---

About Vivek : LinkedIn

Check out website: Surpassmd

FULL TRANSCRIPT BELOW

Shamanth: Welcome to the Mobile UA Show. Today’s episode is part of our series Intelligent Artifice, a deep dive on how AI is transforming performance marketing. Since 2018, over 250-plus episodes, we’ve been deconstructing how the best performance marketers actually win. Every week I sit down with the operator shaping it, or we tear apart the ad systems behind the highest-performing advertisers in the world.

My guest today is Vivek Girotra, co-founder of SurpassMD, a metabolic telehealth company built around GLP-1 medications, with HRT, TRT, longevity treatments, and peptides on the roadmap.
He spent over a decade leading marketing teams on the agency and client side before making the jump to founder.

I thought this would make for a great episode because the analog world of healthcare seems as far from all things AI as you can imagine. Vivek’s argument is that the brand, the trust, and the full user experience are the only real differentiators, and AI is what lets him build all of these faster than a team ten times his size.

In today’s episode, Vivek talks about how he asked AI to interview him before he wrote a single line of his brand messaging, why his first batch of creatives took fifteen rounds of revision, why a nurse’s unscripted hospital video beat anything his team generated with AI, and the standing instruction he gives every agent he runs. This is a great episode that gets into the nitty-gritty of building an AI-first, AI-native company in a space you wouldn’t typically associate with AI. Let’s get into it.

I’m excited to welcome Vivek Girotra to Intelligent Artifice. Vivek, welcome to the show.

Vivek: Thanks, Shamanth. I’m excited to be here.

Shamanth: I’m excited to have you because you’ve spent a number of years leading and driving dramatic results in marketing teams, and now you’re the co-founder of an AI-enabled healthcare startup.

There are so many aspects of that jump that I find fascinating, some of which we talked about in preparation for this call, and I’d love to go deeper. But before we get into AI, tell us a bit about your company.

What are you building, and why did you choose this specific problem?

Vivek: SurpassMD is a metabolic health company built to help people achieve lasting outcomes for their health.

We’re starting off with GLP-1 medications, and as we get traction there, we’ll move into other areas like HRT, TRT, longevity medications, and hopefully, as the regulatory pathway clears up, peptides as well.

The way it happened is really a confluence of three factors. I’d been in the marketing world for a while, first on the agency side, then on the client side, and I felt I’d reached a career inflection point. I could get another VP or CMO-level job, but it felt like more of the same. Second, as someone in his 40s, I’ve started to see how health is fundamentally the foundation of your life.

If you’re lacking in health, you can’t enjoy any of the other things you have, whether that’s money, travel, family, or relationships. It’s fundamental to how we live our lives. Third is the category itself.

GLP-1 medications have been a game changer for people with obesity and diabetes, and every day we discover some new health benefit, because so many diseases are downstream of obesity and cardiovascular issues. And what’s been happening with COVID is that telehealth has become the norm, whereas previously you’d almost always go to a doctor for a visit.

Now telehealth has become the default. So the drugs created a step-function change in the market, and access has become democratized. This felt like a great time to step into this market and make our play for it. And it’s also just fun to build something from the ground up by yourself.

There’s a certain element, I think, that activates a different part of your brain. When you’re always working for somebody else versus doing something on your own, there’s a different kind of drive and motivation that comes up.

Shamanth: You told me in preparation for this call that making this AI-enabled was a big part of your plan, and it still is. And yet you’re not making yet another GPT wrapper. So talk to me about how you thought about building an AI-enabled company in a sector like healthcare, which isn’t just analog, but also super regulated and super licensed.

Vivek: I think the question isn’t where’s the easiest place to build an AI-enabled company. It’s more about where the right confluence of opportunity, potential, and asymmetric advantage is, where my skill set intersects with what AI can help with. The way I think about it: where is there a surmountable moat around the business, one that’s not so easy that anyone can jump in, but not so high that someone like me, using AI tools, can’t jump over it? There’s a balance you have to figure out.

Especially where healthcare and AI intersect, AI plays a great role in automating coordination, though obviously not in clinical judgment, at least for now. If you’ve ever spoken to anyone in healthcare, there are so many pieces to the puzzle.

Even right now, every part of the funnel, whether it’s a landing page, an intake form that has to be HIPAA-compliant, a physician network, a pharmacy, logistics and cold-chain shipping, a patient EHR, or a CRM system, all these systems have to be connected and compliant with existing laws and regulations.

AI plays an important role in making sure that happens, but you always need a human there to make sure it’s done correctly. The last thing I’d say is that if you zoom out, in pharma or compounded healthcare, telehealth is fundamentally about marketing. The compound itself is literally the same thing everyone’s selling. I can’t change the formula of a medicine. So what really matters is your brand, your positioning, the access you can bring, and the trust you can build with your audience. And from a mobile point of view, what’s the full-funnel experience you can build for your user? That’s how I think about it, and AI plays a very important role in helping design that full ecosystem.

Shamanth: When you talk about the brand, the positioning, and the end-to-end user experience being the moat, you were able to leverage not just your past experience, but also accelerate it with AI. Can you talk about how you used AI to develop the brand palette, content pillars, and everything else, while keeping in mind that everything has to pass regulatory scrutiny? Walk me through that process.

Vivek: Sure. I think at its core, what’s really helpful is asking the AI to interview you. That’s an amazing process, because sometimes you have something in the back of your mind, a vision for a brand or a project, but you can’t necessarily articulate it. When I asked the AI to interrogate me exhaustively, I was able to distill things like our brand truth.

What’s our vision? What’s our mission? For example, one thing that emerged is that there’s a big movement around Bryan Johnson and the immortality project, living forever, and so on. I love Bryan Johnson, by the way, and I think what he’s doing is exciting and will help a lot of people.

But our goal isn’t to help you live forever. I don’t think that’s worthy enough. I want to make sure people are able to live a healthy, amazing, fulfilling life in the time they have. Just trying to increase the numbers on the clock is of no consequence if you’re miserable throughout all those hundred years you’re alive.

That was a very interesting insight that emerged only because I was able to go back and forth with the AI and build a foundation for what our brand stands for, our vision and mission, and so on. Once you have that, you encode it into the AI’s context, and then you start building other layers on top of it. For example, we have a skill for platform-level policy, one for healthcare-level policy, and one for operational policy.

Platform-level is, for example, Meta’s regulations around healthcare and telehealth advertising: what you’re allowed to show or not show in your ads.

The second is the clinical piece: compounded drugs aren’t FDA-approved, so you shouldn’t be making certain claims in your ads either.

The third is operational: what’s my shipping policy? If my drugs take three business days to reach someone, I can’t promise “deliver it tomorrow,” because that would be misleading. Those things are grounded in your business and can change depending on your partners.

So I’ve been able to build the core of the brand and the layers on top of it that guide the messaging, with the help of AI. It’s an iterative process. Sometimes it gets too aggressive in policing your messaging, so you tell the AI to turn the dial. Maybe it’s at an eight out of ten, and you pull it back to a six, because ultimately you want a mix of driving user conversion while staying compliant and regulated.

Shamanth: So a lot of the regulatory work we’re talking about, at the platform level or overall, is really about codifying the changes to your ads or social posts so they’re as bulletproof as possible before they go out into the wild.

Vivek: Yeah. But the important piece is being able to tune that mechanism, because if you leave it entirely to the AI, it’ll neuter your ad and make it so ineffective that nobody clicks on it.

Shamanth: And speaking of iteration and giving feedback to the AI, you mentioned that one of the big inflection points was when you moved from providing feedback with screenshots to AI being able to access a computer directly. Tell me about that change and how it affected your workflow.

Vivek: Absolutely. I think this has been one of the biggest “aha” moments for me in the past six months.

My primary interface with the computer has shifted. Now I’m pretty much in Codex or Claude for ten hours a day, as opposed to sitting in Chrome. The AI chat window has become the primary interface for knowledge work, and Chrome and the browser feel almost like an archaic piece of human interface that the agent has to contend with to get work done.

The big moment for me was when computer use came into play, because earlier AI was more of an advisor, and now it’s become an operator.

Before, I’d have to paste screenshots and say, “Here’s what I did, here’s what the output looks like, what should the next step be?” Now it can not only use the browser, which was already possible, but it can actually use a computer, which is a completely different paradigm, because now it can modify files, upload, download, and do things for you.

That’s been amazing, a huge shift, because my role has gone from executing to supervising, and I honestly didn’t anticipate that change happening so fast.

Shamanth: It’s interesting how being able to use a computer came after being able to use the internet. A few years ago I would have said the computer was the easy part and the internet was harder, but it happened the other way around.

It sounds like you built a lot of your infrastructure yourself, your CRM workflows, your website, and so on. And as you built with code, you mentioned that getting to the first 80% is relatively easy, but crossing that finish line is much harder. Can you talk about that?

Vivek: Yeah, absolutely.

Especially in the current state of AI, it gets you to 80% of the output, but it’s the human’s job to get you from 80 to 100.

The problem is that the difference between 80 and 100 is the difference between slop and greatness. That’s the core issue because the 80% output looks polished and plausible enough, most people just ship it. But that’s why so many AI ads now look undifferentiated, like what people call slop. You can identify these ads almost instantly.

So this is where truth, taste, and judgment become important.

Speaking of the social content you mentioned, my first round of creatives took almost fifteen revisions, back and forth, because I’d say, “The text isn’t readable here,” or “This user’s face looks AI, it doesn’t look authentic,” or “The product label is misaligned.”

These small things. Sometimes it would regress, fixing one error and reintroducing a previous one that was already fixed. But once you’re able to work through these fixes, you go back and tell it, “Now build a skill based on how I evaluate social creatives,” so that next time, instead of fifteen rounds, it takes eight or ten, and it keeps reducing with every consecutive batch. That’s been key, and I think it’s the part most people miss. Everyone gets caught up in the FOMO of “my AI just one-shot these creatives while I was sleeping,” or “marketers will now be judged on their GitHub.” That’s nonsense. Those are people who are either selling you a system or are just content producers, unserious people who aren’t actual practitioners driving business outcomes.

Their job is to sell AI, not to use AI to drive business outcomes, and I find that completely unserious.

Shamanth: I’d agree. At least as of today, AI isn’t ready to take you across the finish line, and that’s been my experience too. Taking it from 80 to 100% takes a lot of time, and that’s the real difference between a hobbyist and something people take seriously.

Vivek: One additional point: even as AI gets smarter, unless you’re able to ground it in the truths of your business, your context, your workflows, you end up hitting the same wall.

Shamanth: When you say “make it smarter,” talk to me about the feedback loops you’ve built that you’ve found effective.

Vivek: Sure.

One, as I mentioned, is that whenever there’s an issue, or you give it feedback you think could be replicated in the future, for example around building social media content or editing a video, you can tell it, “Go back through this conversation, look at all the feedback I gave you, and build that into a skill focused on what I’m talking about.” The second thing is your architecture: how do you structure this? There are different ways to do it.

Right now I use Codex a lot, so all the different things are housed in different projects. Grok, I think, has a new system where each bot is a “GrokBot,” treating each agent like an employee: one is your creative strategist, one is your media buyer, and so on.

So you have to figure out your preferred architecture and ground that as your context, things like: this is the company we’re operating, these are our values, this is our brand palette, so that whenever the agents are working, they have that ground truth to refer to.

Lastly, there’s a saying: by the time there’s a playbook, there’s no alpha left. So part of it is just staying on top of trends and what’s working. If we’d spoken two months ago, I would have said Claude was crushing it, but with Opus 5 or something, it got a bit nerfed, and Fable uses up all my tokens in half a day, so I moved to Codex. Codex has been absolutely crushing it for me. Part of it is honestly just riding the wave of intelligence and using it to your advantage.

Shamanth: Understanding the models, as you described. I think it was you who wrote online, correct me if I’m wrong, about using something like 5.6 Terra or Luna within Codex, something slightly less expensive, to get dramatically more mileage. That’s a skill in itself, managing models, not just agent skills.

Vivek: Absolutely. If you understand that, you can give a standing instruction to Codex: “For all grunt work, shift the sub-agents to Luna and operate at max intelligence level.” That lets me be so much more efficient with my tokens and not blow past my budget, because I don’t want to spend $10,000 a month on tokens. That would defeat the purpose of not hiring people.

Shamanth: Certainly. You also mentioned hiring a lot of human creators. Talk to me about why, especially since you mentioned earlier that you were able to make AI creatives work with multiple iterations. What inspired you to use human creators rather than AI ones?

Vivek: Sure. At the previous company, we experimented a lot with AI-generated UGC and so on. In general, I think AI UGC works, but it takes a lot of shots on goal to get the mechanics right, the inflections right. There’s still an uncanny-valley effect with AI, something slightly off about the creator, the way they talk. I think it’s most useful when all you need is a talking head.

But if you want to tap into someone’s existing distribution and organic reach, that’s where real influencers come in. What we’ve found is that all the brands running at scale right now run about 25 to 30% of their ads via sparked posts or collabs, primarily because whitelisting your posts to these accounts lets you reach an audience the algorithm wouldn’t otherwise consider yours. That distribution and social-graph reach is a huge advantage of working with human creators.

Another point: working with human creators builds a level of authenticity that comes through naturally, and it opens up a different box of ideas. If you’re sitting in your own room saying, “I want the influencer to do A, B, and C,” you’re limited by what you already know to ask for. For example, we were talking to a creator who doesn’t have a big audience but is a registered nurse, and she had this idea: “What if I do this video after my shift? I’ll be in my scrubs, in the hospital, find a corner, and record it there.”I thought that was fantastic. I wouldn’t have thought of it myself, but a nurse sitting in what obviously looks like a hospital, talking about a healthcare brand and GLP-1 drugs, feels so much more authentic than anything AI could have cooked up in my room.

There’s still a lot of value here. There are different models emerging: content rewards, pure creator videos, and various influencer or creator models. I think you should try both, but there’s still a lot of value in working with real human beings and bringing that cross-pollination of ideas to the table.

Shamanth: That’s so interesting, it’s almost like you’re unlocking completely new ideas, the way you did with that creator. I’d say that with AI creatives, even though quality is improving (you don’t see six fingers anymore), we had an instance on our team where we made a ninety-second video with AI using very high-quality models.

The only problem was shepherding those models from a sequence of eight-second clips up to ninety seconds, which took multiple days. We joked it might have been easier to animate it by hand. I agree that even though it takes a lot of time and effort today, and maybe that’ll change with more photorealistic models, right now it’s not quite there, and it doesn’t unlock the kind of human intelligence you described. You’ve done extensive work with AI, so are there examples of things you got wrong?

Vivek: Definitely a lot of things. One is that I always thought AI would be like an assistant, and I never thought I’d give it so much access and permission as I have. I thought that boundary-shifting would take a lot more time than it has. I’m amazed at how quickly I’ve let it take control of my browser, connect to my email, Slack, and Notion, all the software we use. The only thing I don’t give it access to right now is my bank account.

Shamanth: Which also makes sense, since I don’t think it would even accept that. I’ve tried, and it doesn’t accept it.

Vivek: There are places, for example, you can do agentic trading on Robinhood now, and you can give your bots a Stripe link or a linked card, a separate debit card. So there are ways to do it. But I’m still a bit hesitant to cross that line right now.

The second thing I’d say is that just a year ago, everyone was talking about prompt engineering. I even bought an e-book on it at some point, watched all the webinars, and it’s all gone away. There are no permanent rules in this space. Right now, my method of prompt engineering is, I ramble. I put on voice-to-text and just free-flow, think out loud, and let AI figure out the rest. Part of the learning is that it’s good to dive deep into best practices, but don’t hold onto them so tightly that you’re still doing them when they’re no longer needed.

Shamanth: Right. I remember being freaked out by “dangerously skip permissions,” and now I don’t even think twice about it. I remember when I couldn’t write code, using Cursor and Windsurf, and now you write code without bothering to check what’s in there. You’re right that the key is not holding on too tightly to an older paradigm.

Vivek: Yeah, same thing with GitHub. I used to worry so much about it. There are some projects where I do need GitHub, but now I just give the API key to Codex and it handles it. I don’t really worry about it.

Shamanth: Yeah, it complains, “You pasted an API key, please rotate it,” and then just goes ahead. I’m like, “Okay, go on, do it.”

Vivek: I have a standing instruction to my AI: “I’m not going to run errands for you.” I learned this: don’t be someone who runs errands for your AI, like “paste this key” or “enter this to confirm.” I’ve confirmed it once. Just do it.

Shamanth: I’m going to steal that one. I try to tell the AI not to repeat mistakes, and this is a good one to add.

Vivek: It’s like the meme: “Make no mistakes. Just build, make me a billion dollars, make no mistakes.”

Shamanth: Exactly, here’s my bank account, go. For somebody looking to launch a business in a traditionally not-AI-friendly vertical, what non-obvious advice would you give?

Vivek: Start with the problem you understand, don’t start with AI. My friend and I had been familiar with peptides for a while. We were part of the biohacking crowd, so we had experience with these medications. I personally know people who’ve been on GLP-1 medication and had life-changing transformations. Marketing is a big piece of the puzzle in telehealth, so this was a space I was reasonably familiar with, and I felt it was a good place to bring AI acceleration capabilities to speed up the learning curve. It wasn’t “I have AI, now what should I do?” It’s the other way around.

Second, the model is rented; your surrounding system is the advantage. How you establish your operating context, your business tools, your specific workflows, your stylistic judgment, and how deeply you incorporate that into the model’s way of working, that’s going to be the differentiator.

Third, how fast can you compress execution while still having checks and balances? For example, for legal work, we have counsel who signs off on agreements. But instead of going directly to the lawyer to draft an agreement from scratch, since lawyers here cost a few thousand dollars an hour, I can go to them with a draft and ask, “Do you think this is good, or would you make changes?” You can use AI for the non-clinical layers to compress execution, while still having checks and balances to make sure you’re not deviating from the correct path.

Shamanth: It’s important to start with the problem, because it’s easy to run really fast in the wrong direction, or to use AI to do something you don’t actually like, and end up asking, “Why am I spending years of my life doing something I hate?”

Vivek: Right, to that point, I think you posted somewhere about all these vendors saying, “You can make 10,000 creatives in 10 minutes.” I’m like, cool, but what’s the point? Who’s going to test 10,000 creatives? It’s actually worse than having ten, because with ten I can test them and understand what’s working. When you’re making 10,000, there’s nothing to do with them.

Shamanth: 100%. I was actually talking to someone about the paradox of choice with AI lately, a product manager who said, “Everyone on my team comes with two hundred ideas, and I have no idea how to evaluate it all, because half of it is slop.” How do you think about the temptation to make tens of thousands of things versus making fewer, better ones?

Vivek: I think two things. One, have an operating rule within your company. I’m pretty blunt with my team: “Nobody sends me AI slop and asks me to read, digest, and figure out the insights myself. You’re taking the AI output, it’s your job to run through it, digest it, figure out what’s useful, and give me the derivative output. Your job isn’t to copy-paste.” If your job is just copying and pasting, forwarding emails, then you’re not needed, it’s as simple as that.

Second is prioritization. When you’re thinking through ideas, how do you decide priority? For me right now, since we’re bootstrapped, I have to be frugal with creative testing, so I try to go as wide as possible: what are the top customer personas I can create, so that in its creative-diversity search, the system can find the right audiences for the widely diverse personas with different pain points. It’s not about using AI to make 10,000 creatives that all say more or less the same thing. That’s not useful to anyone.

Shamanth: 100%. On our team, we ran into similar challenges, partly because, as CEO, I was pushing everyone to be AI-first, AI-forward, which naturally led to people going so big on AI that they weren’t doing human review. Something I’ve learned is that I need to find a balance even while pushing the team to adopt AI, and the way I’ve framed it lately is, if you present something built with AI, you need to stand behind the conclusions.

I pushed the team to use Claude Code, everyone’s on the terminal now, which is great. But sometimes I see an elaborate statistical analysis and think, “I’m sure this is mathematically accurate because a computer did it, but I’m also sure you can’t explain or stand behind it.” So our standing rule is, if you can’t explain or stand behind a conclusion, you’re not presenting it, and you’re not sending it to a customer. I’ve had to be very careful as a leader here, because it’s easy to point to things like the CEO of Shopify spending 25% of his time coding as inspiring, and it is. Leaders do have to set the right example. But my experience is that you can go too far, and you have to rein in the team by reining in yourself as a leader.

Vivek: I think that’s a big danger. It’s almost like a dopamine hit, seeing the machine do things you never imagined. I sometimes find myself spending hours in front of the computer going down rabbit holes, and then I have to check myself: is this P0 or P1 on my list of priorities? If it’s P3, why am I wasting time on it right now? It is very seductive.

Shamanth: It’s become very easy to do very well things you should never be doing in the first place. I’ve been guilty of that. I was actually glad when we got a slower model, because there’s no way to get instant gratification out of it, so for the first time I started asking myself if I really needed to spend two hours letting something run. The alternative was to use something less powerful, which also forces you to be more intentional about what you’re doing. Otherwise I’d have thrown every problem at it and let it run wild.

Vivek: Isn’t that a Seth Godin quote? Something like, nothing is worse than doing something well that shouldn’t have been done at all.

Shamanth: I think it’s Peter Drucker, I’d have to look that up, but I remember the sentiment even if I forget the attribution. Well, this has been incredible, and it’s not often I speak with people who are actually building in the trenches while also being this thoughtful about it. Thank you for taking the time to chat today. Before we wrap, can you tell people about yourself and how they can find out more?

Vivek: Thanks for your time today, this has been great. I call this trading notes with a practitioner’s notebook, people who are actually doing the work every day and driving business outcomes, not just making content for a living. You can find out more about us at SurpassMD.com. We have a GLP-1 medication live for anyone who thinks they might qualify. You fill out our online questionnaire, and a doctor or clinician will get back to you within a few hours. For anyone listening, we have an offer for RocketShip HQ listeners: with the code ROCKETSHIP, you can get 20% off your first subscription, of any duration, up to a twelve-month subscription if you’d like. I’m pretty active on social media, you can find me on LinkedIn.

Shamanth: Wonderful, and you do write some very thoughtful posts. Thank you for the code, we’ll include that in the email newsletter and the episode post when it goes live. Excellent, Vivek, this is a great place for us to wrap. Thank you again for your time.

Vivek: Thank you.