<?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 Paul Velonis on Medium]]></title>
        <description><![CDATA[Stories by Paul Velonis on Medium]]></description>
        <link>https://medium.com/@velona?source=rss-16573b8b8152------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Ncjz7BjRYhvxRb7xgR9hwA.jpeg</url>
            <title>Stories by Paul Velonis on Medium</title>
            <link>https://medium.com/@velona?source=rss-16573b8b8152------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 04:10:26 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@velona/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[Sleeping on Grok Bot]]></title>
            <link>https://medium.com/@velona/sleeping-on-grok-bot-4ff75de37eae?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ff75de37eae</guid>
            <category><![CDATA[grok-ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 00:28:16 GMT</pubDate>
            <atom:updated>2026-11-19T00:28:16.075Z</atom:updated>
            <content:encoded><![CDATA[<p>After my post last week about Instinct, a close colleague who also tried Instinct told me they still prefer Grok Bot.</p><p>I hesitated at first, I’m not exactly short of AI tools. I’ve built a Python-based “second brain” at home, experimented with OpenClaw, and use Perplexity and Jan.ai pretty heavily. So I didn’t really want to take on another subscription.</p><p>Less than a week in, I think I’ve been sleeping on Grok Bot. In the first couple of days it helped me migrate several of my websites from a paid Hostinger plan to Cloudflare Pages’ free plan, and redesigned them in the process ( I also love Cloudfare btw). It then restructured my wordpress site to create content pillars and re-categorised all my blogs according to content area. A job I probably never would have got to.</p><p>I still think Instinct is the right fit for my personal agent needs. But Grok Bot is making a strong case for the work side of things.</p><p>The agent landscape is really heating up.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ff75de37eae" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The AI frontier is getting harder to catch.]]></title>
            <link>https://medium.com/@velona/the-ai-frontier-is-getting-harder-to-catch-5152260e26c2?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/5152260e26c2</guid>
            <category><![CDATA[venture-capital]]></category>
            <category><![CDATA[ai-frontiers]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[ai-startups]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 22:16:23 GMT</pubDate>
            <atom:updated>2026-11-17T22:16:23.670Z</atom:updated>
            <content:encoded><![CDATA[<p>All the talk is that AI is getting cheaper, and for users and developers, it generally is.</p><p>But building the frontier models is a different story. Back in 2020, GPT-3’s training compute was estimated to cost roughly US$4–5 million. Since then, estimated training costs for leading models have climbed into the hundreds of millions, it is anticipated that next year should see billion-dollar training runs.</p><p>Whilst I was encouraged to come across some labs (e.g. <a href="https://pathway.com/"><strong>Pathway</strong></a> ) pursuing alternative architectures, including approaches beyond transformers. But my impression from the discussions was that the prevailing mindset remains, scaling works. Until another approach proves itself, keep scaling.</p><p>That creates an increasingly sharp divide. If your ambition is to build a general-purpose foundation model that competes at the frontier then you need to spend and therefore raise hundreds of millions right from your seed round.</p><p>The counterargument to this was that smaller more specialised models would emerge in various domains, which will out-compete the generalised frontier models. Well this seems to be challenged by the latest frontier releases. One discussion in San Francisco indicated that founders were reconsidering their pitches because a new model release from OpenAI had challenged the advantage they were building around.</p><p>Specialisation alone may not be the moat many hoped it would be.</p><p>In the previous Web paradigm, PayPal is the example often quoted as the company that helped produce a generation of founders who went on to build significant businesses. Could OpenAI and Anthropic become similar breeding grounds? Not simply employees leaving to build applications, but teams leaving with frontier-lab experience, raising substantial seed rounds and training models designed to lead in specific domains.</p><p>Could the next wave of specialist AI companies look less like lean startups and more like heavily funded spin-outs from the frontier labs?</p><p><a href="#FrontierAI">#FrontierAI</a> <a href="#VentureCapital">#VentureCapital</a> <a href="#AIStartups">#AIStartups</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5152260e26c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The State of AI: A Conversation with an Old Uni Mate]]></title>
            <link>https://medium.com/@velona/the-state-of-ai-a-conversation-with-an-old-uni-mate-a76b5677ab12?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/a76b5677ab12</guid>
            <category><![CDATA[frontier-tech]]></category>
            <category><![CDATA[ai-doomers]]></category>
            <category><![CDATA[the-fourth-turning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Wed, 16 Sep 2026 02:54:29 GMT</pubDate>
            <atom:updated>2026-09-16T02:54:29.014Z</atom:updated>
            <content:encoded><![CDATA[<p>I was catching up with an old uni mate this week. Like me, he is at a point in his career and life where he finds himself pondering many philosophical questions.</p><p>We covered a lot of ground but ended up on the topic most preoccupying everyone these past couple of weeks:</p><ul><li>Will AI kill us all?</li><li>Will it take all available jobs? Will new jobs be created, or will AIs take those too?</li><li>What should we tell our kids to study? Will this change lead to abundance or ruin? Even in the abundance scenario what will people do for purpose?</li><li>Are the AI labs asking for a pause because of dangers? Are they doing this altruistically, to lock in their advantage, or have they simply run out of money?</li><li>Can competing companies and geopolitical actors come together to put some guardrails in place?</li></ul><p>I’ve seen a bunch of posts on social media with people asking how is AI going to kill us exactly. Let me tell you what I’m worried about.</p><ol><li>The AI labs racing towards AGI, which many believe we have already achieved</li><li>The work Unitree is doing with their robotics, specifically how fast and agile they are getting, not to mention teaching them martial arts</li><li>Tesla’s Optimus project, particularly the work around hand dexterity when the Optimus hand surpasses the human hand then really there is no advantage we retain</li><li>Neuralink research working on a human brain chip interface</li></ol><p>Each of these poses a different risk but put them all together and it leads to a pretty dystopian future: AIs taking all thinking jobs, robots with AI chips taking all doing jobs, robots being better at craft than humans, and humans becoming subservient to frontier AI via Neuralink chips.</p><p>If I were setting guardrails here is what I would do specifically. First, halt any further AI training towards AGI, direct investment instead towards specialised intelligence not general intelligence. Second, set dexterity limits to 90 percent of the human hand. Third, put guardrails around robots and don’t train them to fight or kill sadly it’s very hard to see this domain being restricted in any way as nation states race towards drone warfare. Fourth, limit Neuralink purely to disability use cases.</p><p>We are living through a very interesting time. It brings me to a book I read some years ago called <a href="https://www.goodreads.com/en/book/show/670089.The_Fourth_Turning">The Fourth Turning </a>which describes a cycle of four generations creating four distinct phases that ultimately lead to an unravelling. The final phase is usually book-ended by a financial system collapse, a world war, a pandemic or large technological change after which everything has changed.</p><p>The cycle before ours came to an end with the two major world wars at the beginning of last century and the new cycle began with the end of WWII. We are now 80 years from that point, which is the typical span of a cycle according to the book, and we are beset by geopolitical risks, pandemic risks, financial system risks and now technological risks.</p><p>It is difficult to avoid the anxiety about a major societal change, but we’ve been resilient enough to survive all the ones previous, I’m sure we will get through the next one also.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a76b5677ab12" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Leave it at the door!]]></title>
            <link>https://medium.com/@velona/leave-it-at-the-door-29d0231f1d86?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/29d0231f1d86</guid>
            <category><![CDATA[work-from-home]]></category>
            <category><![CDATA[meeting-etiquette]]></category>
            <category><![CDATA[schools]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Thu, 13 Aug 2026 03:16:58 GMT</pubDate>
            <atom:updated>2026-08-13T03:16:58.751Z</atom:updated>
            <content:encoded><![CDATA[<p>I was on a school tour this morning, when it was mentioned that phones are “leave at the gate” at this particular school, and that classrooms were moving away from devices and back to pen and paper. This was a trend we were told, expect to hear more of this type of thing across the board from all education institutions. The explanation was that teachers felt that devices were distracting students in the classroom and inhibiting engagement.</p><p>Yet walk into many business meetings, and many people come in and plonk their laptops down. “I’m dealing with a crisis..” “I’ve got to get a sales deck out” “my team need to contact me” or any other of a plethora of reasons, while supposedly also being present and contributing to the meeting.</p><p>How long before going to meeting involves a “leave it at the door” policy for your laptop and phone? How would this even work for remote working?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29d0231f1d86" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Vibe Coders Should Rethink Dependencies — Real Velona]]></title>
            <link>https://medium.com/@velona/why-vibe-coders-should-rethink-dependencies-real-velona-349a957443f0?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/349a957443f0</guid>
            <category><![CDATA[replit]]></category>
            <category><![CDATA[llm-agent]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[coding-agents]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 00:26:23 GMT</pubDate>
            <atom:updated>2026-07-31T05:29:34.143Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nl7KC53dL-cxraPt.png" /></figure><p>Coding agents change the economics of software, and that makes dependency minimisation a lot more attractive than it used to be.</p><p>There was a time when dependencies were obviously the right move. Every package you didn’t have to write yourself was a win. It saved time, reduced boilerplate, and let you move faster. You didn’t want to hand-roll a date picker, a validation library, a modal system, or a pile of tiny utilities if someone else had already solved it well. That made sense.</p><p><strong>But vibe coding changes the equation</strong>.<br>When you have coding agents doing the work, the old tradeoff starts to shift. The thing that made dependencies attractive in the first place was the reduction in effort. It was expensive to build everything, some of the code was plan boring. Who loves spending a day re-implementing a small helper that already exists in a package you could easily use.</p><p><strong>Agents don’t care nearly as much.</strong><br>If you ask an agent to build something simple, it can usually write the code quickly enough that the “cost” of hand-rolling it is much lower than it used to be. In many cases, the token cost is modest, the dollar cost is modest, and the result is often easier to understand than a pile of third-party packages with hidden behavior buried inside them.<br>That is the shift.</p><p><strong>Dependencies used to buy speed. Now they often buy complexity.</strong><br>For a solo founder, indie hacker, or vibe coder, a dependency is not just a shortcut. It is a commitment.<br>Every package you add may bring:<br>* Transitive dependencies you didn’t ask for.<br>* Security vulnerabilities you now inherit.<br>* Breaking changes you’ll eventually have to deal with.<br>* Upgrade churn that arrives right when your product is finally working.<br>* Debugging time spent inside someone else’s abstraction instead of your own code.<br>That last one is the killer.<br>You are borrowing trouble from the future. It feels productive today and turns into maintenance debt tomorrow. This was always true, but the trade-off is much more marginal and basically not worth it anymore.</p><p><strong>The security argument is stronger than people think</strong><br>This is where the dependency discussion stops being aesthetic and starts being practical.<br>If you have fewer packages, you have a smaller attack surface. That is just true.<br>Fewer dependencies means:<br>* fewer vulnerable packages,<br>* fewer update alerts,<br>* fewer surprise regressions,<br>* fewer places for malicious code to hide.<br>And for a vibe-coded app, this matters even more, because the whole point is often speed and iteration. The more moving parts you add, the more your app starts depending on ecosystems you do not control.<br>That can be fine in mature engineering teams with strong process.<br>It is much less fine in a fast-moving prototype, startup, or solo-built product where nobody has time to babysit 47 packages and their nested children.</p><p><strong>Agents make hand-rolling viable again</strong><br>This is the key insight.<br>A human developer might avoid hand-rolling because it is too slow. An agent does not have that same bottleneck.<br>You are not saving much by installing a package if the agent can generate a clean local implementation in a few seconds. In some cases, you are actually losing efficiency because the package introduces setup friction, import complexity, version mismatch, or runtime bloat.<br>So the old default of “just install a package” is no longer automatically the best move.</p><p><strong>It is not just about tokens. It is about total cost.</strong><br>People might assume using dependencies is cheaper because they save tokens and code generation time.<br>The bigger issue is not token cost. It is regression risk. A dependency may look like a shortcut today, but it can quietly introduce behavior that is hard to test, hard to observe, and hard to fix later. In vibe-coded projects, that is where things can go sideways fast: a package update, a transitive change, or an abstraction you don’t fully control can create bugs that are subtle, difficult to reproduce, and painful to untangle even with an agent’s help.</p><p><strong>The practical rule</strong><br>Here is the rule I’d give anyone using coding agents or vibe coding:<br>Assume you should not add a dependency unless you have a good reason not to hand-roll it.<br>That forces better thinking.<br>Before adding a package, ask:<br>* Can the agent write this cleanly?<br>* Is this problem actually hard?<br>* Will this package create upgrade pain later?<br>* Does this solve a real risk, or just save five minutes today?<br>* Will this package still be worth it in six months?<br>If the answer is weak, skip the dependency.</p><p><strong>Why agents still overuse dependencies</strong><br>Coding agents do not start from a blank philosophical slate. They have learned from a huge amount of human code, and human code has long treated dependencies as the default answer to many problems. That means agents often inherit the same bias: package first, build later.<br>In other words, the agent may be fast, but it is not automatically minimalist.<br>That is why the prompt matters. It gives the agent a clear operating rule: prefer native code, prefer local implementations, and only add dependencies when the tradeoff is genuinely worth it.</p><p><strong>A prompt you can add to your repo</strong><br>If you want your Agent to follow this approach, add the following to your project documentation as an engineering rule.<br>Place it in one of these files:<br>* readme.md if you want it visible to everyone immediately.<br>* agents.md if your repo uses agent instructions.<br>* project_rules.md if you want a dedicated policy file.<br>* docs/engineering.md if you keep project standards in documentation.</p><pre>Project Dependency Policy<br><br>1. Default policy<br>- Dependencies are discouraged by default.<br>- Prefer built-in browser APIs, Node standard library, framework-native capabilities, and small local utilities first.<br>- Do not add packages for trivial tasks, syntactic sugar, or convenience alone.<br>- Keep the dependency tree small and avoid unnecessary transitive dependencies.<br><br>2. When a dependency is allowed<br>- A package may only be added if it solves a real problem that should not be hand-rolled.<br>- Prefer hand-rolled solutions for simple utilities, formatting, validation, fetch wrappers, state helpers, and basic UI behavior.<br>- Avoid packages for work that can be implemented cleanly in about 50–100 lines.<br>- Prefer one well-maintained library over multiple overlapping libraries.<br>- Do not add duplicate libraries for similar purposes.<br><br>3. Required justification before install<br>Before adding any dependency, provide:<br>- why native code or existing project code is not sufficient,<br>- the security tradeoff,<br>- the maintenance tradeoff,<br>- whether it is runtime or dev-only,<br>- the expected transitive dependency impact,<br>- a no-dependency alternative.<br><br>4. Approval rule<br>- Do not install new dependencies until they are justified and approved, unless they are already part of the project’s approved core stack.<br><br>5. Post-change review<br>After coding:<br>- list every dependency used,<br>- justify each dependency,<br>- identify redundant or removable packages,<br>- remove unused packages,<br>- flag outdated, unmaintained, or high-risk packages.<br><br>If uncertain, choose the simpler implementation with fewer dependencies</pre><p><strong>Why this works</strong><br>This policy gives the agent a default bias toward simplicity, while still leaving room for real exceptions.<br>That matters because not every dependency is bad. Some problems are genuinely better solved by a mature package. But the burden of proof should be on the package, not on the codebase.<br>That is a better default for vibe coding than the old “npm first” mindset.</p><p><strong>The bottom line</strong><br>Dependencies were a brilliant answer to a human problem.<br>Agents change the problem.<br>When the builder is no longer a human typing every line, the argument for external packages becomes much weaker for a lot of everyday tasks. In many cases, it is simpler, safer, and not meaningfully more expensive to let the agent write the code directly, keep the dependency tree small, and avoid creating a future maintenance headache.<br>That is the real shift vibe coders should understand.<br>Not “dependencies are bad.”<br>But: dependencies should now be justified, not assumed.</p><p><em>Originally published at </em><a href="https://realvelona.com/2026/07/28/why-vibe-coders-should-rethink-dependencies/"><em>https://realvelona.com</em></a><em> on July 28, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=349a957443f0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The talent to build is already inside your organisation]]></title>
            <link>https://medium.com/@velona/the-talent-to-build-is-already-inside-your-organisation-fba14df527f1?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/fba14df527f1</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[corporate-innovation]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Thu, 14 May 2026 03:52:56 GMT</pubDate>
            <atom:updated>2026-05-26T02:24:32.564Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qLdpefup4W4R0vgs.png" /></figure><p>What being a Startmate First Believer taught me about the gap between corporate innovation and the startup world.</p><p>Serving as a selector for the most recent Startmate cohort gave me a perspective I did not expect: how little actually separates the world of startups from the corporate innovation work I have spent years facilitating.</p><p>I went into the process carrying the same vague mystique about founders that most people carry; a sense that building a company from nothing requires some rare combination of traits, a certain recklessness or genius that the rest of us simply lack. What I found was almost the opposite.</p><h3>The myth of the exceptional founder</h3><p>There is a narrative around startups and the people who build them, the college dropout, the obsessive visionary, the person who bets everything on a single idea. That narrative does not make for an accurate picture of what builders and building actually looks like in practice.</p><p>What I observed inside the Startmate accelerator was a process that would feel remarkably familiar to anyone who has worked seriously in corporate innovation. The same structured problem framing; the same rapid iteration between assumption and evidence; the same relentless pressure to stay close to the customer to learn and reshape what you are building. The dynamics were not foreign to me. I had helped facilitate versions of them across a multitude of organisations with people with a variety of job titles.</p><p>The mystique around founders is largely just that, mystique. The difference between a founder and a skilled corporate innovator is often structural, not one of capability.</p><h3>The capability is already there</h3><p>The people I have worked with across corporate innovation programmes over the years are genuinely exceptional; skilled in precisely the disciplines that building a company demands. They know how to move from ambiguous problem to testable prototype. They are immensely skilled at product design and customer experience. They are practised at working across functions, managing stakeholders, and finding momentum inside systems that resist it.</p><p>Many of them are, in every meaningful sense, already doing the work of founders. The difference is that they are doing it inside a company rather than building one of their own. That distinction matters but it does not say anything meaningful about their readiness or their potential.</p><h3>So why are they still inside?</h3><p>The honest answer is that the path outward is not always visible, and the support structures that make early-stage building survivable are not always obvious from the inside of a large organisation. The leap feels larger than it is, not because the work is foreign, but because the context is unfamiliar and the safety nets are not obvious.</p><p>This is precisely where programmes like the Startmate accelerator matter. Not because they teach people to be founders from scratch; but because they provide the environment, the community, and the structured support that makes the transition from skilled practitioner to builder genuinely viable. For someone who has spent years navigating corporate innovation and has something they have quietly started building on the side, Startmate is worth serious consideration. It is one of the best environments I have seen for turning that early work into something real.</p><h3>There has never been a better time</h3><p>Beyond the individual opportunity, I find myself thinking about what it would mean at a broader scale if more of this talent found its way outward. The tools available to early-stage builders today are extraordinary, the access to capital, community, and customers has never been better, and the costs of getting started have never been lower.</p><p>We spend considerable energy inside our companies asking how to make innovation happen. I wonder how we could find better ways to identify the people who are already equipped to go and build, and giving them genuine permission and support to do so.</p><p>The talent is there. The conditions are right. The only thing that remains is unlocking it.</p><p><strong>Applications for the next Startmate cohort are open.</strong></p><p>If you work in corporate innovation, strategy, or new ventures and have started building something; even something early; the accelerator is worth a serious look. The process is more familiar than you might expect, and the support on the other side is real.</p><p><a href="https://www.startmate.com/accelerator/program">https://www.startmate.com/accelerator/program</a></p><p><em>Originally published at </em><a href="https://realvelona.com/2026/05/14/the-talent-to-build-is-already-inside-your-organisation/"><em>https://realvelona.com</em></a><em> on May 14, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fba14df527f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build First: What Heidi Health’s Culture Says About Where Software Is Heading]]></title>
            <link>https://medium.com/@velona/build-first-what-heidi-healths-culture-says-about-where-software-is-heading-eb08ffa97ab9?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/eb08ffa97ab9</guid>
            <category><![CDATA[startup-lessons]]></category>
            <category><![CDATA[healthtech-startup]]></category>
            <category><![CDATA[heidi-health]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Thu, 07 May 2026 00:16:13 GMT</pubDate>
            <atom:updated>2026-05-26T02:21:27.083Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FDjWLdrM91Dgm8pY86kaIA.jpeg" /></figure><p>Last week I attended the Blackbird Sunrise Startup Festival in Sydney, my two favorite sessions of the day were the Anthopic keynote at the beginning of the day, that I wrote about <a href="https://realvelona.com/2026/05/04/how-anthropic-structures-their-technology-teams/">earlier this week</a>, and the closing session by Dr Tom Kelly, co-founder and CEO of Heidi Health. I’d been looking forward to this one. I’m relatively new to the Australian startup ecosystem, but it hasn’t taken me long to notice that Heidi comes up constantly. In almost every conversation I have in this space, someone has a connection to it, a reference to it, an opinion on it. That kind of credibility, especially for a company still in its growth phase, says something about the gravity of the people building it. Tom is a former vascular surgeon who built an AI care partner to give clinicians back the time that administration was taking from them. That origin story alone tells you a lot about why the product works; it was built by someone who felt the problem first hand.</p><p>What immediately became obvious was that Tom is not your regular founder, I mean to start with he is both a surgeon and technologist. He has that rare quality in a founder where the conviction is obvious without being performative. I left the session understanding instinctively why the company is where it is, and why it will almost certainly be Australias next unicorn.</p><p>I’ll admit that I often struggle at conferences these days, there usually isn’t that much I can take away from the sessions, but this one was different. Tom laid out what I believe is a fairly radical approach to the way they build at Heidi.</p><p><strong>The concept he put on the table: Build First Culture.</strong></p><p>The framing was simple. Everyone builds. Not some people, not the engineers, everyone. And crucially, everyone ships to production, regardless of role.</p><p>What he’s describing is a deliberate dismantling of the intermediate steps that most organisations treat as non-negotiable. The discovery phase. The design phase. The stakeholder alignment meetings. The approval gates. In a build first culture, you don’t move through those stages sequentially; you skip to the end and build the thing. The artefact is the proposal.</p><p>It’s a genuinely provocative idea, that I’ve been ruminating on since his talk.</p><p><strong>What does it actually mean to throw away the intermediate steps?</strong></p><p>Most organisations use those intermediate steps as a form of risk management. You analyse before you build so you don’t build the wrong thing. You design before you implement so the implementation is coherent. You get approvals so that no one person is accountable if it goes sideways. There’s a logic to it, even if the process tends to calcify over time into something that slows everything down without actually reducing risk in any meaningful way.</p><p>What AI changes is the cost of building. If generating a working prototype costs a fraction of what it used to, the economics of all those intermediate steps look different. The risk of building the wrong thing and then throwing it away is no longer the same risk it once was. Which means the justification for the process that exists to prevent you from doing that is no longer quite so solid.</p><p>But this raises some real questions. How do you stop it becoming, as Tom put it, a slop factory? If everyone is shipping to production, what are the guardrails? How do you maintain coherence across a product when the constraint of careful upfront design is removed? These aren’t rhetorical questions; they’re the ones any leader thinking about adopting this model needs to sit with seriously.</p><p>Tom acknowledged this directly. The answer seems to involve building robust review systems; not as a gate before you ship, but as a check after. The sequence changes. You build, you review, you refine; rather than analyse, design, build.</p><p><strong>What changes in hiring.</strong></p><p>Interestingly, Tom noted that Heidi is still hiring across all the traditional roles. This isn’t a story about eliminating functions. What has shifted is the lens. They’re looking for people who are curious and who genuinely enjoy using the tools. That’s a softer criterion than it sounds. Curiosity about tooling in a world where the tooling is changing monthly is actually a strong signal about how someone will adapt over the next few years; probably a stronger signal than any specific skill set.</p><p><strong>The question I keep coming back to.</strong></p><p>Is build first a culture, or is it a capability? Because the instinct of most organisations when they hear something like this is to try and adopt the culture without doing the underlying work to make it safe. You can tell people to just build it without creating the review infrastructure to catch what goes wrong. You can remove the approval gates without removing the conditions that made those gates feel necessary in the first place.</p><p>The organisations that will get this right are the ones that treat it as a systems design problem, not a mindset shift. The mindset is almost the easy part. The hard part is building the scaffolding that makes it safe to move that fast.</p><p>Heidi is clearly doing something right. Whether the rest of us can borrow from that playbook without also borrowing the discipline underneath it; that’s the more interesting question.</p><p><em>Originally published at </em><a href="https://realvelona.com/2026/05/07/build-first-what-heidi-healths-culture-says-about-where-software-is-heading/"><em>https://realvelona.com</em></a><em> on May 7, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb08ffa97ab9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Anthropic Structures Their Technology Teams]]></title>
            <link>https://medium.com/@velona/how-anthropic-structures-their-technology-teams-15f6f6dd3b6f?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/15f6f6dd3b6f</guid>
            <category><![CDATA[anthropic-claude-code]]></category>
            <category><![CDATA[anthropic-ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Mon, 04 May 2026 01:34:25 GMT</pubDate>
            <atom:updated>2026-05-26T02:18:24.167Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lFTwur25hrXnDJoF.jpg" /></figure><p>I was at the Sunrise Startup Festival last Thursday, and one of the more interesting sessions was the opening keynote and panel by Anthropic, who recently opened their Sydney office. Angela Jiang, Head of Product for the Claude Platform, and Katelyn Lesse, Head of Platform Engineering, were the ones on stage.</p><p>Australia was an interesting market for Anthropic to plant their flag in, and sending their Head of Product and Head of Platform Engineering to a startup festival here says something. Their reasoning was pretty clear; we rank among the highest users of Claude on a per capita basis in the world, and they wanted to come and see first hand what the local market was actually doing with it.</p><p>While there was plenty of product vision in the talk, what I kept coming back to afterwards was the commentary about how Anthropic actually operates internally. It’s rare to get a genuine look inside one of the leading AI labs, so I thought it was worth pulling that thread a little.</p><p><strong>Team size hasn’t changed; composition has.</strong></p><p>Teams at Anthropic still sit at around seven or eight people. By their own admission, that’s not really driven by any scientific reasoning; it’s more cultural and historical, and it just seems to be the size that works. Anyone who has spent time building or running technology teams will probably nod at that.</p><p>What has changed meaningfully is what those eight people actually look like. The old model, one team lead, one product manager, and four to six individual contributors, is giving way to something different. The new shape is one Team Lead and closer to a 50/50 split between what they’re calling System Designers and Implementers. What wasn’t entirely clear from the talk was whether the System Designer role is a purely technical one, a senior engineer with architectural responsibility, or whether it absorbs what product management used to do. My instinct is that it’s probably a blend.</p><p>What was made explicit by Angela is that PMs need to be more technical now. Not engineers exactly, but technical enough to ship simple features themselves. That’s a real shift. If you’re a product manager who has been leaning on engineering capacity to execute your ideas, that dynamic is changing.</p><p><strong>The innovation cycle is being compressed deliberately.</strong></p><p>One of the more striking things Angela also spoke about was the deliberate effort to collapse the time between idea and shipped feature. The expectation used to be that a feature takes a month. The goal now is to genuinely explore how to get it done in a day.</p><p>Some of that compression comes from better planning and faster iteration. But a lot of it, and they were reasonably candid about this, comes from handing large parts of the implementation effort directly to Claude. One-shotting feature development, where you prompt Claude with enough context and it produces something close to shippable code, is clearly becoming a meaningful part of how their teams actually work.</p><p><strong>Agents are the next frontier internally, not just in the product.</strong></p><p>There was mention that Claude agents are getting significantly closer to being self-learning, improving at operating autonomously for much longer periods without needing a human in the loop. This isn’t just a product pitch; the implication was that Anthropic is exploring what that looks like for their own internal workflows too.</p><p>The framing they used was around systems that run 24/7, monitoring and responding without waiting for a human operator to wake up and make a call. As an example they used a Great Barrier Reef monitoring system, which is admittedly a more compelling illustration than a software deployment pipeline, but the underlying point applies equally well to the latter.</p><p><strong>What this means if you’re not Anthropic.</strong></p><p>The honest takeaway from all of this is that the productivity gains they’re describing are real, but they require genuine structural change to realise. You don’t get a month’s worth of feature development into a day just by giving your existing team access to Claude. You get it by redesigning how the team is composed, what each role is responsible for, and how much technical fluency you expect across the board.</p><p>The shift towards System Designers and away from a traditional PM and TL split is an expression of that. So is the expectation that product people can ship. These aren’t cosmetic changes to job titles; they’re a different theory of how software gets built.</p><p>I doubt most technology organisation are even starting to have these conversations yet, but based on what was on stage yesterday, the organisations that figure it out first are going to move at a pace that makes the rest look like they’re standing still.</p><p><strong>A couple of quick asides.</strong></p><p>One piece of advice that landed well was around SaaS pricing. The argument was that platforms need to move away from seat-based models and towards feature-based pricing tied to token usage. The thinking is that AI features are now both the most valuable and the most expensive to serve, and that needs to be reflected in how they’re priced. It’s a sensible point; and probably some advice that more than a few established SaaS companies in Australia (Atlassian/Canva) should be taking seriously.</p><p><strong>on Mythos.</strong></p><p>In the Q&amp;A part of the panel, someone in the audience asked whether Anthropic’s newest model, Mythos, is genuinely as capable as the reports suggest, or whether it’s mostly well-managed hype. It was the most interesting moment of the session, not because of the answer, but because of the reaction before the answer came. Katelyn paused in the way people do when they’re trying to find language for something they haven’t quite been able to articulate even to themselves. I’m reasonably good at reading people in those moments; and my strong impression was that this is not a incremental improvement dressed up as a leap. It genuinely seems to be a step change in a different class entirely. Worth keeping a close eye on.</p><p><em>Originally published at </em><a href="https://realvelona.com/2026/05/04/how-anthropic-structures-their-technology-teams/"><em>https://realvelona.com</em></a><em> on May 4, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=15f6f6dd3b6f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ten Years Electric, and I’m Going Back to Petrol. Here’s Why. — Real Velona]]></title>
            <link>https://medium.com/@velona/ten-years-electric-and-im-going-back-to-petrol-here-s-why-real-velona-bf28a645692c?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/bf28a645692c</guid>
            <category><![CDATA[ev-charging-station]]></category>
            <category><![CDATA[australia]]></category>
            <category><![CDATA[ev-road-trip]]></category>
            <category><![CDATA[roadtrip]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Wed, 08 Apr 2026 03:48:58 GMT</pubDate>
            <atom:updated>2026-04-08T04:11:28.731Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*b8lptBORkSWjdg2u.jpeg" /></figure><p>It started in Bairnsdale. We pulled into the Tesla Supercharger to find every bay full; frustrating, but not unusual, and we were next in line so it was manageable. Six more Teslas arrive one after another within 10 minutes of our arrival while we were still waiting, a queue forming behind us almost instantly. This immediately created the issue of tracking who was next in line, luckily the Coles carpark was somewhat empty, so each car naturally lined up in each bay one after the other. We were traveling in a convoy with friends so as soon as this situation transpired I did a quick search on Plugshare.com and sent them a text to warn them: skip Bairnsdale, head straight to the NRMA charger at Lakes Entrance where there are 3 of 4 charges available, we’ll meet you there.</p><p>Lakes Entrance turned out to be a lovely accident. We found a winery, sat down for lunch, and did what you do on a long weekend away; we relaxed, we lingered, we ordered another glass. It was exactly the kind of unplanned afternoon that makes a road trip worth taking. We stayed longer than we should have. We knew it at the time, in that vague, easy to ignore way you know things when the food and wine is good. We weren’t in a rush it would be fine. Or so we thought…</p><p>We arrived at Cann River rain streaking across the windscreen, and noticed quite a few people congregating around the Tesla chargers. We found a spot to park and walked over to discover some general confusion around the order of things. With some gentle encouragement we were able to get everyone to identify themselves and their vehicle and at this point found ourselves number thirteen in the queue with the three Tesla Superchargers already full. I immediately started doing the math in my head. 4 cars per slot, 20–30 minutes per car minimum, plus our time to charge. I passed on the bad news to the family, luckily there was a public toilet and a playground for the kids.</p><p>What followed was a couple of hours of standing in the rain, coordinating a makeshift queue system with eleven strangers, watching the weather and collectively hoping nobody had a charging fault. We were lucky. The rain stayed light. The charger held up. Everyone got through, it was even jovial at times as everyone was sharing the same miserable experience. But the whole time, the thought sitting quietly in the back of my mind was: this is not good enough.</p><p>Standing there in the rain, I kept thinking back to our early road trips in the Tesla, nearly 10 years ago now, when the network was sparse and the cars were rare and pulling into a Supercharger to find another EV already there was genuinely exciting; you’d wander over, compare notes, talk about where you’d come from and where you were headed. It felt like a club. The infrastructure hadn’t caught up yet, but neither had the demand, and somehow it balanced out. Twelve people in a queue in the rain, coordinating on their phones, is a very different kind of club.</p><p>If the rain had increased in intensity it would have become impossible to manage the queue, and perhaps the jovial nature would have changed to something else. Even worse if the charger had failed everyone would have been stranded. With no other charging locations within 150km, on a public holiday, with local accommodation at capacity, it would have either been a boon for tow truck companies or we all would have ended up sleeping in our cars.</p><blockquote><em>“After ten years and more road trips in an EV than I can count, I am moving back to petrol for the car we use to travel in. That is not a sentence I ever expected to write.”</em></blockquote><p>The return trip introduced a different kind of problem entirely. We’d watched enough charging queues by that point to decide we were not going to repeat the experience on the way home. The plan was simple: charge to one hundred percent before leaving, skip every Supercharger on the route back, and rely on the NRMA fast charger at Lakes Entrance as our one stop.</p><p>It was a reasonable plan. On paper, it was actually a good plan. What we hadn’t accounted for was what happens when a ten-year-old Tesla meets a brand new NRMA charger and they decide, at the worst possible moment, that they are not going to cooperate. The charge session started. Something went wrong. And when we tried to disconnect the cable, the charge port jammed, not stuck in the inconvenient way, but stuck in the way where you’re standing in a car park in Lakes Entrance gently trying not to make things worse while quietly catastrophising about the 340 kilometres between you and home.</p><p>After calls to NRMA, and Tesla which weren’t that much help truth be told, we eventually got it free. Unfortunately we damaged the port in the process, bent, misaligned, just functional enough to let us disengage, but not functional enough to inspire any confidence about what came next. We left Lakes Entrance with 240 kilometres of charge showing on the dash. We had 340 kilometres to cover. The maths was not complicated, and it was not good.</p><p>What followed was the kind of drive that rewires you a little. Luckily Tesla shows you the state of every supercharger from within the cars navigation system. Every Tesla Supercharger between Lakes Entrance and Pakenham was full and very busy. We didn’t stop. We couldn’t risk losing the range detouring nor sitting in the car burning energy while waiting in another queue. We drove carefully, on Teslas ‘Range Mode’ , keeping speed down, and trying to get as much aerodynamic assistance by driving behind trucks and SUVs, watching the percentage tick down with the particular focus of someone who has done the arithmetic and knows exactly where it lands. We pulled into Pakenham with four percent battery remaining. Four percent. On a trip where the original plan had been to skip chargers entirely and coast home in comfort.</p><p>This was crunch time, could we charge or would we need a tow. We managed, to carefully work the charge port back into a usable state — enough to charge at Pakenham and get home without a tow truck. Phew!! Crisis averted.</p><p>This is not how a modern transport network should function. A charging cable that physically jams and cannot be disconnected is not an edge case to be dismissed; it is a hardware failure on critical infrastructure with no fallback. In a regional town, at night, with a car that cannot move, managed is the difference between an inconvenient story and a genuinely dangerous one.</p><p>The charger count narrative talks about new installations, about government funding, about progress. What it does not talk about enough is whether those new chargers are being tested against the diversity of vehicles on Australian roads, including vehicles that are five, eight, ten years old, which is the actual fleet that will be pulling in to charge on any given long weekend for the next decade.</p><p>I want to be clear about what this piece is, and what it isn’t. It is not an anti-EV rant. I still believe in the technology. I am not switching because EVs are bad; I am switching because the infrastructure required to use one confidently on regional and rural Australian roads is not where it needs to be, and I no longer have the patience to plan trips around that gap.</p><h3>The infrastructure story we keep being told</h3><p>The headline numbers sound encouraging. Tesla reached 150 Supercharger sites across Australia as of late March 2026. The federal government has committed significant funding; the Driving the Nation Fund is partnering with NRMA to deliver a national backbone network of over 100 EV charging stations on key highway routes, spaced at an average interval of 150 kilometres. Progress is real, and I do not want to dismiss it.</p><p>But 150 kilometres between chargers, as a planning benchmark, assumes everything goes right. It assumes the charger you arrive at is working. It assumes nobody else is already using it. It assumes the weather is fine and your actual range matches your theoretical range. On a public holiday weekend, driving through the kind of regional Victoria terrain between Melbourne and the New South Wales south coast, none of those assumptions holds particularly firm.</p><h3>The problem isn’t just the number of chargers</h3><p>EV sales in Australia are currently outpacing charge point installations by more than three to one. That ratio matters enormously in places like Cann River, because you are not just competing with the number of chargers available; you are competing with an ever-growing pool of EV drivers who have all made the same calculation about where to stop on the same long weekend. A single stall in a regional town was never going to scale to meet that demand, and yet here we are.</p><p>As more drivers came to rely on public fast charging, reliability and congestion issues became harder to ignore, particularly during holiday travel periods; this has shifted conversations away from charger counts and toward uptime, redundancy, and user experience. That is exactly the right conversation to be having, but it is one the industry should have started having several years ago.</p><p>There are hundreds of millions of dollars being spent on chargers using taxpayer money, but there is quite little consideration given to the lifespan and longevity of electric chargers, and to making operators accountable for their own infrastructure. Accountability is the word. A broken charger in the CBD is an inconvenience. A broken charger in Cann River, on a public holiday, with twelve people in the queue and the next option 150 kilometres away, is something closer to a stranding risk.</p><h3>The routes that work, and the ones that don’t</h3><p>For city-based drivers who take occasional road trips on the east coast, a pure BEV is entirely manageable. For those whose travel regularly extends into areas where charging infrastructure is sparse, a PHEV reflects the current reality of Australian roads more honestly. That is a measured and honest summary from people who spend a lot of time thinking about this, and it broadly matches my own experience over the past decade.</p><p>The Melbourne to Sydney run? Fine. Adelaide? Fine. The network on the major east coast corridors has genuinely improved and, on a regular weekday, the experience is largely seamless. But get off those corridors, head toward the coast, into the hills, down to the south coast of New South Wales, and the picture changes quickly. The gaps that exist in those areas are not just logistical inconveniences; on a public holiday with a full car and children onboard, they are genuine safety considerations.</p><h3>A decade is a long time to wait</h3><p>I have been patient. I bought my first EV when the charging network was genuinely sparse and the technology was genuinely limited, and I adapted my travel habits accordingly. I planned meticulously. I left buffers. I turned inconveniences into stories. For a long time, I believed that patience was the right posture; that the infrastructure would catch up, and that early adopters had a role to play in demonstrating demand.</p><p>What I did not anticipate is standing in the rain in Cann River in 2026, doing the same mental calculations I was doing in 2016, just with a longer queue.</p><p>The Electric Vehicle Council is confident the vast majority of Australia will have access to public EV charging infrastructure before 2030. I hope they are right. But 2030 is four years away, and I am not willing to spend another four long weekends negotiating queues in the rain on the off-chance the charger holds up.</p><p>“The technology deserves better infrastructure. Australian EV drivers deserve better infrastructure. And until we have it, I think we need to be honest that the promise of the electric road trip is still, for many routes, exactly that: a promise.”</p><p>The transition is coming. The direction of travel is right. But the gap between the narrative and the lived experience of trying to drive from Melbourne to Eden on Good Friday is, right now, very wide; and for families making practical decisions about which car to buy and how to travel, that gap is what matters most.</p><p>I will be back in an EV for road trips eventually. But not yet.</p><p><em>Originally published at </em><a href="https://realvelona.com/2026/04/08/ten-years-electric-and-im-going-back-to-petrol-heres-why/"><em>https://realvelona.com</em></a><em> on April 8, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bf28a645692c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Software Engineers Are Wrong About AI Disrupting Enterprise SaaS]]></title>
            <link>https://medium.com/@velona/why-software-engineers-are-wrong-about-ai-disrupting-enterprise-saas-126ccdc830a1?source=rss-16573b8b8152------2</link>
            <guid isPermaLink="false">https://medium.com/p/126ccdc830a1</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[saas]]></category>
            <category><![CDATA[salesforce]]></category>
            <category><![CDATA[disruption]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Paul Velonis]]></dc:creator>
            <pubDate>Mon, 09 Feb 2026 22:06:20 GMT</pubDate>
            <atom:updated>2026-02-13T16:25:32.903Z</atom:updated>
            <content:encoded><![CDATA[<h4>There’s a narrative gaining momentum in engineering circles that goes something like this: AI coding tools have made developers so productive that the entire SaaS stack is about to get disrupted. Engineers are shipping features faster, and they’re extrapolating that into a future where traditional SaaS platforms like Salesforce, ServiceNow, and SAP become obsolete.</h4><p>I’ve written about this acceleration before. In my post when I was <a href="https://realvelona.com/2025/05/19/how-i-built-my-dream-gtd-app-in-hours-with-replits-ai-agent/">exploring one of the vibe coding platforms Replit</a>, I explored how AI-powered development tools are genuinely transforming what’s possible for prototyping and building certain types of applications. I built my own GTD app for under $100, something that would have cost tens of thousands just a few years ago. The engineers aren’t wrong about the productivity gains; they’re wrong about what those gains actually mean for enterprise software.</p><p>In my opinion enterprise SaaS platforms won’t be disrupted by AI. They’ll be the primary vehicle through which organisations adopt AI in the first wave of transformation. I’ve written about this in a previous post about <a href="https://realvelona.com/2025/07/25/where-are-we-with-ai-adoption-and-is-this-the-end-of-white-collar-professions/">where we are with AI adoption</a>.</p><h3>The End for Non-Enterprise SaaS</h3><p>Let me be clear: there absolutely will be disruption. The overpriced, underutilised SaaS tools I wrote about in my Replit piece are vulnerable. You know the ones; the apps that charge a fortune for features nobody uses, where the price-to-functionality ratio makes you wince every time the renewal comes up. Those are ripe for replacement with custom-built alternatives using AI development tools.</p><p>But Salesforce? ServiceNow? SAP? Those aren’t going anywhere. Not in this first phase of AI adoption, anyway.</p><h3>Why Enterprise SaaS Will Be the Vehicle for AI Adoption</h3><p>In my piece on AI adoption and white-collar professions, I outlined three paradigms of AI transformation. We’re currently in Phase 1: the Human-First AI Transformation, where humans use AI tools to enhance productivity. This is the ‘sprinkle AI on everything’ phase.</p><p>Enterprise SaaS platforms will dominate this phase for three fundamental reasons that have nothing to do with technical capability and everything to do with how large organisations actually operate.</p><h3>1. Procurement Inertia and Existing Vendor Relationships</h3><p>Large organisations have established procurement patterns. They have vendor management processes, contractual frameworks, and relationships that have been built over years or even decades. When Salesforce offers an AI upgrade, it fits perfectly into the existing buying pattern. The vendor is already approved, the legal framework is in place, and the procurement team knows exactly how to process the purchase.</p><p>Compare that to the alternative: convincing leadership to embark on a massive transformation programme to replace core systems with custom-built alternatives. Which path do you think a risk-averse enterprise will choose?</p><p>It’s far easier to buy the AI upgrade than to disrupt and re-engineer your entire operating model. This isn’t a technical decision; it’s an organisational one.</p><h3>2. Cost Arbitrage in the Short Term</h3><p>Yes, AI development tools can make building software cheaper. But you know what’s even cheaper in the short term? Clicking ‘upgrade’ on your existing SaaS subscription.</p><p>A wholesale transformation programme comes with massive costs: change management, training, integration work, data migration, the inevitable delays and overruns. Even if you’re using AI to accelerate development, you’re still looking at months or years of effort and significant investment.</p><p>Meanwhile, the SaaS provider offers AI capabilities as an add-on or upgrade. The CFO does the maths and it’s not even close. The transformation programme might have better economics in the long run, but ‘in the long run’ is doing a lot of work in that sentence. Most organisations will take the certain, immediate cost savings of the upgrade over the uncertain, distant payoff of a transformation.</p><h3>3. Risk Transfer and Accountability</h3><p>This is perhaps the most underestimated factor. When you buy AI capabilities from your SaaS provider, they’re on the hook for making sure it works safely, securely, and in compliance with regulations. They assume the liability.</p><p>When you build it yourself, even with AI accelerating the development, you own all the risk. If something goes wrong, if there’s a security breach, if the AI makes a mistake that costs the company money or reputation, that’s on you.</p><p>For regulated industries, this calculation is even more stark. You can’t just spin up an AI agent that touches customer data or financial records without proper governance, compliance, and audit trails. The SaaS vendors are already compliant, already audited, already certified. They’ve done the heavy lifting on regulatory frameworks.</p><h3>The Skills Gap Nobody Wants to Talk About</h3><p>The engineer’s perspective assumes that organisations have access to competent technical teams who can wield these AI development tools effectively. This is, to be generous, an optimistic view of the current state of enterprise IT.</p><p>Most large organisations don’t have the engineering talent to build and maintain sophisticated AI solutions. They’re already struggling to keep their existing systems running, to hire good developers, to retain the talent they have. The idea that they’re suddenly going to pivot to building bespoke AI-powered replacements for their core systems is fantasy.</p><p>The SaaS providers, on the other hand, do have that talent. They have entire teams dedicated to integrating AI capabilities into their platforms. They can amortise that investment across thousands of customers. It’s the classic economics of software at scale.</p><h3>What Actually Gets Disrupted</h3><p>So if the enterprise SaaS platforms aren’t getting disrupted, what is?</p><p>The bloated, overpriced tools that deliver marginal value. The niche SaaS products charging enterprise prices for commodity functionality. The internal tools that were built by consulting firms years ago and have been limping along ever since. Those are all vulnerable to being replaced by AI-accelerated custom development.</p><p>I’ve seen this firsthand. In my recent businesses, there were multiple instances where we paid thousands of dollars monthly for SaaS tools that delivered functionality we could now build ourselves in a weekend with Replit or similar platforms. The price-to-functionality ratio was so skewed that the economics of building our own made sense.</p><p>There’s absolutely a shake-up coming for the SaaS industry. Companies will start scrutinising their tool stack more carefully, identifying what’s truly irreplaceable and what could be rebuilt in-house. The mid-tier, feature-poor, overpriced SaaS companies should be worried.</p><p>But the enterprise platforms? They’re in a different category entirely.</p><h3>The Organisational Transformation Reality</h3><p>Having lived through the digital and agile transformations over the last 20 years, I can tell you that organisational change takes far longer than anyone predicts. Technology is easy; changing how organisations work is hard.</p><p>In the early phases of agile adoption, organisations dabbled rather than truly transforming. They stuck ‘agile’ on the front of everything but underneath, it was the same old processes. It took years, sometimes decades, before genuine transformation happened, and only then through either fear of disruption or visionary leadership.</p><p>We’re seeing the exact same pattern with AI. We’re in the dabbling phase, the ‘sprinkle AI on everything’ phase. True transformation, the kind that would involve replacing core enterprise systems, is still years away. We can’t even say we’ve reached late mainstream adoption of the Human First AI paradigm yet, because established patterns and methodologies are only just starting to emerge.</p><h3>What This Means for the Future</h3><p>The engineer’s narrative about AI disrupting the entire SaaS stack isn’t wrong; it’s just premature and incomplete. Disruption will happen, but it will be selective and slower than predicted.</p><p>In the first phase of AI adoption, enterprise SaaS platforms will strengthen their position by becoming the delivery mechanism for AI capabilities. They’ll upgrade their platforms, add AI features, and make it easy for organisations to adopt AI without the risk and complexity of transformation programmes.</p><p>Later phases might look different. As we move into the Human-Supervised AI paradigm and eventually toward more autonomous AI systems, the economics and capabilities might shift enough to make replacement viable. But that’s years away, and by then the landscape will look completely different anyway.</p><p>For now, if you’re an engineer convinced that Salesforce is about to be disrupted by AI-powered startups, I’d suggest you’re underestimating the power of organisational inertia, the complexity of enterprise needs, and the value of established vendor relationships.</p><p>And if you’re working at an enterprise SaaS company, I’d suggest you’re in a stronger position than the doom and gloom narratives suggest, as long as you’re aggressive about integrating AI capabilities into your platform. The organisations that succeed will be those that make AI adoption as frictionless as possible for their enterprise customers.</p><p>The revolution is coming, but it’s going to be an upgrade, not a replacement.</p><p><em>Originally published at </em><a href="https://realvelona.com/2026/02/10/why-software-engineers-are-wrong-about-ai-disrupting-enterprise-saas/"><em>https://realvelona.com</em></a><em> on February 9, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=126ccdc830a1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>