Synthesia has a great story, going from 0 to 50,000 customers. Accel Partner Philipe Botteri and Synthesia’s co-founder and CEO Victor Riparbelli deep dive into the lessons learned about building an Enterprise-focused Generative AI company and scaling it.
Where did Victor’s journey begin? Let’s find out.
Synthesia’s Beginnings
During his childhood years, Victor had a love of computers and sci-fi. He learned you could turn those interests into a career during his teenage years. He started building websites for local businesses and graduated into working in the Danish startup ecosystem.
His sweet spot was thinking about the product and how that connects with the customer. From there, he decided to build his own company. Copenhagen wasn’t the ideal launch pad, so he moved to London, where he met one of his co-founders doing a research paper on a neural network generating highly realistic video.
After seeing that magic, Victor had two thoughts.
- Producing video content would be possible behind your desk, with no cameras, microphones, lighting equipment, etc.
- Once that was possible, you could generate video via code.
New technology always breeds new media formats, and for video, that probably means less linear and more personalized and interactive.
The Long-Term Vision
What will be possible with this technology in 10 years? Going from “Wow, this is a cool tool” to actually getting rid of all hardware is a process, and Victor believed we would be there within ten years of the company’s conception. When you set a long-term vision, how do you sequence a company along the way with a product customers will buy now?
Synthesia’s team first thought about what they could realistically build and sell within 12 months, the length of their runway. They began with AI dubbing, dubbing real video into different languages and changing the lip synchronization and voice-over.
They went to video production professionals, agencies, and even Hollywood, but the tech hadn’t matured enough in 2017. They were selling a vitamin, not a painkiller.
After a couple of years of this, they sat down to determine if it was the right path for the business. Did they want to become a visual effects studio with proprietary technology? No, they didn’t.
Learnings From The First Three Years
The people they were selling to, their houses weren’t on fire. They learned there are millions, maybe billions, of people not making videos today who are desperate to. Their house is on fire, and these people range from a random person on the street wanting to start a YouTube channel to the world’s largest companies wanting to make video content but without the budget to do so.
The thesis for Synthesia was to focus on all of these people and give them faster, more affordable solutions, which led them down the path of developing avatar tech that they’re famous for today.
It took three years to hit this inflection point despite the product almost barely working. There was such a big pull from the market that they wanted a better way of making videos without cameras, and it didn’t matter if they were of lower quality.
The people interested in the product were in two camps:
- The ones who thought it was cool and fun.
- The ones with a real business need.
The second group of people took Synthesia down the path of a B2B SaaS platform for video creation, and today, they’re the biggest in the world doing it.
Fast food companies that need to train a lot of people to work at their restaurants could create videos instead of a 40-page handbook. If you have 2,000 salespeople globally, you can inform them about pricing and the competitive landscape through bite-sized videos instead of long, boring emails.
How Mark Cuban Became an Angel Investor
In 2017, they tried to raise their first round. Victor thought, “This is amazing. It’s going to be so easy to raise money. We clearly have an insanely cool idea and a great team!” Of course, no one loved it and likely thought they were crazy. They got turned down by around 100 investors because the tech was very early.
They almost gave up until another co-founder sent a cold email to Mark Cuban, who sat at the corner of media and technology.
He responded back in 5 minutes, and they proceeded to have a 14-hour exchange over email. Mark doesn’t do calls. Around 4 a.m. U.K. time, he agreed to invest a million dollars into the company and kickstarted everything.
You need grit as an entrepreneur and founder, especially if you’re doing something that isn’t what people are funding at that moment in time. You need to find someone who already shares your thesis because it will be difficult to convince someone cold.
How AI Companies Differ From Traditional SaaS
In software, two things matter. You need to build and sell a product. In an AI company, you need to build and sell a product, but before building, you need to do the research. You need an engineering team and a research team, which requires a different budget and types of people to manage.
Back when Synthesia started, they were heavily focused on AI only. If they couldn’t make that work, they couldn’t make anything else work. Once they built the first product and hit the inflection point, they hired a team to build the whole SaaS platform.
Looking back, they would have started the product engineering side earlier if they had raised more money. What’s unique to AI businesses is that if you want to do the science, you should because it unlocks a new market. The biggest issue with research is that there’s so much uncertainty.
Running a Team of AI Researchers
What Synthesia has found to work is that you almost need to take a VC-like approach to your research. There’s a problem you’re trying to solve, and many different ways people try to solve it. You don’t know which way will work, so you need a spread of different teams working to solve the same problems from different angles.
It gives you wide coverage and hedges against not picking the right direction. Once you find something that works, you have to be good at killing your darlings. Shut down the other projects and pull those teams into that project.
This is different from a traditional SaaS app. You wouldn’t have three different teams trying to build a billing system. With AI companies, you’re dealing with a lot more uncertainty.
How to Evaluate an AI Research Team
When you have people knocking around ideas and experimenting as their job, how do you evaluate them on performance? It’s really hard! You can look at velocity and how smart someone is about how they do things.
Many researchers will naturally be drawn to the hardest problem, and it’s not always the most valuable thing for your business. For example, at Synthesia, they’re building avatars. The research team has found things that work, but they’ve also figured out that for people with a big red beard, sunglasses, and a pirate hat, the tech doesn’t work.
Researchers might say, “Oh, that’s interesting,” and get sucked into trying to solve that extreme outlier case that doesn’t matter for your business. To measure performance, you can see how well the IC and managerial team think about the product and make choices.
You can measure how many experiments they run per week or month and how well they evaluate them. If you don’t measure anything, you risk having a high school science lab with no oversight.
How to Strategically Focus AI Research
A lot of things are being developed in AI, and you could use off-the-shelf products or push the boundaries of science. How do you strategically focus the research when building today?
For many years, there was no choice but to build it yourself. There wasn’t anything on the market, and the open-source community was working on things different from what the Synthesia team needed. “Today, you want to be very intentional about what narrow slice of AI you can be uniquely the best in the world at,” Victor says.
The bar is high, and you probably won’t be the best at building a fully generalized LLM model unless you’re Anthropic, OpenAI, or Google. So, instead of trying to be the best at general AI video, Synthesia focused on being the best at people presenting to the camera.
Instead of trying to be the best at a broad category, hone in on that narrow slice of whatever you’re solving for and be the best at that.
Off-The-Shelf vs. Building It Yourself
With OpenAI and the hyperscalers, companies no longer need to do as much research. Will there always be some areas where you should continue researching and building yourself?
“The world is messy and complex, and when you build systems that work in production, there are a million different things you need to be able to do,” Victor shares. “There will be value in research for a long time.”
You need to be intentional about whether you want to be an infrastructure layer where you truly have to be the best in the world or if you see yourself in the application layer. If you’re operating in the application layer, there are advantages to building yourself because it provides the opportunity to take things to market first.
Maybe that matters less over time, but it’s probably worth spending millions of dollars in research if you can go to market 6-12 months before competitors.
You’re Not Just Building an AI Model
When you build a product, the sole value of what you’re building is not only the AI model. If you take ChatGPT, an interface to a model, OpenAI needs to be the best at the thing. But if you’re building a customer support chatbot that can respond to customer inquiries, that technology is more than the model itself.
Your customers aren’t buying an AI interface. They’re buying a full solution that sits around it. The traditional rules of business don’t change. You want to build a platform that solves a problem for your customers and doesn’t just give them a piece of technology unless you’re an infrastructure company.
Data Will Be Your Advantage
Data will give you a competitive advantage. We see this now when people train massive models that are powerful and capable, and people are trying to make them good at specific tasks.
This work mainly involves curating the data you train the base model on, and fine-tuning it to the task you’re trying to accomplish. The publicly available data on which these models are trained don’t have much information.
There probably aren’t a million MRI scans or confidential spreadsheets from big businesses. That’s the data that’s valuable and not available to the public, which will be important going forward.
There are two parts to data that matter:
- Your existing data.
- Acquiring new data
Curated data is where we’ll likely see the most progress in all modalities over the next 12 months.
AI Adoption in the Enterprise
How do you make it happen? Distribution is almost as important, if not more important, than building the product. There’s always some channel that’s significantly better than the others and constantly changes.
When Victor started in tech more than ten years ago, SEO was all the rage. Many billion-dollar companies use SEO as the primary way of driving traffic. In AI, it’s clear that word-of-mouth is the channel.
There’s a magical element to AI, a huge wow factor. People feel threatened by it, want to understand it, and want to post about it on LinkedIn. This last 12-18 months has seen a lot of virality. To achieve virality, you need to make your product:
- Accessible
- Fun
- Shareable
You can build a word-of-mouth groundswell if you’re good at those things. Most successful AI companies today have used an element of virality because that distribution channel has been strong.
From Inbound to Outbound
Synthesia first became a force of nature through word-of-mouth and virality on social media. People wanted to pay for it and were excited, but most of those people are known as AI tourists. They’re willing to pay to try out the product, but they will likely churn.
So, you want to build a business for the long-term use cases while drawing those people in through virality. Phase one is the viral moment to draw in interesting people. Phase two is outbound with paid campaigns to target the specific ICP you discovered during the first phase.
Doing outbound might be less effective if you aren’t sure who your target ICP is. So gathering as much information as you can about them early will help you shape cold outbound for the right user later.
The Right Channels For an AI Business
As soon as you feel like you have actual product market fit, you need to think about how to build a scalable distribution engine. You don’t want to be in a situation where that moment hits; you’re being flooded with interest and not ready.
Eventually, the virality moment ends in a month, six months, a year, or two years. When that ends, you don’t want to leave your sales team with no inbound. Start thinking about your distribution strategy before that so you’re in control of your own destiny.
Whether in AI or any other business, you should start doing SEO if people are searching for your solution. If you’re building a new category, you do SEO early on. Even if it’s less impactful in the short term, you could own important keywords when they become mainstream.
Your job is to harness word-of-mouth, especially in AI, where there’s so much public interest. But don’t get pulled in too much by the hype where you’re only focused on short-term experimental revenue vs. long-term Enterprise revenue.
Long-term revenue is more painful, but it’s essential to invest in it.
Key Takeaways
- Focus on driving utility over novelty. Short-term revenue is great and can be positive for your business, but don’t lose sight of what’s valuable for your customers long-term.
- Be intentional about when to do the science yourself, when to implement something open-source with a timeline around it, and when to buy off the shelf.
- Think about what’s going to happen to the market in the future. The AI landscape is changing rapidly, so you want to stay ahead of the curve.
- Every new market has growth hacks. For AI right now, it’s creating viral content that’s accessible, fun, and shareable.
The post How to Build a GenAI Company for the Enterprise with Synthesia’s CEO appeared first on SaaStr.
via https://www.aiupnow.com
Amelia Ibarra, Khareem Sudlow
