D2C Revenue Rocket Podcast
Episode 11: Michael Bervell, TestPilot
AI-Powered ADA Compliance, Accessibility & Legal Risk in E-Commerce

Zak (00:08)
Welcome to the D2C Revenue Rocket Podcast, brought to you by Klaviyo and ECD Digital. Today we have Michael Bervell, founder and CEO of TestPilot. Michael, welcome to the podcast.

Michael (00:19)
Thanks, Zak. Great to be here, and thanks for having me.

Zak (00:21)
I'm really excited because every time we have conversations, I thoroughly enjoy them. So I'm excited to have this conversation and to be able to share it with all the listeners.

Michael (00:33)
Yeah, I feel like we end up laughing a ton. I learn a bunch about you, your life, your perspective on life, office dogs. I think last time we chatted, we debated whether we should get an office dog at our office. I didn't realize that someone has to take it home — it doesn't just live at the office. It's a dog that's at the office and also at home. So yeah, we have lots of fun things. I'm excited to see where this combo goes.

Zak (00:55)
That was pretty good. So to give a little more background — Michael and I were talking about office life, and he says, 'Yeah, we're gonna get an office dog. That's really cool.' So I asked, 'Well, whose dog is it gonna be?' He goes, 'Well, it's gonna be the office dog.' Okay, awesome — well, where's the dog gonna go at nighttime? Who's gonna take care of it on the weekends? 'Well, it's an office dog.'

Michael (01:20)
Someone will adopt it on certain days, I guess. I never grew up with pets. My parents — we had a rabbit maybe, and I'm pretty sure it ran away when I was about eleven. So we weren't great pet parents, but I do love animals. So yeah, I'm open to an office dog if someone's willing to take care of it.

Zak (01:46)
The office experiment.

Michael (01:49)
It's amorphous — like the tragedy of the commons, I guess.

Zak (01:53)
Michael, for anyone who doesn't know, can you give us the 30 to 60 second overview of TestPilot?

Michael (02:00)
Yes! TestPilot — we stop e-commerce businesses from getting hit with lawsuits. That's the big 30-second overview, and we do it across a bunch of different industries and spaces. Our first marquee product is ADA compliance. There's a bunch of websites that get hit with ambulance-chasing lawsuits for not having their websites be usable for people who are blind or deaf, and our whole business is to make websites more usable for those users. There's a whole other skew of products that a lot of your listeners probably know about — around CIPA and CCPA. Are you asking for consent before you do marketing and advertising? There's a bunch of ambulance-chasing lawsuits around that too. A lot around strikethrough pricing laws — how long can you discount your prices, and how do you track that? And subscription laws — how do you get consent from someone before they subscribe to your platform, so you know what they're subscribing to and you're following all the rules? So there are so many legal things in e-commerce, and there's really no company tackling that at a systematic level. That's our business, starting with ADA compliance.

Zak (03:02)
We've seen this come up with a handful of our clients over the years — they get hit with lawsuits, and when we look into it, it's generally a law firm out of New York. They'll file fifty or a hundred lawsuits on the same day, with the same plaintiff, against random websites. It can hit anybody, and it sucks when that happens.

Michael (03:29)
Am I allowed to screen share on this? Let's try it — okay, here we go. This is some of what we see in terms of the most common law firms and plaintiffs. Exactly what you just mentioned. Sometimes these law firms will go out and hire a plaintiff to file a hundred or two hundred of these cases or demand letters. As soon as there's a settlement, that person gets paid out $500 to $1,000 for filing and essentially being the face of the case.

Zak (03:39)
Yes.

Michael (03:59)
And then on the other side, the law firm pockets the rest of the money — sometimes nine, ten, eleven, twelve thousand dollars. From what we've seen, a lot of these lawsuits are growing over time. We saw five thousand lawsuits last year, and we're expecting more this year. Let's assume they all settle for ten or fifteen grand — that's a fifty million dollar industry in public settlements, and they send out even more letters. If those five thousand settlements represent, say, a twenty percent settlement or filing rate, that's probably somewhere around twenty thousand demand letters. So every business in e-commerce has seen this, I'm sure. We hear it from people all the time, whether it's their current business or a past one. And it's kind of sad, because the purpose of accessibility is to grow your market and make your website usable for all people — and yet it's been co-opted by these ambulance chasers. It ends up creating more stress than it does the joy of building a great website.

Zak (05:00)
Yeah, I want to dig in a little more, but before we keep going, can you tell us about your background? Because you have an interesting one.

Michael (05:10)
Yeah, I grew up with no dogs. I grew up in Seattle — I was the youngest of three, and both my siblings were older than me. They're both doctors now. My parents were immigrants; they came from Ghana with essentially nothing. And I loved business from an early age. Some of my early businesses were back in fifth grade — my mom would make me my lunch for school, and it was jollof rice with plantains and chicken. I'd ration it out to my friends and sell it for a dollar — I'd make about eight dollars off my lunch so I could go buy chicken nuggets, because I ate ethnic food at home all the time and never got mac and cheese. So that was my first business. I did a lot of other businesses throughout middle school and high school, and started a nonprofit about twenty years ago that's still operating today, with people in a bunch of different countries. Then I ended up applying to college — I was big into music and jazz, so I wrote my college essay about being a jazz drummer and what that meant for leadership and philanthropy, and I got into Harvard. I studied philosophy and computer science there, which is a whole funny story — the only reason I studied philosophy is because all my other grades freshman year were terrible. Imagine coming from being a 3.9 student, and by the end of my first year I had like a 2.6 or 2.7 GPA. The only class I didn't get below an A-minus in was philosophy, so I said, 'This is my new major.' I got a minor in computer science because I liked building things. After that I spent some time working in venture capital, then at Microsoft, then at Twelve, and eventually went back to Harvard for business school and started TestPilot during the summer between my first and second year.

Zak (07:14)
Very cool. Did you start it by yourself, or with somebody else from school?

Michael (07:18)
I actually met my co-founder through my ex. He was one of her best friends from undergrad — they both went to Princeton together. She said, 'This guy's great, he's super smart.' I met him about six years ago now. The relationship between me and my co-founder has lasted longer than the one between me and my ex, which is always funny to think about. But yeah, I started the company with him, and he'd seen the problem firsthand because he was working at Amazon, at Twitch, when they got sued for ADA compliance, and as a new grad at the time, his job was to fix the lawsuit. They basically dumped this PDF report on him and said, 'Fix this.' There was no tooling, no automation — it was 2022, AI had just been coming out. We thought, maybe there's a market here. So we sent out a bunch of cold emails to chief accessibility officers at places like Delta, Walmart, United, NBC Universal — sent out about a hundred cold emails and got fifteen calls. For people who do cold calling and cold email, a good hit rate is like one call per thousand. So the fact that we got fifteen out of a hundred off my Gmail, we thought, okay, there's something here. And that was even before we discovered e-commerce — at first we were only selling to the Fortune 1000.

Zak (08:35)
That's amazing. So you did that and thought, okay, we've got something here.

Michael (08:40)
Yeah, you follow the signal — where there's smoke, there's fire. We sent out those emails, found the smoke, and thought, okay, there's gotta be something here. So we've just been building and building. We raised two years ago; the business is about three years old now, and as of this morning we just passed 127 paying customers. So we went from zero to a hundred-plus in about two years, which has been an insane growth ride. Nothing like Anthropic and Claude, or Cursor, or whatever — but we're happy with where we're going.

Zak (09:18)
Yeah, it's still early.

Michael (09:20)
Yeah, still early.

Zak (09:22)
You're leveraging AI enormously in what you're doing. Can you talk to us about that?

Michael (09:30)
Yeah, I was playing around last night with a lot of our AI tools — I still code randomly on my local computer, then send it to my engineering team and say, 'Look what I made in three hours, can you guys make this better?' One thing we think about a lot is what's called agent loops. Normally when you use an AI agent, you prompt once — you say, 'Hey Claude, look at my emails and figure out any dropped leads,' or 'Hey ChatGPT, build this image.' What we've seen growing at companies like Anthropic is the agent loop, where instead of prompting once and getting a response, you prompt once and it keeps going — prompting and prompting until you hit a certain result. That's relevant for us because our result is really clear: we want to scan a website and have it show zero issues. So we build our agents to scan a website, find issues, and loop over and over until we get to a website with no issues. Because we now have two years of training data on what 'good' looks like, our agents are really precise. Our whole business is AI-focused — we take website code, add it directly to the codebase, fix it using AI, then push the changes back to the team for review. We call it AI-plus-human-in-the-loop to make all these code changes at scale.

Zak (11:05)
Awesome. I want to come back to that, but first I want to touch on something — there have been companies around for a while that do ADA compliance remediation on websites. But all the ones I've interacted with work like this: they come in, fix all the issues on your website, then hand it off and say, 'Here you go, have a wonderful day, Michael.' Why isn't that sufficient?

Michael (11:37)
Yeah, I honestly wish what you described did exist, because before us, it didn't. What actually existed was: they'd give you a list of the problems. They'd audit your site and say, 'Based on this audit, go fix it.' If they were doing what you just described, I don't think we'd have a business — because what you described is what we do: we give you a list of fixes you can review and merge live. When we think about the accessibility industry, there are two types of products for e-commerce. One is overlay widgets — tools like AccessiBe, UserWay, part of our own AudioEye offering. You install them with one line of JavaScript, and they have toggleable features — you've probably seen that little icon floating in the corner of a website. You click it and can adjust color contrast, text size, whatever. The problem is those tools only solve about 40 to 60 percent of accessibility issues. Maybe your heading structure is off — a widget can't fix that. Maybe your images don't have alt text — a widget won't fix that. Maybe your links say 'click here' instead of 'click here to learn more about products' — widgets won't solve that either, because it's purely visual. The better solution, which also tends to be super expensive, is to hire someone to manually test your site — you hire a person from Level Access, Allyant, Deque, or 216digital, and they'll manually audit your site and give you a report to fix. That's exactly my co-founder's experience at Twitch — Level Access did an audit, he got the report, and thought, 'What do I do with this?' There's no training for it — no one in school learns accessibility when they're studying engineering. They learn Python and SQL, and none of that has to do with ADA compliance, which is kind of a forgotten topic. So we tried to build a product somewhere in the middle — the ease of the overlays, but the accuracy of manual testing, and build that into a product that runs continuously — weekly or monthly — with fixes delivered by our team.

Zak (13:57)
Why do you need to do this monthly? Explain what's changing that makes that necessary.

Michael (14:04)
The best websites we see are A/B testing all the time — half the time you don't even know what's going to work in e-commerce, so you just keep trying things. Let's say you build your website to be perfectly accessible today; six months of A/B testing and tweaking colors and text sizes later, it's naturally going to break. I describe it like cleaning your room — I wish I could clean it once a year and be fine, but that would get pretty grimy. So you clean periodically, once a week, or for me, once every other week, and that keeps you in a good position. The other benefit we've seen is improved GEO and AI search visibility, and increased SEO. By making your site accessible, sure, you avoid lawsuits — that's table stakes. But beyond that, if you're changing the code so your website is more readable by a screen reader — what a blind person uses to navigate a website — it's also going to be more readable by an AI agent, Google search, or any other indexer. So it becomes a no-brainer from a revenue standpoint, in addition to reducing the cost of a lawsuit.

Zak (15:16)
Have you guys been able to start tracking the impact on GEO yet?

Michael (15:22)
Yeah, we did a small study on our first hundred customers — we asked them for their Shopify analytics and looked at the three months before TestPilot and the three months after, to see how much their search traffic increased. We saw about a 14.5 percent increase across the board. So say you're getting a thousand ChatGPT views a month — with TestPilot, you'd get about 1,140. Then we look at those additional 140 people: what's your conversion rate, maybe 10 percent? Now you have 14 new people buying, at a $100 cart size — those 14 people brought you $1,400 in net new dollars. And you avoided lawsuits, and you're ranking higher in SEO. So we can actually tie a dollar value to what it means to be accessible, and that's why most of our customers see four, five, six-X returns — we price as a cost-reduction tool, but the value we provide is also on the revenue side.

Zak (16:20)
I love that — I was hoping you'd bring up the SEO and GEO impact, because I figured it had to be really big. There's really nothing too sexy about ADA compliance on its own — you've done a lot to make it sexier — but when you tie in the revenue-driving aspect, that's when it gets really exciting.

Michael (16:46)
For really big businesses, even a small increase in search traffic is huge. We work with a couple hundred-million-dollar-plus brands, and if we can prove a two percent increase in organic search traffic, that's tens of thousands of dollars to the bottom line — their whole annual contract pays for itself within a month or two, just from that two percent increase. This is actually kind of funny — we were talking with one customer, a flip-flop company that's a conglomerate of five different sites and stores. We do check-ins every other week for the first couple months of a partnership, and at one of our last check-ins, he said, 'Hey, we're trying to figure out why our search traffic went up 22 percent compared to last year — we haven't done anything differently.' And I said, 'Well, it's accessibility.' We went back in the timeline, and it was the exact day we put in our code changes. For us it was an obvious no-brainer, but for him it was a complete surprise. So yeah, it does have a real impact.

Zak (17:57)
That's awesome. And now it sounds like you've started layering on other products in the same category as well.

Michael (18:07)
Exactly — that's the long-term strategy and vision. I go to so many conferences and see all these marketing, attribution, and subscription tools, which are great for revenue generation. But one of the biggest line items for e-commerce founders is legal expenses. You talk to founders and they say, 'Yeah, we just have a line item to pay out these lawsuits and demand letters — it's the cost of doing business.' Once you hit a certain scale, you hold a reserve just to settle these things. To me, that seems like a broken system. The challenge with this whole industry is that sometimes it costs a lot less to settle than to fight, even if you're right. Take accessibility: say your site is perfectly accessible and someone still files a demand letter. If you wanted to hire a lawyer and prove you're right all the way through court, that's going to cost fifteen, twenty, twenty-five grand. So if someone's asking for a $5,000 settlement, the economics often make sense to just settle. That's really the existential crisis of our business — the question we ask is, what can we do to make the cost of fighting so low that it falls below what people would typically settle for? That's not just an ADA question, it applies across any part of the industry where you might get hit with a lawsuit. That's where most of our products are focused — building solutions that make it cheap to prove you're right, so that if you need to fight, it's still economical.

Zak (19:44)
Now you have the data too — it's not just something that makes it cheap, it's actually a revenue-driving product as well.

Michael (19:53)
Exactly. When we look at some of these other segments, we're trying to figure out how it becomes both a cost-cutting story and a revenue-generation story. Take CCPA and privacy — a lot of people have been getting hit with California privacy, invasion-of-privacy lawsuits. There's a really prolific filer named Vivek Shah — if you've gotten one of these, you've probably gotten it from him. Essentially, when he files, he says, 'This person used your website, and the Meta ad pixel fired before they gave consent.' So the fix is a consent banner — anyone who visits your site has to say, 'I accept being tracked by Meta.' That data feeds back into Meta, and now you get less data, and your revenue starts to spiral. So the economic game they're playing is: would you rather your Meta ads revenue and data drop by hundreds of thousands, or settle for $20,000? That's both a tech problem and a strategy problem — we're trying to figure out the best tech solution to stay legally compliant on consent without completely ruining your ads data and insights, which is the harder question.

Zak (21:07)
Nice, I like that you're tackling that. You're an AI-native company in every sense — everything you do is with AI. A lot of our listeners are D2C founders and operators, and you go to a lot of conferences and talk to a lot of brands. So — what are some ways you're seeing D2C brands leverage AI for their own operations?

Michael (21:40)
Yeah, I'll say three things I've seen that are pretty interesting. One is on the content side. Before, if you wanted new content or to test new creative, you basically had to hire someone to create it each time. What's beautiful about AI is that if you have a framework or a design system, creative can be created so much faster — Claude plus Figma, for design, is unparalleled — and you tie it into your Meta ads. Now you can run a thousand different tests and SKUs and have AI analyze which ones are working. I'm not a marketer by trade, but it really lowers the bar for what that takes. We do a lot of advertising, especially cold email, and that's how we've optimized our cold emails to get response rates of five, ten, fifteen, twenty percent. That's one big way I've seen it used. The second big way is in the day-to-day operations of the business — there are so many metrics in a Shopify dashboard that you can get lost in it, but if you just download the data and upload it into Claude and ask, 'What am I missing here? What are the opportunities?' — that's huge. Beyond marketing and revenue growth, for your day-to-day operations you no longer need to hire a data analyst; you can do it all internally. And third, we're AI-native in that we record every call we take and feed it into our AI — we have a database of every recording. As the business grows and changes, we can query where we were at any point. It's this infinite record and infinite memory to learn from. It's great for onboarding — when I hire someone new, I just say, 'Ask Claude whatever questions you want, it has all the context on the company.' You should probably ask Claude first and me second, honestly, because it knows more than I do at this point. We use it a lot for internal training, onboarding, getting customers up to speed, and writing case studies. But what matters most is whether your data source is really strong. The best companies have really strong data sources — ours comes from recording meetings natively, but other customers get theirs from whatever first-party data source they have, whether it's Shopify directly or something else.

Zak (24:08)
So is all that call data stored inside a Claude project, or do you have it in a vector database? Talk about the architecture a bit.

Michael (24:18)
We have a call recording tool we use — some people use Fireflies, there's another one called Grain, and even Zoom recordings go into a folder somewhere. We have that tied to Claude through an MCP server, so we can say, 'Hey Claude, look at every recording of a call from this customer and summarize what I need to know going into my next meeting.' We even do this for in-person meetings — we take super detailed notes when we meet someone in person, then upload those notes into HubSpot and into our call recording database, and it's all directly tied into that MCP server.

Zak (24:58)
What is a piece of software that you're currently using, or one you used to use, that you want to replicate?

Michael (25:07)
I think a great piece of software — this is great for D2C listeners too — is a company called Clay. Think of it like Microsoft Excel, but on steroids. In Excel, you have columns left to right, and each row is its own piece of data — maybe a person's name, how much they've spent in the last 12 months, their location, their phone number. What Clay does is let you add new columns where each column is its own AI prompt. So you can say, 'Hey Clay, look at these rows — the person's name, spending habits, how many products they've bought — and tell me what you think they'll buy next and what offer I should give them.' It outputs a result, and you can turn that into a campaign or whatever you want. So Clay is basically Excel where every column is an AI prompt. We use it constantly for cold email, for LinkedIn outbound, for enriching leads and lead scoring. But the way I'd imagine your audience using it is: look at your customer list and figure out what unique, personalized offer you can send to each person. You can get really detailed — a little scary, honestly, but cool — like, 'I only want to email people who bought my product last month, live in Massachusetts, spent over a hundred dollars, and whose last purchase before that was six months earlier.' You build that into a table and have a specific campaign for that exact person. Takes about half an hour.

Zak (27:06)
[Sponsor read] Want to see where your e-commerce brand is leaving money on the table? Book a free revenue audit with the team at ECD Digital Strategy and get a clear plan to increase revenue. Go to Klaviyo.com/ECD to schedule yours today.

Zak (27:22)
That's pretty neat — and that all happens inside Clay itself?

Michael (27:26)
Yeah, and you can export all those results into wherever you need — another data source.

Zak (27:31)
That's neat. Klaviyo's new product they're releasing is called Composer, and essentially it's going to handle a lot of the execution that brands and agencies typically do — and it'll have its own MCP server directly into Composer. So something you could do is take the analysis you were just describing from Clay, connect it to Composer, and have Composer build out the campaigns or flows and add people to them based on the data piping in from Shopify — it could all just be one loop running.

Michael (28:06)
Exactly. One of the challenges with tools like Composer is that the data may not be fully enriched — you might have one customer where you have every data point you need: name, email, phone, buying history, address. And then other customers where you're missing half of that. When you're missing half of it, what do you do? That's where enrichment companies like Clay fill the gap. I feel like a Clay salesperson all of a sudden — they're good friends, I like them a lot, and we were an early customer. They're based in New York, same as me, so I actually went to their Series A or Series B party. Great team, really cool.

Zak (28:50)
That's awesome. Let me back up — I talk to a lot of brands who are at very different places with their AI adoption. Some are using it like a search engine, and some are doing really complex stuff with it. For brands that aren't doing much with it yet, it can be pretty overwhelming. What suggestions would you give brands on where to get started?

Michael (29:24)
What I sometimes do, about once a month since these models change so quickly, is ask my AI, 'Based on everything you know about me and the market, what new project should I build?' Sometimes we try too hard to be creative — if we don't have the perfect idea, we feel like we can't even use AI. But AI is broad enough that you can ask it how it would use AI. This ties back to the agent loop idea — level-one prompting is asking it to do a specific thing; level-two prompting, which is where I think the world is headed, is asking for an output or a goal and letting the AI figure out its own way there. So I'll ask AI, 'I want to improve my conversion rate by half a percent — give me ten experiments to do that.' Then I'll spin up ten different Claude prompts and say, 'Here's what one of my Claude agents told me to do, go build it.' You can open ten tabs and do that — it takes almost no time. Within half an hour you have ten examples you can evaluate and actually make live, all aimed at improving your conversion rate by half a percent. That's how we use Claude internally, and that's what I'd recommend to someone completely new — use Claude, ChatGPT, whatever tool you've got. Sometimes being 'stupid' about it is the smartest approach — just ask it, 'What would you do if you were me?' It'll give you all these ideas, and you just go do them.

Zak (30:54)
Sounds good — that's a great example. What you just walked through was three, four, five prompts, tops, and you're on your way to increasing your conversion rate by half a percent, one percent, two percent.

Michael (30:58)
Yeah, exactly.

Michael (31:15)
Exactly. Some of it works, some of it doesn't — the difficulty is someone still has to implement it. At a company like Anthropic, the AI implements everything, but we're not that AI-built yet. We still have our team review everything and manually spot-check. But yeah, you can go really deep in this category.

Zak (31:40)
Let's do some future-looking questions — I want to get your vision on a couple of things. Where do you see the future of e-commerce agencies in 12 months, 24 months from now?

Michael (32:00)
I think if you're still billing based on hours, you're probably not going to exist in twenty-four months. The reason is, if you tell a client, 'This took me three hours,' and the client says, 'I just used Claude Code and did it in thirty minutes,' you're going to be stuck in an infinite battle trying to prove value. The value to consumers is whatever solution you're providing — they want something completed. Maybe they're hiring you because they want a website built with a two percent conversion rate — that's the outcome. They don't really care how you get there, as long as you deliver that outcome. So I think agencies are going to move toward outcome-based pricing instead of input-based pricing, where the input has traditionally just been work hours. I don't think that's sustainable. The challenge with outcome-based pricing is that project scope starts to feel infinite — if someone says, 'I want a website with a two percent conversion rate,' and you go build it and end up spending $10,000 in Claude credits, they might say, 'No, I wanted this to be a $2,000 project.' That's why billing by hours makes sense — you say, 'This will take ten hours, anything beyond that costs extra.' But if you're billing by outcome, what's the cost of getting there? Then it's on the agency to figure out how to actually deliver that outcome within a reasonable cost. The benefit is that as you get really good, your margins increase exponentially — if you deliver an outcome the client thinks will cost $10,000, but you're so efficient it only costs you $500, you make that $9,500 margin you never could have made as a traditional agency. But that's the challenge and the arbitrage of any services business in the AI era, especially as token costs keep changing — the more you use it and get hooked on it, the more you're at risk of price increases.

Zak (34:12)
Right now we're in a time where tokens are really inexpensive — it's smart, they've got to make it cheap, get everyone hooked, and then the pricing has to go up eventually. When that happens, there will be more changes. It may come down to: if people don't have the expertise to do something efficiently with AI, it might actually be cheaper to hire someone to do it manually — it'll take more time, but the cost will end up similar to an employee. So to be really efficient, you need to really know what you're doing with the tools.

Michael (34:53)
Exactly. That's actually funny, because now, when I'm hiring, that's one of my first questions: what AI tools are you using, can you show me? I hired an EA, and my EA uses AI like nobody's business. We had a meeting on Monday, and I said, 'Hey Janet, I need you to go back through all our Grain recordings — we've recorded a bunch of case studies but never published them, and I want you to write all of them up.' She went back and found twenty. It's Monday at 1 p.m., and I said, 'Can you get this done by tomorrow at 9 a.m.? I need all twenty case studies written.' Normally that would be a week-long project. She had it back by the end of the day Monday — all twenty case studies written out. I also had AI go back and check the accuracy, and we have an AI agent that sends it out to customers to confirm they approve it and are good to launch. So within twenty-four hours, we went from having all this unused recording material to having twenty case studies scheduled to go out over the next twelve weeks.

Zak (39:59)
I love it. You'd mentioned you have siblings who are doctors — are they asking you about AI and how you're using it? Do you see—

Michael (40:16)
Yeah, brother and sister. My brother is actually a social media influencer — over a million followers across his platforms — and he's pretty anti-AI. I'm always saying, 'Why don't you use AI to make your videos? You have so much content, just upload the recordings and let AI make the videos.' And he's like, 'I don't believe in that — I think the reason I go viral is because I make everything manually, and I am human.' And because both my siblings are doctors, and honestly they're very close to people at the end of their lives, I think when you're at that point, you're not really thinking about technology — you're thinking about what's human, what you can touch and be around. So it's funny — I'm so AI-pilled, I talk about it all the time, I use Claude, I use ChatGPT. My sister messaged me yesterday, 'Michael, can you help me redo my website?' And I redid it with Claude in ten minutes, and she was so impressed, she'd never seen anything like it. I think because we're thinking about AI all the time, it's easy to assume the whole world is on this wave — but I'd guess it's probably only five or ten percent of people using these tools daily, at anywhere near full capacity. Anthropic has about four thousand employees; Microsoft has a hundred and fifty thousand. We're so early in what's possible, even for people who've already unlocked a lot of it. When I look at my siblings, I realize — yeah, these are such early days.

Zak (41:48)
I get surprised sometimes talking to companies and asking what they're doing with AI, and it's almost nothing. A lot of times it's these bigger companies you'd think would have a whole AI department, and they're not doing anything.

Michael (42:07)
It's a weird technology — I call it emergent. There's no book on how to use AI, and even the people building it don't fully know how to use it yet. Like the Fable and Mythos models that got released a couple weeks ago — they got rolled back because Anthropic didn't fully anticipate what they could be used for, and it turned out they had some pretty serious cyberattack-related risks, so they rolled them back. Because it's an emergent technology, it's hard for an established corporation to hand down a top-down mandate on how to use it. So I think the best companies using AI are very bottoms-up — figure out how to automate your own workflow, then share your best prompts with the company, and the company builds that into a system that runs in the background for everyone to benefit from. That raises the baseline of what 'bad' and 'good' look like. For us, it's almost impossible to send out a bad cold email anymore, because AI reads all of our emails — even the worst person on our team is probably going to send a cold email that's ninety-five percent good.

Zak (43:11)
This has been an awesome conversation — talking about what TestPilot's doing, talking with you, and talking about how you're using AI. It's always exciting for me when we get to chat.

Michael (43:25)
Yeah, but I don't think AI can take care of office dogs — that's the only problem.

Zak (43:30)
Not yet, not yet. I've got a couple rapid-fire questions before we wrap up — you ready? One mistake founders make early?

Michael (43:43)
Not focusing on distribution.

Zak (43:46)
Hardest part of building a startup?

Michael (43:48)
You never know if what you're doing is right.

Zak (43:50)
One thing you'd do differently?

Michael (43:53)
Honestly, things are going pretty well. But I'd say focusing more on marketing and brand building.

Zak (44:04)
Okay — and what's one thing founders should be paying attention to that they're not paying nearly enough attention to yet?

Michael (44:12)
I think every founder should find a friend who's completely AI-pilled and talk to them once a week — someone who works at Anthropic or OpenAI. Just DM someone and say, 'Can I pay you $500 a week just to ask how you use these tools?' It's going to be way more valuable than almost anything else you could do. I say that because I had dinner with a friend who got hired at Anthropic three weeks ago — we had dinner two days ago, and I just asked, 'How do you guys use Claude internally?' He showed me, and it was insane — I realized I'm basically level one.

Zak (44:48)
Like, the systems they have in place.

Michael (44:51)
Yeah, and it's all internal — unless you see it firsthand, no one's posting about it on Twitter. By the time it becomes public, you're months delayed. So if you're trying to be on the cutting edge, find a friend who's willing to be an AI mentor at one of these companies, and ask them to show you what they do. You'll be pretty amazed.

Zak (45:15)
I love it. Awesome, Michael — thank you for coming on the show. It's been great having you here.

Michael (45:21)
Thanks so much, Zak. Good to see you.

Zak (45:24)
[Sponsor read] Before we wrap up — if you're serious about growing your e-commerce brand, go grab a free revenue audit. We find the revenue you're leaving on the table. We'll analyze your store, marketing, and retention channels to show you exactly where the biggest revenue opportunities are. It's completely free, and you'll walk away with a clear growth plan. Just go to Klaviyo.com/ECD and book your audit today.