with domain-specific knowledge, creating precision, reliability, and measurable ROI in industries that once resisted automation.</p><p>The opportunity is massive. Brodetskyi estimates that while traditional SaaS captures just 1–5% of an employee’s value, Vertical AI can capture 25–50% by automating high-value, cognitive work from drafting contracts to processing insurance claims. That expands the addressable market from a $450 billion software industry to an $11 trillion labor market.</p><p>As Euclid Ventures puts it, “<em>The future of AI is specialized.</em>” The most valuable companies will be those that apply intelligence to real-world problems, integrating deeply into professional workflows, handling unstructured data, and delivering measurable outcomes.</p><h3>Building Moats: What Makes Vertical AI Companies Durable</h3><p>Vertical AI’s long-term value lies in its defensibility. These companies don’t compete on size or hype. They compete on precision, embedded workflows, and proprietary data.</p><p>As Alex Niehenke from Scale Venture Partners explains in <em>The Next Decade of Software is Verticals and AI</em>, vertical markets were never “slow adopters.” They lacked software that justified the switch. “<em>Solutions weren’t providing enough value, so they didn’t succeed.</em>”</p><p>Now, AI provides the breakthrough. These systems deliver step-change ROI, automating work instead of just digitizing it. By “<em>capturing the value of labor rather than software spend,</em>” as Niehenke notes, even small vertical markets can now support billion-dollar outcomes.</p><p>Regulation, once a barrier, is becoming a moat. Compliance-heavy sectors like healthcare, law, and finance demand explainability and audit trails. As Niehenke writes, “<em>Using OpenAI to make a loan decision is not possible today, since it cannot provide an audit on bias for regulators.</em>” Vertical AI companies that bake compliance into their logic are better positioned to win and to keep winning.</p><h3>The Compounding Economics of Vertical AI</h3><p>From an investor’s standpoint, the appeal is clear: capital efficiency and rapid monetization.</p><p>Mike Schatzman, Ohad Tzur, and Kasra Davar highlight this in <em>Why Investors Should Pay Attention to Vertical and Agentic AI</em>, noting that early-stage vertical AI companies are growing 400% year-over-year, maintaining 65%+ gross margins, and reaching 80% of traditional SaaS contract values, all on seed rounds of $2–8 million.</p><p>They’re not chasing foundation-model scale. They’re executing with focus. As Bessemer Venture Partners observes, “<em>In AI, access to proprietary data isn’t just an advantage, it’s a barrier to competition.</em>” These models learn from specialized data sources that others can’t access, and they embed directly into daily workflows, making them difficult to dislodge.</p><p>This embeddedness also drives retention. DelMorgan &amp; Co., in <em>Vertical AI Startups: Why Investors Are Paying Close Attention</em>, reports that “<em>Vertical AI companies with strong workflow integration can achieve customer retention rates 30–50% higher than their horizontal counterparts.</em>” Each workflow completed strengthens the model, compounds the data advantage, and reinforces the moat.</p><h3>The Rise of Vertical AI Agents</h3><p>2025 is the year of agents: autonomous, domain-trained systems that act on behalf of professionals.</p><p>As Anjali Chaudhary explains in <em>Vertical AI Agents: Unlocking Efficiency, Automation, and Scale</em>, “<em>The era of general-purpose AI and traditional SaaS solutions is giving way to vertical AI agents — intelligent systems purpose-built for specific industries.</em>” These agents don’t just assist, they execute.</p><p>The vertical agent era has arrived, and its impact is already visible across real-world applications: Abridge turns doctor–patient conversations into structured clinical notes. EvenUp drafts legal demand letters. Feedzai processes trillions in payments while detecting fraud in real time. Gartner predicts that by 2026, 80% of enterprises will adopt vertical agents, and Bessemer expects the category’s market cap to grow 10x larger than legacy SaaS.</p><h3>How Vertical AI Builds Long-Term Value</h3><p>The long-term playbook for Vertical AI is consistent across every perspective. The best companies own the action layer, acting first on raw, unstructured data before it reaches legacy systems. They price on outcomes, charging per claim, report, or transaction rather than per seat, aligning revenue directly with customer success. They embed deeply within core workflows, creating switching costs and feedback loops that continuously refine their models.</p><p>They also codify compliance, building regulation into their architecture rather than treating it as an afterthought. And above all, they compound data. Each workflow completed improves accuracy, strengthens insight, and attracts more customers, creating a cycle of learning and defensibility. As Andrii Brodetskyi writes, these companies do not sell software; they sell industry-grade intelligence that learns and scales with every interaction.</p><p>Over time, Vertical AI evolves from tools to infrastructure. By embedding intelligence into mission-critical systems, these platforms become the operational backbone of entire industries. Each iteration deepens integration and strengthens the models that guide real-world decision-making. As they move from automating tasks to orchestrating outcomes, they capture a growing share of industry profit pools, shifting value from human labor to intelligent execution at scale.</p><p>The most successful Vertical AI companies will not just make industries more efficient. They will define how those industries operate. Their systems will guide underwriting, diagnose patients, draft contracts, and manage logistics, becoming the decision layer that modern enterprises depend on. In doing so, they will create lasting value measured not in efficiency gains, but in ownership of the workflows that power the global economy.</p><h3>Conclusion: The Future Belongs to Vertical AI</h3><p>We like Vertical AI because it represents the most practical and enduring form of applied intelligence. These companies are solving complex, regulated, high-friction problems with precision, trust, and measurable results.</p><p>General-purpose AI will keep advancing, but the true winners will be those who understand their industries as deeply as their models. These companies won’t just automate workflows; they’ll automate decisions, becoming the intelligent operating systems that power modern enterprise.</p><p>The next generation of software giants will not be generalists. They will be specialists, companies that combine domain mastery with machine intelligence to reshape how industries operate at their core. Their impact will extend far beyond efficiency, redefining how value is created, captured, and compounded across the global economy.</p><h3>References</h3><p><strong>Brodetskyi, Andrii.</strong> <em>The Next Trillion-Dollar Opportunity: Why Vertical AI is the Future of SaaS.</em> Medium, 2025.</p><p><strong>Niehenke, Alex.</strong> <em>The Next Decade of Software is Verticals and AI.</em> Scale Venture Partners, January 2025.</p><p><strong>Schatzman, Mike; Tzur, Ohad; Davar, Kasra.</strong> <em>Why Investors Should Pay Attention to Vertical and Agentic AI.</em> Venture Forward Capital, August 2025.</p><p><strong>DelMorgan &amp; Co.</strong> <em>Vertical AI Startups: Why Investors Are Paying Close Attention.</em> DelMorgan &amp; Co. Insights, July 2025.</p><p><strong>Chaudhary, Anjali.</strong> <em>Vertical AI Agents: Unlocking Efficiency, Automation, and Scale.</em> Turing, May 2025.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c38fbe8055a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Five Percent Playbook: How to Turn Generative AI Pilots into Profit]]></title>
            <link>https://medium.com/@elvia_perez/the-five-percent-playbook-how-to-turn-generative-ai-pilots-into-profit-1a6524e97b9b?source=rss-5d512ed77981------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a6524e97b9b</guid>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[mit]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Elvia Perez]]></dc:creator>
            <pubDate>Sun, 21 Sep 2025 06:39:57 GMT</pubDate>
            <atom:updated>2025-09-21T07:18:49.832Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OnmqZVt7O5YiLuV3" /><figcaption>Photo by <a href="https://unsplash.com/@muzammilo?utm_source=medium&amp;utm_medium=referral">Muzammil Soorma</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction: The GenAI Divide</h3><p>Enterprises have poured <strong>$30–40 billion</strong> into pilots and proof-of-concepts. Yet according to MIT’s new report <em>State of AI in Business 2025</em>, <strong>95 percent of these pilots fail to produce measurable ROI</strong>.</p><p>MIT’s analysis makes an important distinction: the failures are <strong>not about the AI technology itself</strong>. The core models are powerful and ready.</p><blockquote><strong>The breakdown comes from poor implementation, brittle workflows, and a lack of integration into daily operations.</strong></blockquote><p>The key takeaway is: the challenge is <strong>integration and governance</strong>, not the technology, and the right response is a more pragmatic approach to adoption rather than retreat.</p><p>At first glance, the 95 percent failure rate sounds like a bubble about to burst. But the report tells a different story. The problem isn’t the models. It isn’t regulation. The real issue is how companies implement and integrate AI into daily work.</p><p>MIT calls this gap the <strong>GenAI Divide</strong>. On one side are organizations stuck in high adoption but low transformation: they deploy generic chatbots for demos and simple tasks but never rewire the workflows that create true value. On the other side is a small group — just <strong>five percent</strong> — that embraces friction as a design principle and builds adaptive systems with memory, feedback, and deep workflow integration.</p><p>This is the story of that five percent, and the investor and founder playbook for crossing the divide. The headline number “95 percent failure rate” made waves precisely because it fits a popular narrative of AI hype. But MIT’s research is far more nuanced.</p><blockquote>The report highlights that <strong>most generative AI tools don’t retain feedback, adapt to context, or improve over time</strong>. They look impressive in a demo but remain brittle when exposed to compliance hurdles, messy data, and human judgment.</blockquote><p>It also has a striking shadow economy. While only <strong>40 percent of companies have official AI subscriptions</strong>, <strong>90 percent of employees already use personal AI tools at work</strong>. This bottom-up adoption is quietly saving some firms <strong>$2 million to $10 million a year</strong> in reduced outsourcing and agency costs, even when official pilots are labeled failures.</p><p>The lesson is clear: <strong>technology isn’t the barrier. Integration and governance are.</strong> Success depends on treating AI not as a plug-and-play app but as infrastructure that must learn, remember, and fit into the grain of everyday operations.</p><h3>Friction as a Feature, Not a Flaw</h3><p>MIT researchers describe a fundamental paradox. <strong>Pilots that glide frictionless from demo to deployment never build the muscle to scale.</strong> They crumble when they meet the texture of real enterprise life, politics, approvals, legacy systems, and compliance.</p><p><strong><em>By contrast, the five percent of winners lean into resistance.</em></strong></p><p>Rick Kiley, founder of Soho Experiential, captures the idea:</p><blockquote>“No behavioral change is possible without friction. People don’t come in thinking one thing and walk out thinking another unless they’re challenged. In experiences, friction is what creates the memory. GenAI is no different. If it’s too smooth, it fades. If it pushes you, it sticks.”</blockquote><p>Industries where mistakes carry high cost, such as aerospace, healthcare, finance, are already adopting this <strong>humility-first design</strong>. They see friction as the raw material for trust.</p><h3>From Generic Tools to Adaptive Systems</h3><p>Most enterprises still deploy <strong>generic chatbots</strong> that answer questions but can’t learn. They achieve early adoption rates of more than 80 percent for small tasks but stall when workflows demand context, compliance, or persistent memory.</p><p>The five percent that scale past pilots do something different. They <strong>engineer GenAI into high-value workflows</strong> and build <strong>learning loops</strong>:</p><ul><li><strong>Memory and feedback layers</strong> allow systems to retain context and improve over time.</li><li><strong>Governance and data boundaries</strong> ensure trust and regulatory compliance.</li><li><strong>Outcome-based contracts</strong> incentivize vendors to deliver measurable learning, not just seat licenses.</li></ul><p>Instead of eliminating resistance, these companies <strong>calibrate it</strong>. Every moment of friction, whether human, organizational, or technical, becomes a signal to redesign processes and harden the product.</p><h3>Where the 5 Percent Focus</h3><p>The MIT study finds that most GenAI spending still flows to sales and marketing because it is easy to measure engagement metrics. But <strong>real ROI hides in the back office</strong>:</p><ul><li><strong>Claims processing</strong> and <strong>accounts payable/receivable</strong></li><li><strong>Procurement and contract review</strong></li><li><strong>Compliance reporting and risk monitoring</strong></li></ul><p>These are friction-heavy zones where automation yields <strong>direct cost savings and rapid payback periods</strong>. MIT estimates that back-office deployments can save millions annually and cut agency spend by <strong>30 percent or more</strong>.</p><p>Successful firms also build <strong>external partnerships</strong> rather than attempting purely internal development. According to the report, <strong>custom solutions developed with specialized vendors are twice as likely to succeed</strong>.</p><h3>A Practical Playbook for Founders and Investors</h3><p>MIT distills these lessons into a pragmatic framework that every founder and investor should internalize.</p><ul><li><strong>Measure absorption, not adoption<br></strong> Don’t count logins. Count the number of workflows redesigned and the percentage of decisions now automated and auditable.</li><li><strong>Fund the memory layer<br></strong> If a system can’t retain context and learn, it cannot scale beyond a demo.</li><li><strong>Redesign contracts<br></strong> Demand that vendors price against learning milestones and ROI outcomes, not monthly seats.</li><li><strong>Channel shadow GenAI<br></strong> Instead of banning unofficial employee tools, formalize and secure them. Grassroots use is often the first sign of true value.</li><li><strong>Treat friction as a signal<br></strong> If a pilot feels too smooth, it is probably too shallow.</li></ul><p>These principles echo a larger shift: <strong>the next decade of AI will be about designing friction wisely, not erasing it.</strong></p><h3>What This Means for Investors</h3><p>From a venture perspective, several themes stand out.</p><ul><li><strong>Focus on adaptive systems<br></strong> Agentic AI, systems with persistent memory and feedback loops, is emerging as the architecture of choice. Thin wrappers on large language models will not endure.</li><li><strong>Real ROI lives in the back office<br></strong> Despite the glamour of customer-facing tools, the biggest economic gains are hidden in procurement, finance, and operations.</li><li><strong>Distribution matters more than demos<br></strong> MIT’s interviews show that <strong>referrals, procurement relationships, and board-level introductions</strong> predict adoption far better than feature checklists.</li><li><strong>The window is closing<br></strong> Enterprise AI budgets are consolidating. The next few quarters will lock in vendor relationships that are hard to unwind. Founders who can secure early footholds will define the stack for years.</li></ul><h3>Guidance for Founders</h3><p>For founders building in this space, the implications are equally sharp.</p><ul><li><strong>Start narrow, expand wide<br></strong> Win a single workflow with deep customization before scaling horizontally.</li><li><strong>Build for integration, not replacement<br></strong> “If it doesn’t plug into Salesforce or our internal systems, no one’s going to use it,” one executive told MIT researchers. The lesson: design to fit, not to disrupt.</li><li><strong>Sell like a services company<br></strong> Enterprise buyers evaluate AI more like consulting or BPO than SaaS. They expect proof of value and measurable outcomes.</li><li><strong>Embrace agentic design<br></strong> The next generation of systems will “think, see, and do”, retaining memory, adapting in real time, and collaborating with other agents. Building these capabilities early creates a durable moat.</li></ul><h3>Looking Ahead: Agentic GenAI and the Next Wave</h3><p>MIT’s report closes with a forward-looking insight: <strong>the next phase of enterprise AI will be agentic. </strong>These systems will not just generate content; they will <strong>discover tasks, negotiate with other agents, and coordinate actions across networks</strong>.</p><p>The shift will challenge companies even more to design for friction. Agents with long-term memory and autonomy require careful governance, clear audit trails, and thoughtful human-in-the-loop oversight. Those that succeed will turn AI from a demo into the backbone of their operations.</p><h3>Thank you so much for reading!</h3><p>I truly enjoy connecting with others who are exploring this space, and I’d love to hear your thoughts, comments, or suggestions for future articles. If you’re building in AI, I’d be happy to connect. You can email me at <strong>elvia@lasolasvc.com</strong>.</p><h3>References and Further Reading</h3><ul><li><a href="https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/"><strong>MIT Finds 95% of GenAI Pilots Fail Because Companies Avoid Friction</strong> </a>— Jason Snyder, <em>Forbes</em></li><li><strong>The GenAI Divide — State of AI in Business 2025</strong> — MIT Report</li><li><a href="https://www.marketingaiinstitute.com/blog/mit-study-ai-pilots"><strong>That Viral MIT Study Claiming 95% of AI Pilots Fail? Don’t Believe the Hype.</strong></a> — Mike Kaput</li><li><a href="https://hbr.org/2025/08/beware-the-ai-experimentation-trap"><strong>Beware the AI Experimentation Trap</strong></a> — Nathan Furr and Andrew Shipilov, <em>Harvard Business Review</em></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a6524e97b9b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Defensibility in the Age of AI: Five Attributes Every Vertical Startup Needs]]></title>
            <link>https://medium.com/@elvia_perez/defensibility-in-the-age-of-ai-five-attributes-every-vertical-startup-needs-14f0d99eff60?source=rss-5d512ed77981------2</link>
            <guid isPermaLink="false">https://medium.com/p/14f0d99eff60</guid>
            <category><![CDATA[vertical-ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Elvia Perez]]></dc:creator>
            <pubDate>Mon, 01 Sep 2025 03:40:15 GMT</pubDate>
            <atom:updated>2025-09-01T15:29:28.331Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2Ux7CAcGtILdCPQ0" /><figcaption>Photo by <a href="https://unsplash.com/@omilaev?utm_source=medium&amp;utm_medium=referral">Igor Omilaev</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The proliferation of AI startups has created a paradox: it has never been easier to launch a product, and it has never been harder to build one that lasts. With open-source LLMs approaching parity with proprietary models, MVPs being built in days, and well-funded startups and incumbents ready to follow fast, every founder and investor must now ask: What makes this business truly defensible?</p><p>I spent this past weekend diving into pieces from Euclid Ventures, Bowery Capital, NFX, and Abraham Thomas, and a common thread emerged: five attributes consistently show up as the foundation of defensibility in vertical startups.</p><h3>1. Workflow Ownership — The Authoring Layer</h3><p>Euclid Ventures, in <em>Emerging Playbooks in Vertical AI</em>, emphasizes the ‘authoring layer’, the point where critical work begins. This is where the first inputs are created, like a doctor dictating notes, a lawyer drafting a contract, or a contractor measuring a roof. Because everything else in the process flows from this starting point, the software becomes the foundation of the workflow rather than just a bolt-on tool. In other words, the product isn’t an add-on; it <em>is</em> the workflow itself, which makes it deeply embedded and hard to replace.</p><blockquote>Startups that capture the very first step of work (where information is created) become the foundation of the workflow. Because everything else flows from that starting point, the product becomes essential and deeply embedded, not just a nice add-on. By embedding at the source of value creation, startups capture structured data and become indispensable. As Euclid notes, this early wedge doesn’t just automate, it sets the stage for a broader <strong>System of Intelligence</strong> where the platform evolves into the daily operating hub for an industry.</blockquote><p>A <strong>System of Intelligence </strong>is software that doesn’t just store information (like a database or spreadsheet). <em>It actively analyzes information, learns from context, and offers insights that help users make smarter decisions.</em> When a startup builds this kind of system, its product often becomes the central hub.</p><p><strong>Founder lesson:</strong> Don’t just integrate around workflows. Own the entry point where work gets done.</p><h3>2. Data Gravity — Control and Loops</h3><p><em>If your product gathers and holds a lot of valuable data, it not only becomes sticky and hard to replace, but it also enables compounding feedback loops that make the product better and more defensible over time.</em></p><p>Abraham Thomas, co-founder of Quandl, reframes data defensibility in <em>Data and Defensibility</em>. He describes two key moats:</p><ul><li><strong>Data Control:</strong> exclusive access to proprietary or fragmented datasets.</li><li><strong>Data Loops:</strong> feedback cycles where usage generates better data, which improves the product, attracting even more usage.</li></ul><p>For example, a Vertical AI platform that captures sensitive medical records, legal documents, or logistics data directly from its users is building a moat through control, since competitors can’t easily access the same datasets. At the same time, if every new customer interaction improves the system’s accuracy or insights, that creates a loop: the product grows stronger with scale, which then attracts even more users. Over time, these effects compound into a powerful advantage that’s hard to replicate.</p><p><strong>Founder lesson:</strong> Build your product so every interaction generates data that competitors can’t easily replicate.</p><h3>3. Distribution and Ecosystem Lock-In</h3><p>As Euclid points out in <em>Dude, Where’s My Moat?</em>, defensibility doesn’t always come from the model. It often comes from distribution, such as integrations with legacy platforms, compliance rails that reduce procurement risk (the built-in rules, guardrails, and processes that ensure software meets legal, regulatory, and security requirements, making it easier and less risky for a company to buy and adopt), or partnerships that become difficult to unwind.</p><p>Pete Flint of NFX adds a useful analogy: think of early-stage distribution as the <strong>bailey</strong>, characterized by fast-moving, brand-building, and customer-acquiring momentum. True defensibility, however, lies in the <strong>motte</strong>, the harder-to-replicate advantages like switching costs, ecosystem lock-in, and customer trust.</p><p>The <strong>bailey</strong> -&gt; early-stage tactics that help you grow quickly (distribution hacks, brand buzz, fast customer acquisition). They’re important for momentum, but not very defensible, as competitors can copy them.</p><p>The <strong>motte</strong> -&gt; the deep, durable advantages (switching costs, data moats, customer trust, ecosystem lock-in). These take longer to build, but once you have them, they protect the business over the long term.</p><p><strong>Founder lesson:</strong> Distribution isn’t an afterthought. You should design for channels, compliance, and partnerships in mind.</p><h3>4. Domain Expertise and Trust</h3><p>Loren Straub, in <em>Rethinking VC Evaluation</em> (Bowery Capital), observes that technical sophistication is no longer scarce as AI has democratized that advantage. Instead, <strong>domain expertise + AI-native execution</strong> has become the new moat.</p><p>In industries like healthcare, finance, and law, accuracy and compliance are non-negotiable. Startups with deep domain knowledge build trust that outlasts any single innovation. Founders who have lived the workflows of their industry know the real pain points and can design solutions that actually fit. This credibility not only helps win early customers, but it also shortens sales cycles and makes the product harder to replace. Over time, domain expertise becomes as important a moat as the technology itself.</p><p><strong>Founder lesson:</strong> Domain credibility is not optional. It shows founders deeply understand their industry’s workflows and can design solutions to real problems they’ve experienced.</p><h3>5. Speed with Discipline</h3><p>Speed is still essential, but to a certain extent. James Currier of NFX argues in <em>Speed &amp; AI</em>:</p><blockquote>“Don’t be the person who gets it right, be the person who learns faster than anyone else.”</blockquote><p>Straub pushes this further: the best AI startups now ship new features within <strong>24–48 hours</strong> of customer feedback. This quick cycle creates impactful moments, keeps customers engaged, and builds early momentum.</p><p>But as Euclid reminds us, speed alone is not durable. At scale, companies slow down.</p><blockquote>The moat comes from what you do with that speed, using it to capture workflows, generate proprietary data, and establish trust.</blockquote><p><strong>Founder lesson:</strong> Move fast to gain an edge, but channel that momentum into moats that endure.</p><h3>My Perspective: What Founders and Investors Should Be Thinking About</h3><p>For <strong>founders</strong>, the approach is clear:</p><ul><li>Start with a wedge in a revenue-critical workflow. -&gt; Launching with a workflow tied directly to revenue ensures customers feel the impact immediately, making adoption much easier. It also positions your product as a “must-have” rather than a nice-to-have.</li><li>Capture proprietary, first-party data from day one. -&gt; Data collected directly from customer use creates a compounding advantage that competitors can’t replicate. Over time, this data strengthens your product, improves performance, and becomes the foundation of defensibility.</li><li>Bake in compliance, trust, and transparency early. -&gt; In regulated or high-stakes industries, trust is important. Building compliance and transparency from the beginning not only prevents setbacks but also differentiates you as a reliable partner.</li><li>Leverage distribution and partnerships to scale adoption. -&gt; Partnerships, integrations, and distribution channels accelerate growth, reduce friction, and lock you deeper into your industry ecosystem.</li><li>Evolve toward a <strong>System of Intelligence</strong> that becomes the daily operating hub for your vertical. -&gt; Founders need to evolve toward a System of Intelligence because a single workflow wedge is easy to copy, but a platform that becomes the daily operating hub for an industry is far harder to replicate. By centralizing where work begins and decisions are made, the product captures proprietary data, embeds deeply in operations, and creates high switching costs. In short, it transforms from a useful tool into indispensable infrastructure.</li></ul><p>For <strong>investors</strong>, evaluation must change. Diligence should focus on:</p><ul><li>Does this product own the authoring layer? -&gt; Owning the authoring layer means the product is used at the very start of a critical workflow, like writing clinical notes or logging a claim. If you control where work begins, you become foundational to the process and harder to replace.</li><li>Are there clear data control and loop dynamics? -&gt; Defensible AI businesses don’t just use public data; they capture unique, proprietary datasets. If customer usage improves the product and generates feedback loops, the advantage compounds over time.</li><li>How strong are regulatory, distribution, or integration moats -&gt; Regulatory approvals, ecosystem partnerships, and deep integrations create high switching costs. These moats slow competitors down and make it riskier for customers to switch providers.</li><li>Does the team bring domain expertise -&gt; Surface-level knowledge isn’t enough in high-stakes industries like healthcare, finance, or law. Founders need lived experience or deep insight into the workflows, regulations, and pain points that outsiders can’t quickly replicate.</li><li>Are they using speed to compound into moats? -&gt; Speed is only valuable if it builds toward lasting defensibility. Investors should check whether speed is generating proprietary data, customer trust, and workflow lock-in.</li></ul><h3>Closing Thought</h3><p>The AI boom has made it easy to launch a product, but far harder to build one that endures. In Vertical AI, the real winners won’t just move fast, they’ll earn trust, capture data that compounds, and transform everyday workflows into platforms their industries can’t live without.</p><h3>Thank you so much for reading!</h3><p>I truly enjoy connecting with others who are exploring this space, and I’d love to hear your thoughts, comments, or suggestions for future articles. If you’re building in AI, I’d be happy to connect. You can email me at <strong>elvia@lasolasvc.com</strong>.</p><h3>References</h3><ol><li><a href="https://insights.euclid.vc/p/dude-wheres-my-moat">Dude, Where’s My Moat?</a> by Euclid Ventures</li><li><a href="https://insights.euclid.vc/p/emerging-playbooks-in-vertical-ai">Emerging Playbooks in Vertical AI</a> by Euclid Ventures</li><li><a href="https://www.nfx.com/post/speed-and-ai">Speed X AI</a> by James Currier</li><li><a href="https://bowerycap.com/blog/insights/rethinking-vc-evaluation">Rethinking VC</a> Evaluation by Loren Straub</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14f0d99eff60" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The True Cost of AI: How Unit Economics Will Define the Next Generation of Startups]]></title>
            <link>https://medium.com/@elvia_perez/the-true-cost-of-ai-how-unit-economics-will-define-the-next-generation-of-startups-ae264e0076e2?source=rss-5d512ed77981------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae264e0076e2</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[unit-economics]]></category>
            <dc:creator><![CDATA[Elvia Perez]]></dc:creator>
            <pubDate>Sun, 24 Aug 2025 04:43:08 GMT</pubDate>
            <atom:updated>2025-08-24T19:02:16.240Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fUJBHG-gT3lG6rpcWEOOmQ.jpeg" /></figure><p>AI is facing a unit economics crisis, where efficiency will matter just as much as innovation. The true winners of this era will be the companies that create value while building sustainable businesses.</p><h3>The Hidden Economics of AI</h3><p>When people think of AI, they usually imagine software. In reality, AI is an <strong>ecosystem of expensive moving parts</strong>:</p><ul><li><strong>Software (the model):</strong> The program that takes input (like a question) and produces an output (like an answer), essentially the “brain” of AI.</li><li><strong>Hardware (the GPUs):</strong> The advanced computer chips that give AI the raw computing power to process massive amounts of information at once.</li><li><strong>Infrastructure (Cloud, On-Prem, or APIs):</strong> The strategic decision of how to access that compute power, renting through APIs or cloud platforms vs. owning in-house systems, each with major cost trade-offs.</li><li><strong>Data:</strong> The information AI is trained on accuracy and quality of this data largely determines how well the system performs.</li><li><strong>Energy:</strong> The significant electricity required to run AI systems, especially during training, makes them far more resource-intensive than traditional software.</li><li><strong>Integration:</strong> The step of embedding AI into existing systems and workflows so it drives real outcomes rather than staying a stand-alone tool.</li></ul><h3>Data: The Most Defensible Asset</h3><p>One of the most critical aspects of AI is data. The accuracy and value of any system are only as strong as the quality of the data it is trained on. Clean, well-structured datasets produce reliable results, while biased or fragmented data can cripple even the most advanced models.</p><p>This is why many investors view <strong>data moats</strong>, proprietary, high-quality, and often regulatory-grade datasets, as more defensible than the models themselves. The companies that win will be those that treat data as a core strategic asset.</p><p>It’s also worth noting that data is not free. In industries like healthcare and finance, acquiring, cleaning, and maintaining compliant datasets can be one of the most expensive inputs, but it’s also what creates lasting differentiation.</p><blockquote>For investors, the critical question is: <em>what data advantage does this company own that competitors cannot easily replicate?</em></blockquote><h3>Hardware: The GPU Squeeze</h3><p>And then there’s the hardware aspect. As Matt Biringer explains in <em>AI’s Cost Crisis</em>:</p><p>Most people think of AI as just software, but it’s really powered by hardware. Traditional Computers run on CPUs (Central Processing Units), which are like two-lane roads. This is great for handling everyday tasks one at a time. AI, given its complexity, however, needs GPUs (Graphics Processing Units), which are like massive highways with thousands of lanes, processing information all at once.</p><p>This is what allows large AI models to answer questions or generate text in real time, but it comes at a price. Running on GPUs can cost up to 100x more than running on CPUs, which is why the economics of AI are so much tougher than traditional software.</p><p>AI costs are difficult to manage because companies face both unpredictable demand and constrained supply. Companies must commit to GPUs before they know how much they need, often wasting money. At the same time, GPU shortages force long contracts or expensive on-demand pricing. Together, these make spending hard to control. It’s like being forced to sign a year-long gym membership before you know if you’ll ever go, and if you don’t, you still pay the bill.</p><h3><strong>Compute: The Hidden Bill</strong></h3><p>This isn’t the first time tech companies have been overwhelmed by infrastructure costs. In the 2010s, SaaS startups learned the hard way with Amazon Web Services (AWS), the leading cloud hosting platform. Every new customer made their apps use more online storage and processing power, which meant higher monthly bills. Costs piled up so quickly that many companies collapsed simply because they weren’t tracking them closely.</p><p>AI is now facing the same challenge, but instead of cloud servers, the big expense is GPUs, the powerful chips that run AI models. Every time an AI system trains on data, answers a question, or interacts with a customer, it uses GPU power. And every bit of that power comes with a cost.</p><p>That’s why compute, which is the raw processing power behind AI, can’t just be treated as a background expense. It has to be tracked as carefully as revenue and users. In practice, that means knowing exactly how much it costs in compute to serve each customer (for example, the cost per answer an AI model generates). Companies that overlook this may grow quickly on the surface, but underneath, they’re losing money on every single interaction.</p><h3><strong>Infrastructure: Rent or Build?</strong></h3><p>Startups face a critical infrastructure trade-off. Many build quickly on top of APIs from providers like OpenAI or Anthropic, gaining fast access to cutting-edge models without the burden of training or hosting them. But the economics are tough: every user query carries a variable cost they don’t control. As a result, margins are often razor-thin or even negative. TechCrunch recently highlighted AI coding assistants earning $500M in annual revenue, yet still operating with negative gross margins. They are essentially bringing in huge revenue, but losing money on every interaction.</p><p>Other startups take a different route: training or hosting their own models. This gives them more control over performance, costs, and intellectual property. It can also create defensibility, since owning the model makes them less dependent on outside vendors whose pricing or policies might shift overnight. The trade-off, however, is that it requires far more capital and expertise to pull off.</p><p>Once a company decides to build and host its own models, the next question is where to run them, in the cloud or on-premises. With the cloud, companies rent GPUs from providers like AWS or Azure. It’s fast and flexible, but costs can skyrocket as usage grows. With on-premises, companies buy their own GPU servers. The upfront investment is massive, but if utilization is high, it can be cheaper over time.</p><p>It’s a strategic trade-off. Using OpenAI’s API is like renting brainpower by the hour. Training or hosting your own models is like building a brain factory. Both paths tie back to unit economics, flexibility versus cost control, speed versus defensibility.</p><p>For investors, this matters because infrastructure choices made at the earliest stages shape whether a startup can scale sustainably. The companies that balance speed with economic discipline will be the ones capable of turning technical breakthroughs into durable businesses.</p><h3>Where Unit Economics Break Down</h3><ul><li><strong>Forecasting failure:</strong> Startups often commit to expensive GPU contracts before knowing real demand, leading to wasted spend.</li><li><strong>Production volatility:</strong> User adoption can spike unpredictably, sending infrastructure costs through the roof overnight.</li><li><strong>Inference costs:</strong> Popular AI tools like coding assistants often lose money on each user because API fees outweigh revenue.</li><li><strong>Integration overhead:</strong> Connecting AI to outdated systems and workflows adds major technical complexity and hidden expenses.</li><li><strong>Hidden multipliers:</strong> Small inefficiencies, like extra training runs or unnecessary prompt tokens, can double or triple costs at scale.</li><li><strong>Talent + compliance costs:</strong> Hiring AI experts and meeting regulatory requirements adds hundreds of thousands in annual expenses.</li></ul><p>Most AI startups face a cost paradox: the more users they add, the more expensive it becomes to serve them. Each interaction consumes costly computing power, so expenses rise almost one-for-one with usage. Unlike traditional software, where scale improves margins, growth here can actually make profitability worse. It’s like a restaurant where every extra meal raises the price of ingredients. More customers don’t mean more profit; they mean more loss.</p><h3><strong>Integration</strong></h3><p>One of the most overlooked yet critical challenges for AI startups is integration.</p><blockquote>As The True Cost of Artificial Intelligence: Beyond the Hype notes, “Integrating AI into existing business systems is a complex process… over 90% of organizations report difficulties.”</blockquote><p>That statistic is striking, underscoring a fundamental reality: most companies cannot seamlessly integrate AI into their workflows.</p><p>The challenges of doing so are both technical and organizational. Many enterprises still rely on legacy systems never designed with AI in mind. Data is often siloed across departments, making it difficult to consolidate, clean, and prepare. And few organizations have the in-house expertise to handle the complexities of adoption. For startups, this translates into real friction when trying to sell into enterprises.</p><p>What makes integration especially challenging is that the costs aren’t one-time. They become permanent overhead. System integration requires APIs, middleware, and frequent upgrades. In practice, that means building connectors (APIs) so systems can talk to each other, using middleware to manage the flow of data, and continuously updating everything to keep it all working together. AI-generated data itself can be five to ten times larger than the source data, driving up storage costs. And once live, models demand ongoing monitoring, retraining, and support.</p><p>This raises a strategic question for founders: should AI startups design products that plug into legacy systems, or push for “rip-and-replace” adoption? Plug-in models have the clear advantage of faster adoption and lower onboarding friction. This is a major advantage in a market where growth speed remains a critical moat. Rip-and-replace systems may promise deeper transformation, but they face longer sales cycles and higher barriers.</p><p>For investors, the takeaway is clear: the startups most likely to scale are those that strategically think about integration and minimize friction. In markets where efficiency and speed decide the winners, ease of integration can matter as much as the technology itself.</p><h3>What Founders Must Focus On</h3><p>Here are a few areas founders should focus on to build sustainable AI businesses: success will depend not just on innovation, but on scaling with sustainability</p><ul><li><strong>Compute Discipline:</strong> Track cost per inference and cost per customer as rigorously as you would CAC or LTV. (NorthCloud) Think of it as checking the “cost per answer” your AI gives, so you know if you’re making or losing money with each user interaction.</li><li><strong>Start Small, Scale Smart:</strong> Focus on task-specific models that can deliver 80% of the value at a fraction of the cost. (NorthCloud) In other words, solve one problem really well before scaling up to bigger, more expensive models.</li><li><strong>Operational Discipline:</strong> Keep R&amp;D budgets separate from production spend, and adopt practices like FinOps and AIOps early. (NorthCloud, PureStorage) This is about running experiments without letting them bleed into live operations and wreck the budget.</li><li><strong>Data Moats:</strong> Proprietary, regulatory-grade datasets are more defensible than models, which can often be copied or open-sourced. (Walturn) Data is the true “barrier to entry.” If you own it, you own the long-term value.</li><li><strong>Workflow Stickiness:</strong> Build AI into existing processes so it becomes indispensable, rather than an assistant that can be swapped out. (TechCrunch, PureStorage) Stickiness is what turns a demo into a business customers can’t imagine living without.</li><li><strong>Customer-Centric Design:</strong> Focus on solving painful, specific problems for real users rather than building technology in search of a market. This ensures adoption and creates immediate value, not just hype.</li><li><strong>Capital Efficiency:</strong> Be deliberate about when and how you spend on infrastructure, talent, and growth. In AI, burning cash to scale before proving margins is a fast way to collapse.</li></ul><blockquote>At the <strong>business level</strong>, founders need to master the basics of unit economics: compute discipline, data moats, workflow stickiness, and capital efficiency. These are the levers that determine whether growth translates into sustainable margins.</blockquote><p>At the <strong>technical level</strong>, several new ways are emerging to make AI cheaper to run:</p><ul><li><strong>Federated learning:</strong> Instead of sending all your data to one giant system, you process some of it locally (closer to where it’s created), which saves money on storage and transfer.</li><li><strong>Sparse models:</strong> Think of this as trimming the “fat” from large AI models so they only use the parts that actually matter, cutting down training costs.</li><li><strong>AI co-ops:</strong> Companies pool together to share expensive GPU resources, the way neighbors might share tools instead of each buying their own.</li></ul><blockquote>As <em>Beyond the Hype</em> explains: <em>“The key lies in selecting the right scale of models, leveraging pre-trained or open-source options, and ensuring teams are structured to deliver incremental value — rather than betting the farm on a moonshot.”</em></blockquote><p>In other words, winning in AI doesn’t depend on building the largest models or spending the most capital. It requires pairing technical efficiency with business discipline, and treating compute, data, and integration with the same scrutiny as revenue. For both founders and investors, that discipline is what separates hype-driven growth from enduring, scalable businesses.</p><h3>What Investors Must Focus On</h3><p>Investors must also rethink how they evaluate AI companies. Revenue growth and user adoption may look impressive, but without strong unit economics and defensible infrastructure, those numbers really don’t mean much.</p><ul><li><strong>Look Beyond ARR:</strong> A startup showing $500M in annual recurring revenue with negative margins is not a sustainable business. (TechCrunch) Investors need to look deeper than top-line growth and ask: Does each customer actually make the business more profitable, or are margins worsening with scale? Startups with weak gross margins, regardless of revenue, are setting themselves up for collapse once capital markets tighten.</li><li><strong>Back Efficiency Enablers:</strong> The real value lies in companies that make AI cheaper or faster, like chipmakers, inference accelerators, or GPU co-ops. (PureStorage, NorthCloud) By investing in efficiency enablers, we’re backing the infrastructure layer that every AI application depends on, not just betting on the winners of the application race.</li><li><strong>Vertical AI &gt; General AI:</strong> Sector-specific solutions in areas like fintech, healthcare, or legal can command higher prices and offer stronger defensibility. (Walturn) General-purpose AI tools are easily commoditized and expensive to maintain, while vertical solutions solve acute pain points where customers will pay a premium. Investors should favor founders with domain expertise who can tailor AI to highly regulated or high-value industries where adoption has real barriers to entry.</li><li><strong>Integration Layers:</strong> APIs, middleware, and compliance tools that help enterprises adopt AI smoothly are critical for real adoption. (PureStorage) The reality is most enterprises don’t fail because of weak models — they fail because of messy integration into legacy systems. Startups that lower the barriers to adoption and shorten sales cycles will scale faster and create stickier, longer-term contracts.</li><li><strong>Sustainability &amp; Regulation:</strong> Startups that prioritize energy efficiency and compliance with regulations like the AI Act will have a competitive edge. (PureStorage, Walturn) Enterprises won’t risk adopting tools that fail regulatory or ESG standards, and compliance will increasingly become a gating factor for procurement. Companies that build sustainability and governance into their core offering — not as an afterthought — will win enterprise trust.</li><li><strong>Business Model Resilience:</strong> Companies dependent on expensive external APIs (like OpenAI or Anthropic) face margin compression unless they build partial independence. Investors need to ask how much of a startup’s cost structure is under its own control, and whether it has a path to hybrid models, proprietary fine-tuning, or alternative providers. The most resilient businesses won’t be at the mercy of one supplier’s pricing or availability.</li><li><strong>Data as an Asset:</strong> Proprietary, structured, and regulatory-grade datasets will prove more defensible than models alone. Models can be replicated or open-sourced, but unique datasets create lasting differentiation. Investors should ask: what data advantage does this company have that competitors cannot easily replicate?</li><li><strong>Unit Economics Discipline:</strong> The strongest startups will show cost-per-inference and gross margin improvements over time, not just user or revenue growth. In other words, they’re not only growing but getting more efficient with scale. These are the companies that can compound value over the long term rather than flame out under rising compute costs.</li></ul><h3><strong>The Big Picture</strong></h3><blockquote>Matt Biringer said it best: “By understanding the true economics of AI workloads, organizations can build more sustainable paths to innovation.”</blockquote><p>The winners in this era won’t simply be those with the most powerful models; they’ll be the ones who treat compute as a core KPI, track cost per inference with the same diligence as CAC or LTV, and design leaner, task-specific systems that deliver value without burning cash. Over time, the strongest players will diversify their infrastructure strategies, moving beyond pure API dependence into hybrid or owned solutions that improve margins as scale and negotiation power grow.</p><p>For investors, the signal is clear. The companies worth backing are those with business models aligned around unit economics from the start. That means startups that:</p><ul><li><strong>Track compute economics at the customer level</strong> — Measure the true cost of serving each user (e.g., cost per inference or per workflow) to ensure growth actually translates into profit rather than hidden losses.</li><li><strong>Engineer for efficiency</strong> — Prioritize lean, task-specific models that solve defined problems, instead of costly, general-purpose systems that waste compute without adding proportional value.</li><li><strong>Build defensibility through data and workflow stickiness</strong> — Use proprietary datasets and embed solutions directly into customer workflows so the product becomes indispensable, not easily replaced.</li><li><strong>Demonstrate capital efficiency</strong> — Scale infrastructure, talent, and growth spend only after proving margins, avoiding the trap of burning cash before reaching sustainable unit economics.</li><li><strong>Design for integration</strong> — Reduce adoption friction by building tools that connect seamlessly with legacy enterprise systems, rather than forcing customers into expensive rip-and-replace overhauls.</li></ul><p>After finishing this article, I came across <em>&quot;Questioning Margins is a Boring Cliché…</em>&quot; by Sarah Wang and Martin Casado, and I wanted to share a few thoughts on how it relates to this piece. Their perspective highlights that early low margins do not necessarily signal weak or unsustainable businesses. Many of the most successful tech companies began with thin margins but improved over time through pricing evolution, infrastructure cost declines, and enterprise adoption. Similarly, AI applications have unique levers, such as model routing, usage-based billing, and proprietary data, that can strengthen gross margins as they scale. While my focus here has been on the risks of ignoring unit economics, their argument adds nuance: low margins today do not preclude strong, defensible businesses tomorrow. The real test is whether founders can convert early traction into long-term efficiency and resilience.</p><p>For early-stage investors, the best opportunities are in companies that treat technology like infrastructure, are disciplined in cost management, deliberate in product-market fit, and defensible in value creation. The winners won’t just push the frontier of what’s possible; they’ll master the economics that turn innovation into lasting enterprise value</p><h4><strong><em>Thank you so much for reading!</em></strong></h4><p>I truly enjoy connecting with others who are exploring this space, and I’d love to hear your thoughts, comments, or suggestions for future articles. If you’re building in AI, I’d be happy to connect. You can email me at <strong>elvia@lasolasvc.com</strong>.</p><h4><em>References</em></h4><ol><li><a href="https://www.north.cloud/blog/ais-cost-crisis-how-to-avoid-overpaying-for-compute-in-2025">AI’s cost crisis: How to avoid overpaying for compute in 2025</a> by Matt Biringer</li><li><a href="https://blog.purestorage.com/purely-educational/the-true-cost-of-artificial-intelligence/#:~:text=These%20include%20challenges%20in%20keeping,workloads%2C%20the%20investment%20is%20wasted">The True Cost of Artificial Intelligence: Beyond the Hype</a></li><li><a href="https://techcrunch.com/2025/08/07/the-high-costs-and-thin-margins-threatening-ai-coding-startups/">High costs and thin margins threatening AI coding startups</a> by Marina Temkin</li><li><a href="https://www.walturn.com/insights/the-cost-of-implementing-ai-in-a-business-a-comprehensive-analysis">The Cost of Implementing AI in a Business: A Comprehensive Analysis</a> by Flavia Trotolo, Hashim Hayat, Daheem Hayat</li><li><a href="https://a16z.com/questioning-margins-is-a-boring-cliche/?utm_source=substack&amp;utm_medium=email">Questioning Margins is a Boring Cliche…</a> by Sarah Wang and Martin Casado</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ae264e0076e2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Agentic AI is Revolutionizing the Supply Chain]]></title>
            <link>https://medium.com/@elvia_perez/how-agentic-ai-is-revolutionizing-the-supply-chain-f3dda4557522?source=rss-5d512ed77981------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3dda4557522</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Elvia Perez]]></dc:creator>
            <pubDate>Sun, 17 Aug 2025 03:38:17 GMT</pubDate>
            <atom:updated>2025-08-18T21:17:29.346Z</atom:updated>
            <content:encoded><![CDATA[<p>Supply chains are at an inflection point. Agentic AI is the catalyst.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*acIH9tdRMMzI5t3dsskEWg.jpeg" /></figure><h4>Introduction: The Next Inflection Point</h4><p>The demand for supply chain resilience and adaptability has accelerated post-pandemic and amid geopolitical disruptions. There is a critical yet underdeveloped opportunity: the AI-driven execution layer. Highlighted by Forum Ventures, Manhattan Associates, Codewave, and other industry leaders, this layer is rapidly emerging as the foundation for the next wave of supply chain productivity, resilience, and innovation.</p><h4>What Is the Execution Layer and Why Does It Matter?</h4><p>The execution layer represents more than just another tool or dashboard. It’s the missing piece of infrastructure that enables AI agents to bridge the gap between recommendations and real-world action.</p><p>At its core, the execution layer acts as a middle layer that enables agents to:</p><ul><li>Connect and work with outside tools like APIs, databases, web browsers, or terminals.</li><li>Keep track of progress and carry out tasks step by step without breaking.</li><li>Manage things like access permissions, errors, retry attempts, usage limits, and backup plans.</li><li>Run multi-step processes while remembering past actions and maintaining context.</li></ul><blockquote>This layer is the leap from “knowing what to do” to getting it done automatically and reliably. As Forum Ventures puts it, “This is the critical infrastructure that will turn AI from an assistant into an operator, capable of handling the messy, manual, and error-prone work that still dominates enterprise operations”</blockquote><h4>The Limits of Traditional Models and the Rise of Agentic AI</h4><p>Supply chains are vast, interconnected, and constantly tested, from labor shortages to transportation delays. Yet most traditional software still operates in silos, running on fixed schedules: monthly plans, weekly cycles, daily batches. The problem is, the real world isn’t static and legacy systems were never designed to adapt in real time.</p><p>As Keith Moore Chief Executive Officer of AutoScheduler.ai. notes:</p><blockquote><em>“</em>We’re still dealing with systems that operate on fixed cadences… But the real world doesn’t run on a schedule. It throws curveballs: labor shortages, inventory misalignments, transportation delays, demand shifts. Our legacy systems weren’t built to respond in real time.”</blockquote><p>Agentic AI introduces a fundamentally new approach. It allows organizations not just to predict disruptions, but to act on them proactively, turning uncertainty into an innovation lever. Instead of replacing core platforms, it layers intelligent, action-taking agents — powered by generative AI — across every function: demand planning, warehousing, production, and transportation. These agents serve as the “brain” above WMS, TMS, and ERP systems, making decisions, executing tasks, and communicating seamlessly with both humans and other agents in natural language.</p><h4>From Generative AI to Agentic AI</h4><p>Generative AI tools like ChatGPT have become ubiquitous. They are powerful yet ultimately stateless, limited to producing responses without the capacity to act. Sanjeev Siotia, CTO of Manhattan Associates, explains, Agentic AI moves beyond that constraint with a “think, see, do” model:</p><ul><li><strong>Think:</strong> Uses large language models (LLMs) to interpret requests and design strategies</li><li><strong>See:</strong> Continuously monitors and analyzes its environment</li><li><strong>Do:</strong> Executes actions, makes context-aware decisions, and adapts in real time</li></ul><blockquote>As Siotia notes: “Agentic AI isn’t just about responding to queries. It’s about enabling autonomous, task-oriented behavior that mirrors how humans work.”</blockquote><h4>Digital Co-Workers: How Multi-Agent Collaboration Increases Productivity</h4><p>One of the most powerful advantages of agentic AI is the rise of “digital co-workers.” These agents don’t just follow instructions; they act with initiative. They flag inconsistencies, suggest corrections, and collaborate with one another to solve complex problems. For instance:</p><ul><li>A warehouse agent can reprioritize tasks based on order urgency and available labor.</li><li>A transportation agent can reroute shipments on the fly.</li><li>A planning agent can adjust production schedules as demand shifts.</li></ul><p>Together, these agents create a dynamic, adaptive supply chain — where every node grows smarter, every function moves faster, and the entire network becomes more resilient and cost-effective.</p><h4>Agentic AI: Application Areas and Technical Requirements</h4><p>Where Agentic AI Adds Value</p><p>Codewave highlights several supply chain domains where agentic AI is already driving impact:</p><ul><li><strong>Autonomous Inventory Replenishment:</strong> Monitors stock levels in real time and automatically restocks to prevent shortages.</li><li><strong>Logistics &amp; Route Optimization:</strong> Adjusts delivery routes and schedules on the fly based on weather, traffic, or capacity issues to keep shipments on time.</li><li><strong>Procurement Intelligence:</strong> Handles supplier contracts, renewals, and cost savings automatically.</li><li><strong>Smart Manufacturing &amp; Maintenance:</strong> Shifts production schedules and predicts equipment issues before they happen, with little human input.</li><li><strong>Compliance &amp; Risk Monitoring:</strong> Keeps track of changing regulations, carbon emissions, and trade rules, and triggers fixes when problems arise.</li></ul><p>The Technical Stack</p><p>Building an agentic supply chain requires multiple layers of infrastructure:</p><ul><li><strong>Data Infrastructure: </strong>Agentic systems need real-time, detailed data to make good decisions.</li><li><strong>Decision Engines: </strong>This is the “brain” of each agent. Reinforcement learning helps agents get better over time, while heuristic planners solve constraint-heavy problems like routing and capacity. Together, they let agents weigh trade-offs and act toward business goals.</li><li><strong>Agent Orchestration: </strong>Supply chains require many agents — procurement, warehouse, delivery — working together.</li><li><strong>Integration Layer: </strong>Agents need to plug into the systems you already use — ERP, WMS, fleet tools. APIs and connectors make this possible, allowing agents not just to recommend actions but to execute them.</li><li><strong>Monitoring &amp; Guardrails:</strong> Autonomous systems still need oversight. Human-in-the-loop reviews, audit logs, and tracking tools ensure safety, transparency, and trust. Feedback loops also help agents learn from results and improve over time.</li></ul><blockquote>As Codewave explains, “Together, these technologies provide the foundation for turning agentic AI from a research concept into an operational advantage. When architectured correctly, they allow your supply chain to sense, decide, and adapt faster than competitors while preserving reliability and control”.</blockquote><h4>Business Impact, ROI, and Limitations</h4><p><strong>The Benefits</strong><br>Agentic AI and execution-layer systems deliver tangible results:</p><ul><li><strong>Faster delivery, fewer shortages:</strong> Products move through the supply chain more quickly, reducing out-of-stock issues.</li><li><strong>Lower costs through real-time optimization:</strong> The system continuously finds the most efficient ways to allocate labor, transport goods, and manage inventory.</li><li><strong>More reliable service and happier customers:</strong> Companies are better able to meet delivery promises and service commitments.</li><li><strong>Less manual effort, fewer mistakes:</strong> Routine tasks are automated, freeing people to focus on higher-value work while reducing errors.</li></ul><p><strong>The Challenges</strong><br>Effective adoption requires thoughtful execution:</p><ul><li><strong>Data quality:</strong> Flawed inputs lead to poor decisions.</li><li><strong>Security:</strong> Agents connecting directly to live systems must have strong safeguards.</li><li><strong>Trust and oversight:</strong> Human review and transparency are critical to prevent over-reliance on automation.</li></ul><h4>Investment Opportunities in Agentic AI Supply Chain</h4><p>The opportunity for founders lies in building specialized agents and orchestration platforms across high-friction domains such as procurement, warehousing, logistics, compliance, and manufacturing. These are the areas where legacy systems consistently struggle, operating on fixed schedules, siloed data, and rigid rules that break down under real-world volatility. Agentic AI offers a fundamentally different approach: adaptive agents that sense, decide, and act in real time.</p><h4>Conclusion: The Future of Supply Chain Is Agentic</h4><p>The shift toward agentic AI is accelerating as supply chains confront rising volatility and complexity. Entire categories of enterprise software are being reimagined as execution moves from static workflows to adaptive, agent-driven systems. For founders, this is an opportunity to build the foundational platforms of tomorrow’s supply chains. For investors, the mandate is clear: the execution layer is where the next generation of category-defining companies will emerge.</p><p><strong>Thank you for reading! <br></strong><em> I’d love to hear your thoughts, comments, or suggestions for future articles. Feel free to reach out at </em>elvia@lasolasvc.com<em>. I always enjoy connecting with others, exploring this space.</em></p><p><em>References</em></p><ol><li>Forum Ventures Studio <br><a href="https://www.forumvc.com/thought-pieces/supply-chain-ai-vc-opportunities">“A Call for Founders: The AI Execution Layer of Supply Chain.”</a></li><li>AutoScheduler.AI, Supply &amp; Demand Chain Executive<br><a href="https://www.sdcexec.com/warehousing/automation/article/22941858/autoschedulerai-why-the-future-of-supply-chain-execution-is-agentic#:~:text=Because%20the%20Old%20Model%20Is,to%20respond%20in%20real%20time.">“Why the Future of Supply Chain Execution Is Agentic.”</a></li><li>Manhattan Associates by Sanjeev Siotia, CTO<br><a href="https://www.manh.com/our-insights/resources/blog/reimagining-supply-chain-commerce-how-agentic-ai-is-shaping-the-future">“Reimagining Supply Chain Commerce: How Agentic AI Is Shaping the Future.”</a></li><li>Codewave<br><a href="https://codewave.com/insights/agentic-ai-transforming-supply-chains/">“How Agentic AI is Transforming Supply Chains.”</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3dda4557522" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>