In this interview, Michelle Lim, Co-founder at Flint, discusses why websites should move from manual optimization toward autonomous systems that continuously test, adapt, and generate content without human initiation.
You’ve described your time at Warp as a “telephone game,” where shipping even a simple A/B test required weeks of coordination across marketing, design, and engineering. Was there a specific moment or failed launch where you realized this wasn’t a tooling problem but a structural failure of how the modern web is built, and that Flint needed to exist as a new layer entirely?
That’s a deep question, and a good one. We were optimizing for the keyword “best Mac terminal.” We were running Google Ads and trying to improve CAC, and every single experiment required coordinating between six different teams.
After six weeks of human labor we shipped a new page layout and experiment. It showed no improvement at all. Zero. That experience made it very hard to justify running another test, because six weeks of effort didn’t improve business outcomes.
I later spoke with a VP of Marketing who told me she now bans experimentation entirely, because the cost of running a test is simply too high. So fundamentally, the paradigm of the web right now is that a human has to go in and say that they want to change something. That means progress is bottlenecked by humans coming up with ideas.
The cost of testing new ideas should be close to zero. The website itself should be proposing changes and running experiments continuously. Removing humans from the loop ensures that and it’s about making changes constant rather than something you attempt once every six weeks.
You’ve said Flint’s core idea was validated within days. For founders reading this, can you walk us through what those first validation conversations actually looked like? What did you show users, and what signal made you confident this wasn’t just a productivity tool, but a platform worth building full time?
I think about validation in two phases: validating the problem and validating the solution.
The framework I used was one-week sprints. Every week, I’d talk to six to ten people: users, non-users, and decision-makers. In this case, that meant marketing leaders. On those calls, I followed the MOM Test framework to understand whether the problem was important enough to fix.
First, I validated the person. So who would be responsible for this solution? And then what are their priorities for the quarter? What is their process today? And how much is this problem costing them? And then when you validate that this person has this problem that is expensive, then the next stage is to show them prototypes of how we could solve the problem.
I’d show them competitors’ landing pages and ask for raw reactions. That was incredibly useful. I could hear very clearly what they liked, what they wanted, and what they felt was missing. I also clicked through a prototype I’d assembled from open-source repositories, just to walk them through what managing a website with AI could look like.
The most important part of the process was design partners. And when you’re working on a design partnership, you don’t need to have built the entire product yet but you need to have proven that people are willing to spend time and money with you to get their problem solved.
Many AI website tools today automate tasks, but Flint insists on the word “autonomous.” Drawing on your cofounder’s background in autonomous vehicles, how do you technically define autonomy in the context of a website? What signals does Flint’s system observe, and what decisions can it make without human initiation?
We think about autonomy the same way people think about self-driving cars. There are different levels. You’ve got something like Tesla’s cruise mode on one end, and fully autonomous systems like Waymo on the other.
We apply that same framework to websites. At the first level, you have assisted editing where a human is still in the driver’s seat, prompting a website builder to improve content or make changes.
The next level is the proactive pages concept. Here, the website starts doing more on its own. It studies competitors, looks at search trends, understands who your users are, and then suggests which pages you should be building.
The final level is full autonomy. At that point, as a website owner, you don’t need to review or approve every page anymore, you trust the system to produce great content on its own.
What’s interesting is that the signals we use at level two are the same ones we use at level three. The difference is confidence. At full autonomy, there’s enough trust in those signals that the human no longer needs to be involved for the site to consistently produce high-quality content.
Flint claims it can extract a complete brand system from a single URL, not just colors and fonts but reusable components and interaction patterns. From an engineering perspective, how do you distinguish between fragile visual artifacts and stable brand primitives, and how do you avoid the “AI uncanny valley” that plagues most generated websites?
One of the key constraints we place on the AI is that it can only operate within an existing brand identity. So if Apple.com came to Flint, the AI would be limited to Apple’s established components. Its typography, color palette, and design system.
We encode all of that as tokens in our system, and we train the AI to use only those tokens. Because of that, it’s hard for the AI to suddenly introduce things like purple gradients, rounded corners, or that “vibe-coded” look you see in a lot of apps today.
The key point is that we’re not trying to create brands from scratch. We don’t do brand creation. We work with teams that already have an existing website and an established brand, and we help them build within those boundaries.
Unlike prompt-driven tools, Flint is fundamentally data-driven. Users upload spreadsheets, competitor URLs, or persona inputs, and the system generates fully coded pages. Why did you believe data-to-web was the only viable path for enterprise use, and what breaks when AI website generation relies too heavily on prompts?
Running a business is really about taking in inputs from the world and then deciding how to steer the business based on those inputs. Giving a website those “eyes,” almost like lidar sensors on the world, helps close that loop. Instead of humans noticing changes and then manually prompting the website to react, the site can respond directly to what’s happening. That’s why data as an input is so important for websites.
The second benefit of feeding data into a website builder is that it can remove the human from the equation for certain decisions. You no longer need someone to constantly monitor the world, interpret what’s happening, and then translate that into updates on the site.
That said, not all marketing is driven purely by data. Areas like product marketing and positioning can’t always be derived from data alone. If you were living in the early 20th century, the data would just tell you to make faster horses but it takes a human to say, “No, we actually need a car,” and to know it’s time to stop hiring horses altogether.
The data-heavy side of marketing is extremely menial. It’s manual, repetitive, and there’s no real alpha in doing it. Human effort is much better spent on creativity. A core part of Flint’s philosophy is enabling marketers to focus on what makes us human, the creative parts of the job.
One of the trade-offs of relying on prompts is that a human has to be present to submit them. That makes systems less responsive when something happens after hours, or even during the workday when priorities are competing. That’s ultimately the cost of prompt-based workflows.
You’ve argued that traditional SEO is giving way to AEO, Agent Engine Optimization, as AI agents increasingly browse the web on behalf of users. When Flint generates a page today, what structural or semantic decisions are already being made with machine readers in mind rather than human visitors?
The first piece is server-side rendering. A lot of AI bots can’t reliably render JavaScript. On a normal website, content is often client-side rendered, which means you’re sending JavaScript to the user’s browser and relying on it to execute there. That JavaScript often can’t be interpreted by AI engines.
Because Flint is server-side rendered, the browser, search bots, and AI agents all receive fully rendered HTML with the content already in place. So they are able to crawl, understand, and index the site.
The second thing we focus on is structuring content to answer real questions, as opposed to filling keywords. Our customers are seeing strong results because we’ve mastered how AI systems expect pages to be structured.
For example, on comparison pages, it’s critical to include sections like who should choose which option and why. More broadly, comparison pages have become essential for companies, because AI engines rely heavily on them when generating recommendations.
You’ve said, “It’s time to kill the traditional website.” In a world where pages dynamically reconfigure based on visitor intent, industry, or even whether the visitor is human or an AI agent, does the homepage still matter? How should brands think about identity when there is no longer a single canonical page?
It’s worth investing real time and effort in your homepage, because it defines your entire design system, the look, feel, and overall expression of your brand. When you get that right, you’re also giving AI agents the context they need to personalize and adapt the site for different users.
I also think that even for transactional products customers still go straight to the homepage. They want to see how you present yourself to the world.
At a fundamental level, humans are still going to be visiting websites in the future. Because of that, having a strong brand and a clear way of expressing who you are really matters. It’s similar to humans as well, even though most information about me could exist online as words I’ve written, people still want to understand who I am as a person.
Flint’s roadmap includes detecting competitor changes and autonomously generating comparison pages. This crosses from optimization into confrontation. How do you design systems that can act aggressively in market response while still respecting accuracy, brand safety, and legal boundaries?
The key is to define strong guardrails in the form of system prompts and validation checks that prevent certain types of content from going live.
Accel described Flint as “Michelle-as-a-service,” effectively encoding the instincts of a top-tier growth engineer into software. Looking back at your time running growth at Warp, what are the most important heuristics you’ve hard-coded into Flint’s decision-making, and where has the system already outperformed human intuition?
There are well-established best practices around what makes a page convert well, and the Flint agent is really strong at applying them. It can look at any homepage and quickly turn it into something that’s optimized for conversion.
What’s changed is that the cost of building is now incredibly low. That means the agent can analyze massive volumes of sales transcripts and generate highly specific, granular messaging for every persona. We’ve always known intuitively that personalization drives better results but as humans, we simply don’t have the time or mental bandwidth to process that much information and create tailored pages for every repositioning or audience segment.
The Flint agent can. It’s able to programmatically generate personalized content at a scale I never could have achieved manually, even at Warp.
Many enterprise teams worry that large-scale page generation risks becoming SEO spam. How does Flint decide when programmatic content creation is genuinely useful versus when it crosses into low-quality or manipulative territory, especially as search engines and agents become more discerning?
A very good question. We don’t want to be creating SEO spam. It would simply just not do well either. Two things we do right now: We always advise our customers to review the content before going live. We also advise and make it VERY easy for our customers to provide proprietary data that they might have to enrich the agent, so that it can provide net new interesting content to the agent that we would not be able to find otherwise.
In your essay on Marx and alienation, you describe how modern work often separates people from the impact of their labor. Flint automates much of what designers, marketers, and engineers used to do manually. Do you see Flint as risking further alienation, or as a way to remove pixel pushing so humans can focus on higher-level creative and strategic work?
This is more the latter – a way of removing menial labor so that humans can focus on higher-level creative work. It is too repetitive what I’m seeing across companies with the amount of operations that needs to get done on the back-end (regarding marketing operations) or even on the design side. Designers want to be spending their time defining the brand identity, but they don’t want to be spending time extending that for a careers page. Similarly for engineering they want to be focusing on creating a system for components, but they don’t want to be designing a component each time or like changing the schema each time. Marketers don’t want to be hand-watching sales calls; they want AI to be able to pull out those insights instead.
Looking ahead to 2030, if Flint succeeds completely, what does browsing the web actually look like? Do humans still click links, or does the web become an invisible negotiation layer between autonomous agents?
There’s always going to be a dual internet. Humans are going to want to click the link to see who is the human behind the product and want to play around with the product. Websites are going to be more interactive. Agents are also going to be looking at the website, and so they need a more efficient way to view the websites. The web does become an invisible negotiation layer in that case.
Editor’s Note
This interview examines the shift from human-triggered website updates to autonomous systems driven by continuous data signals.
