Rewriting the Startup Playbook for the Age of AI

As generative AI explodes the company-building rules, founders backed by NEA identify 4 key emerging trends.

With interest in generative artificial intelligence reaching a fever pitch last fall, Perplexity AI co-founders Aravind Srinivas and Denis Yarats knew they had to act fast. They didn’t have weeks or months to hire talent, so they dispensed with recruiters, scheduling interviews and culture-fit discussions. Instead, they took a more straightforward approach:

They’d offer prime candidates a paid two-week trial.

The plan worked. Perplexity hired co-founder Johnny Ho, who joined as chief strategy officer after just a few days. Other than fourth co-founder Andrew Konwinski, all of the company’s employees have been hired this way. Besides the time saved, the process weeds out people who weren’t likely to join anyway and does a better job of generating passion for Perplexity’s mission than any speech or incentive package could match. “Now I hear that other companies are using this trial basis recruiting,” says Srinivas.

Since the release of viral chatbot ChatGPT last November, venture-backed startups like Perplexity have had to toss out parts of the company-building playbook and embrace new ways of thinking. From fundraising to product development to customer service, many time-honored practices will seem antiquated in just a few years, says NEA partner Ann Bordetsky: “We’re in the primordial ooze stage, and things are kind of messy and experimental. But you can see that people are assembling a new approach.”

It won’t be the first time a breakthrough technology has prompted profound changes in how great companies are built. Along with inventing the PC, founders like Steve Jobs and Bill Gates invented new rules for making, marketing, and selling complex digital technology. So did Amazon, Yahoo and countless Internet startups in the 1990s, Facebook and other social networking companies in the wake of the smartphone, and Salesforce, Box, and other SaaS providers courtesy of the cloud.

As one of Silicon Valley’s most established venture firms, NEA has been on the frontlines of each of these revolutionary technology shifts. And if there’s one thing the firm has learned over the past four decades, it’s that disorienting times like these often generate some of the most successful, influential companies in the world.

“Each of these technologies enabled business model innovations that weren’t possible before,” says Scott Sandell, NEA chairman and chief executive officer. The internet, for example, made open source viable and made it possible to deliver software as a continuously improving service rather than a licensed product—for free, if one chose. “It changed the model for developing software, for distributing software, and for getting paid for software,” says Sandell. “That’s the kind of impact I believe AI will have, as well.”

If anything, AI’s impact will be even greater, says NEA partner Aaron Jacobson. While previous upheavals involved how and where technology could be used, “AI is actually shifting who does the work,” he says. “That’s never happened before, so this disruption will be faster, fiercer, and bigger than ever since more is up for grabs.”

The generative AI craze is less than a year old, but so much has happened that we can already discern the outlines of how company building will change in the months, years, and possibly decades to come. To better understand this historic transformation, NEA asked four founders of AI startups in its portfolio to share examples of their new thinking. From those discussions, we’ve identified four emerging trends that will likely define company building in the unfolding AI revolution.

Emerging AI Trend #1: Nimbleness is paramount

Even compared to past bubbles, the pace at which the generative AI market is developing is staggering. Incredible new ways of using large language models (LLMs) are released seemingly every week. And tech giants and leaders such as Google, Microsoft, and OpenAI, the owner of ChatGPT, are pouring tens of billions of dollars into creating not just these models but also APIs and other tools to help innovators commercialize their products, often embracing open source rather than pursuing walled garden strategies in search of lock-in and higher margins.

Also, generative AI lets companies—not just AI startups, but companies of all types—move much faster. When Srinivas left OpenAI to start Perplexity AI, the first-time founder used the firm’s own “answer engine” technology to learn things that would have taken hours of Google searches and countless lunches and coffees with experts. “We didn’t know lots of basic stuff that founders need to know, like how to do corporate taxes. We didn’t know you had to issue 1099s to contractors,” he says.

As a result, the company is executing at a furious clip. The team released four versions of Perplexity’s chatbot based on OpenAI’s GPT-3.5 LLM in a matter of months, bringing in more than a million views a day. Then, when OpenAI released GPT-4 in mid-March, co-founder Yarats quickly called the entire company together for an extended hackathon. Two weeks later, they launched Perplexity Copilot, an “interactive search companion” that can ask users clarifying questions, run multiple searches simultaneously, and deliver more accurate results.

Pointing to the success of AI startups that have iterated quickly, such as OpenAI and Midjourney, Srinivas says “whoever executes fast is being rewarded. Speed is the constant here.”

And speed requires a different mindset. When an employee asked during a recent all-hands meeting whether they could institute measurable quarterly goals, Srinivas said he’d be glad to consider it—so long as employees were prepared to have their targets changed every few weeks.

Such nimbleness also calls for a far more flexible approach to technology and innovation. Rather than bet on particular vendors or proprietary products, companies need to be strictly agnostic about the technology they use to develop their products—and even about the products and technologies they create. When Deon Nicholas saw advances in foundation models in 2017, he started Forethought to create an AI-based customer service system. Due to breakthroughs such as OpenAI’s GPT-3.5 and ChatGPT in 2022, the company consciously walked away from much of its homegrown technology stack.

“We went through the five stages of grief, because we’d had something of a technology moat for a few years,” says Nicholas. “But ultimately, we realized that we could continue to lead the way by applying technologies like GPT-4 to stay two or three years ahead of the competition. AI is going to fundamentally change our market, so the important thing is to lead that change.”

Adaptability is imperative, says Dan Siroker, founder of Rewind AI, which makes a service to give people “perfect memory” via a searchable record of all their digital interactions. “The world is changing faster and faster, so the ability to be faster than everyone else is more important than it used to be,” he says. “That’s why we release 11 versions of the product per day. I’d even say that your ability to react and adapt is more important than your ability to proactively predict the future. That may have been what made great founders 10 years ago, but I don’t think it’s as important as being able to pay attention and listen and make better decisions each and every day.”

Emerging AI Trend #2: Productivity reimagined

With a consumer adoption rate that makes TikTok look like a slowpoke, ChatGPT has millions of people around the world rhapsodizing about how much more productive it could make them. Leading startups are already implementing changes, enabling them to get far more done with far fewer people.

This productivity boost creates a foundation for faster growth. For starters, it makes companies more profitable. One insurance technology startup saw its margins jump from 40% to over 50% after some lightweight training of its large language model.

And generative AI helps startups maintain the virtues of small size, such as agility and esprit de corps, even as they grow. For decades, studies have shown that when it comes to developing software, smaller teams are more efficient and productive. That’s one reason Perplexity accepted a $26 million A round when it could have raised far more: to prevent the company from growing headcount faster than is absolutely necessary. “There's a certain magic that happens when there's just the right kind of people working on the right kind of things,” says Srinivas.

That magic matters for recruiting, too. “There is no better way to sell yourself to an employee other than showing your fast progress,” he says. “You can say whatever you want about a roadmap and the vision, but why should anyone trust me as a first-time founder? Because the team is releasing home run software every few months.”

Ultimately, a company’s product strategy will also need to reflect this revolution in productivity. Take the notion of “seats”—licensing software by the number of people who are allowed to use it—as a way to sell business-to-business software. “If every employee is doing 10 times as much work, you’ll want to focus on the output or on whatever other unit of value the customer cares about,” says Siroker.

So how lean is lean enough? There are no clear guidelines yet. Too much is still in flux. But headcounts will be substantially smaller, says Siroker, who founded and sold digital experience platform maker Optimizely before starting Rewind. He says the company is already changing who it hires, choosing more versatile all-around players and fewer people because they have a deep expertise on a particular programming language or task that LLMs are quickly learning to do. When everyone in the company is using generative AI, “it’s like they’re all wearing an Ironman suit.”

“We're building an amazing company with 15 people,” says Siroker. “In the past we might have needed hundreds of people. It’s a different way of thinking.”

The result will be a new breed of very valuable, very small companies. “We're going to see companies that generate hundreds of millions of dollars of revenue with 25 or 50 people,” says NEA’s Jacobson. “It’s going to be wild.”

This hyper-productivity could create new problems, including a major intensification of the war for top AI talent, says NEA partner Vanessa Larco. Already, OpenAI is reportedly paying salaries of more than $1 million. “Generative AI might make a 10x engineer into a 100x engineer, but it’s not going to make a mediocre engineer much better. Like if you gave someone who’s not very good at math a calculator,” she says. “They wouldn’t know how to get the most out of it.”

Emerging AI Trend #3: Build a data moat

Ah, network effects, that wondrous phenomenon that occurs when the value of a product or service to its users grows with every new customer. Facebook is more valuable to people with three billion others on the network than it was with 3,000 or three million. Mega-market cap companies almost always have network effects.

Success in the generative AI era will depend less on who has the most customers out of the gate, and more on who figures out how to get more of the data they need to build a better product than their competitors. “It's all about the data,” says Larco. “You can be the first mover, but if you’re not getting proprietary data sets and gobbling the stuff up, it won’t matter.”

As such, founders in the generative AI era need to make data more of a strategic priority. After all, other traditional sources of competitive advantage are likely not feasible. Given the ferment of innovation around generative AI, no startup is likely to maintain a large technology advantage for long, especially when relying on popular foundation models like GPT-4 that are available to everyone. The best branding and marketing for AI-based products is the correctness and reliability of that intelligence, notes Forethought’s Nicholas. “It’s almost impossible to build a moat the traditional way, with technology or branding, for instance. The only real way to do it is with proprietary data.”

For consumer companies, success depends in large part on having a truly disruptive user experience. “We look for virality,” says Bordetsky. Daily traffic to Perplexity AI, for example, has skyrocketed in part because its “answer engine” offers not only ChatGPT-style answers to queries but also links to sources of the information.

And those user experiences need to be designed with data front and center. “Companies need to be especially great at aggregation, user growth, and engagement to continuously refine their AI-powered product, because that's the real flywheel for the business,” says Bordetsky.

A killer user experience has been Nicholas’ focus since he launched Forethought in 2017 to create a better conversational chatbot for customer service. The company invested to develop the plumbing software to easily integrate data from incumbents such as Confluence, Salesforce, and Zendesk to inform its chatbot. And in March, it rolled out a service called SupportGPT based on OpenAI.

As a result, the company now has more than 50 integrations, which has helped it sign on more than 100 customers, including Marriott and Instacart. Nicholas says the company is building a reinforcement learning system to capture even trivial-seeming interactions, such as whether a help desk agent actually used the Forethought systems recommendations, and whether the end customer was satisfied with that answer.

“Ultimately, it’s about creating a system of intelligence,” he says. To set up a customer service operation today, companies hire agents and arm them with customer data from traditional systems of record. “In 10 years, the first thing companies will do is AI to build a system on top of that proprietary data. So a help desk agent could simply ask ‘what was that interaction we had with Deon a few weeks ago?’”

Emerging AI Trend #4: Think bigger

No technology in history has taken off as fast as generative AI—not search, not the smartphone, and not social media. Unsurprisingly, the level of competition has risen just as fast.

As of May 2023, Dealroom counted more than 250 generative AI startups, and tech’s giants are focusing on the technology like their future depended on it, which it probably does. And because they already have reams of precious data, the incumbents have an advantage over their forebears in holding off the next generation of disruptors. Given the whirlwind environment, “this is not a time for incremental thinking,” says Larco. “You need to do something that's so different and so obvious that people will look back and say, ‘I can’t believe we used to do it that way.’”

That means dispensing with piecemeal product improvements, and instead focusing on creating entirely new categories. For Rewind founder Dan Siroker, the moonshot is to give people that gift of “perfect memory.” Once a person has agreed to give access to their digital activity—the websites they’ve gone to, the texts they’ve sent, what was said on Zoom calls—Rewind’s app can retrieve any interaction even if the person only recalls a few key words that were typed or spoken. (It only works on recent Macs, for now, as they are dependent on Apple’s M1 and M2 chips.)

The plan required coming up with compression technology that crunches data down by more than 3,000 times, making it possible to store years of data on the user’s device. They also had to figure out guardrails to deal with the obvious privacy concerns (for example, all the data remains on the device and is never stored in the cloud).

But a formidable challenge was part of the attraction. “I looked at this less like a business plan and more as a problem I want to spend the rest of my life on,” says Siroker, who compares Rewind’s memory-preserving capability as a “superpower” similar to the hearing aids that restored his hearing when he was going deaf in his 20s. “It was more selfish than strategic.”

Siroker was careful to find investors he felt sure would support his mission. That led him to a group led by NEA, who he praised for our record of investing in companies through all the inevitable ups and downs—sometimes even buying rather than selling at the IPO. “The folks at NEA understand the odds of success at pulling this off, and they’re in anyway,” says Siroker. "They're interested in building a company with long-term value. That's one reason they're one of very few firms that have succeeded in every generation of technology."

Siroker ascribes to the view that it’s easier—not harder—to execute a bold business plan. “The more audacious the idea, the more excited the employees and potential investors are,” he says. “And it helps me stay excited about what we're doing, as well.”

Perplexity’s Srinivas agrees that truly big missions can inspire great people, but stresses that it needs to be backed up with real accomplishments. “I’m realistic that we’re not going to be able to compete with Google on compensation. Nobody can,” says Srinivas. “But right now there are a lot of very talented technical people who are extremely bored at places like Google, and they’re looking for their opportunity to make their mark on the world.”

Crowdbotics is another NEA portfolio company using generative AI to pursue a lofty goal. Simply put, the company intends to leverage AI to reinvent the entire software development process. While scores of low code and no code competitors were creating tools so non-techies could create software for simple applications, founder Anand Kulkarni believed that the foundational models that were just beginning to be discussed in academic journals in 2016 would one day let application development teams accomplish far more using natural-language commands.

“Any software engineer will tell you that writing the code is the easy part,” says Kulkarni. Far more difficult is figuring out what the software should do, and how to express that in terms a computer can understand – not to mention despised tasks such as writing documentation and ensuring that the resulting code is secure. Plus, Crowdbotics has developed processes so companies can save all the software its developers produce into a catalog of reusable components.

Ultimately, the vision is to “drive the incremental cost of developing software to zero” by letting companies produce production-ready code in hours or days, versus the months or quarters it now takes, says Kulkarni. Within a decade, developers will be able to describe the software they’re trying to create, and the company’s “code operations” system will stitch together existing components and help developers write the rest and build in the necessary security, privacy, and bias guardrails.

The process will never be totally automated, as humans will have to review and troubleshoot corner-case issues. Even so, the approach is already freeing up latent innovative power for Crowdbotics’ customers. Through the decades, most ideas for software-based innovation have been shot down, usually over concerns about the cost of engineering. “There’s so much friction just getting ideas to the starting line that most ideas never get considered,” says Kulkarni. “We’re changing those economics by making it efficient and easy.”

Not so long ago, Kulkarni didn’t declare this mission too forcefully. The company downplayed its use of generative AI for fear prospective customers would scoff at the idea that AI could handle anything as complex as building enterprise software. “We’ve had a demo for years to show how our system could build an application for a customer in under a minute, but we rarely used it because we didn’t think the customer would believe it.”

Not anymore. With so much interest in generative AI, the one-minute demo is “front and center in how we talk to customers,” and the company has revamped its marketing and product roadmaps to play up its generative AI capabilities.

“The mindset of the world has changed, and the market is suddenly hungry for what we’ve been building,” says Kulkarni. “When you have moments like this, you need to rise to meet them.”

The information provided in this blog post is for educational and informational purposes only and is not intended to be investment advice, or recommendation, or as an offer to sell or a solicitation of an offer to buy an interest in any fund or investment vehicle managed by NEA or any other NEA entity. New Enterprise Associates (NEA) is a registered investment adviser with the Securities and Exchange Commission (SEC). However, nothing in this post should be interpreted to suggest that the SEC has endorsed or approved the contents of this post. NEA has no obligation to update, modify, or amend the contents of this post nor to notify readers in the event that any information, opinion, forecast or estimate changes or subsequently becomes inaccurate or outdated. In addition, certain information contained herein has been obtained from third-party sources and has not been independently verified by NEA. The companies featured in this post are for illustrative purposes only, have been selected in order to provide an example of the types of investments made by NEA that fit the theme of this post and are not representative of all NEA portfolio companies. The company founders or executives or any other individuals featured or quoted in this post are not compensated, directly or indirectly, by NEA but may be founders or executives of portfolio companies NEA has invested in through funds managed by NEA and its affiliates. Any statements made by founders, investors, portfolio companies, or others in the post or on other third-party websites referencing this post are their own, and are not intended to be an endorsement of the investment advisory services offered by NEA.
NEA makes no assurance that investment results obtained historically can be obtained in the future, or that any investments managed by NEA will be profitable. To the extent the content in this post discusses hypotheticals, projections, or forecasts to illustrate a view, such views may not have been verified or adopted by NEA, nor has NEA tested the validity of the assumptions that underlie such opinions. Readers of the information contained herein should consult their own legal, tax, and financial advisers because the contents are not intended by NEA to be used as part of the investment decision making process related to any investment managed by NEA.