How AI is changing startup leadership, and what founders can do about it.

Startup CEOs today are caught between two realities.

On the one hand, investor expectations for AI startups are sky high. Growth metrics that were considered exceptional in the previous era of SaaS – the 3x, 3x, 2x, 2x, 2x playbook (“T2D3”) – now seem slow. Bessemer has coined “Q2T3” (4x, 4x, 3x, 3x, 3x) as the new 5-year growth benchmark that AI startups should strive for.

The highest velocity AI startups – what Bessemer calls “AI supernovas” – have scaled to $100M ARR in an average of just 18 months. Lovable hit $100M ARR 8 months after launch; Cursor did the same in 12 months. ChatGPT reached 100M users in 2 months. These are the fastest growing apps in history.

At the same time, the business reality remains challenging. Most startups are nowhere near these hyper-growth curves. The vast majority of enterprise customers are still in the early innings of AI adoption – often still experimenting or running pilots. Among the startups we see, there are lots of proof-of-concept demos, but few large production contracts. Customers are excited by AI’s potential, but they often don’t yet know what they actually want or how to integrate AI into their workflows.

So how should CEOs navigate these tensions?

8 principles for leading in our AI moment

Earlier this year, we shared our thoughts on how app-layer AI companies can stay defensible as model providers move up the stack – in short, how you should approach building your product. This month, as a follow-up, I want to share a more personal set of learnings on how to operate differently in the age of AI: how you should show up to work every day as a leader.

This topic has been on my mind since our CEO summit in Bodega Bay. Spending three focused days with our portfolio CEOs drove home the challenges of leading in this moment. The truth is, every software company needs to be “refounded” for our AI era. Even early-stage startups are reworking how they create culture, design their orgs, and make decisions.

What follows is an initial set of “new rules” for CEOs. Some of these are timeless startup truths that take on new meaning and urgency today. Other principles are truly new, emerging from the experience of startups in our portfolio.

I’m particularly grateful to Jonathan Siddharth (Turing) and Mohit Aron (Nutanix/Cohesity). Both participated in our CEO summit, and I share some learnings from their sessions and our follow-up conversations here.

🏋️‍♂️ 1. Hire people who act like owners

Your culture is defined by the people you hire and the behaviors you reinforce. The pace of AI startups is so fast that you need people who will truly do “whatever it takes.” To build this high-intensity culture, hire for an ownership mentality.

Jonathan actively screens for what he calls “raw founder energy.” He hires “less for experience, more for exceptional ability…someone that’s really hungry, intense, hardworking.” In practice, this means Turing might pass on a polished veteran if a scrappier candidate shows more intensity and an eagerness to own outcomes.

You want team members who care so deeply that they never consider any problem above or beneath them. That creates a culture where everyone pushes hard and actively works to make things better, rather than saying “that’s not my job.”

Intensity is contagious. When new hires see their peers working with urgency and taking ownership, it sets the norm. Over time, a virtuous cycle kicks in: people who don’t like this level of intensity tend to opt out, and the ones who thrive in it stick around and attract like-minded talent. We’re seeing CEOs in our portfolio become much more comfortable signaling this publicly: essentially saying, “Our culture is not for everyone. But if it is for you, you can do truly exceptional work here.”

🥊 2. Lean into confrontation

Pushing for speed and excellence will almost inevitably lead to conflict – and that’s okay. As CEO, you have to be willing to bring confrontational energy when needed. Frank Slootman (a 3x guest on the B2BaCEO podcast) stresses that great leaders “exploit every opportunity to step up the pace and expect a higher quality outcome. … Yes, it is confrontational. That is pretty much what CEOs do all the time: confront people, issues and situations.” In practice, this means holding a high bar, pressing for specifics, and refusing to smooth over disagreements just to keep the peace.

Jonathan shared that he had to work to expand his personal “stress envelope” in order to lean into uncomfortable conversations. “You have to push people hard. You have to push the company hard. That’s the only way. … Each intense meeting raises my tolerance. Now I can do them without breaking into a sweat.”

Being confrontational doesn’t mean being a jerk. The key is to critique ideas and decisions, not people. When giving hard feedback, be specific, actionable, and contextualized. Clearly outline what happened, why it matters, and what could improve – always with concrete examples.

When you normalize candid, respectful debate, your teams start arriving at the truth faster. Over time, people learn that a bit of heat in the discussion leads to better outcomes, and they learn not to take it personally. You also build a culture where giving feedback is reserved not only for crises or annual reviews: it becomes part of the daily course of working together (and that includes positive feedback for wins, not just negative feedback for issues).

The result is an org that’s able to debate ideas without lingering resentment, all in service of making the best decisions as fast as possible.

🤖 3. Automate your job as CEO

As a founder, you set the tone for how aggressively your company embraces AI – including in your own workflow. By automating parts of your own job, you send a strong signal to the team: we all use AI to increase our impact, not just as a buzzword.

Jonathan modeled this behavior by spinning up a skunkworks project at Turing to automate the CEO. “In an ideal world, I’d be synthesizing data from sales, engineering, product – but I’m fundamentally token-constrained in how much I can process,” he explains. So he’s trained an AI agent to do the legwork: it ingests and analyzes all of the company’s operational data – Salesforce reports, Looker dashboards, spreadsheets, Jira tickets, Slack updates, Zoom transcripts – and drafts a weekly briefing. The agent even writes code to answer questions. (For example, when Jonathan asked which sub-business was growing fastest among Turing’s AI lab customers, the agent wrote a Python script to analyze Salesforce data and create a chart.)

The impact has been impressive. “My weekly exec meetings have become way more intense,” Jonathan shares. “I show up armed with very pointed, data-driven questions that are very in the weeds. I can ask ‘Why is Project X with Lab Y two weeks behind schedule?’” With the help of AI, Jonathan is able to catch issues that would have previously not bubbled up and hold his team accountable to address them quickly.

Importantly, Jonathan hasn’t removed himself from the equation. Every time he catches the agent misinterpreting something or lacking context, he gives feedback and corrects it. In essence, he’s training his digital twin to think more like him. It’s a virtuous cycle: AI is making him a better CEO, and he’s making his AI agent a better assistant.

To sum up: Lead by example in adopting AI. Show your team what’s possible. Create the norm where AI isn’t seen as a threat but as a force multiplier.

📋 4. Treat AI like a direct report

Management is about delivering outcomes: you have a goal, and you coordinate resources (people, tools, processes) to achieve it.

In the AI era, the same fundamentals apply – but your team now includes AI agents.You can’t just sprinkle AI into your workflow and expect magic. You need to set it up for success (with context) and create a feedback loop that allows it to continuously improve. This means setting goals for your AI systems, defining what success looks like, reviewing their outputs, and giving feedback – much like you would with a human team member.

Jonathan’s “CEO agent” is a great example. The same principles apply to other AI tools. Don’t just toy with ChatGPT to draft an email and then ignore it the rest of the week. Instead, integrate AI into your workflow with defined responsibilities: e.g. “This GPT-4/5 chatbot drafts all of our sales emails,” or “V0 creates the first set of design mockups for new feature ideas.” Then actively manage that integration – review the outputs, refine the prompts and data you feed in, and save successful results so those gains can compound over time.

↔️ 5. Flatten your org

Across our portfolio, “the flatter the better” is a common mantra. Traditional boundaries between roles (product writes specs, design makes mocks, engineering builds, customer success deals with users) are giving way to something more fluid. At an AI-first startup (especially an early-stage one), every employee should ideally view themselves as a builder first and their functional title second.

Jonathan, for example, has dramatically flattened Turing’s org. Between him and every entry-level engineer, there’s only one layer of management, and that manager also writes code. Engineers shadow users on calls, designers comb through data, and salespeople contribute to the product roadmap. Everyone feels directly responsible for what ships. When an engineer hears a customer struggle with a feature, they don’t toss it to a PM – they go fix it.

This flat structure speeds everything up. Without long handoff chains, there’s less lost in translation. Communication becomes tighter and more direct. Teams develop a stronger sense of ownership because no one is waiting for someone else upstream or downstream – each person is responsible for solving the problem in front of them.

One founder told me that once they cut the PM layer, a two-person engineer-designer team began shipping meaningful updates in days instead of weeks. They didn’t write formal proposals or wait on green lights; they just built, tested with the user, and iterated.

This isn’t to say that expertise isn’t important. People should still lean into their strengths. A great designer shouldn’t be expected to suddenly write backend code, just as a backend engineer shouldn’t be judged on their UI mockups. But culturally, in an AI-first org, everyone is expected to care about the whole product and to collaborate across roles without friction. You don’t throw something over the wall to another department and forget it. You stay in the loop because it’s all your job. (This ties back to the “owner” mentality in principle #1.)

The leaner and more blended your team, the faster it can learn and adapt.

🚧 6. Prototypes over decks

“Prototypes over decks” is another mantra for high-functioning AI product teams. In some startups, it’s still common to spend weeks building strategy decks to propose new features. Today, that approach is too slow – you need a much tighter feedback loop with reality (both the tech’s reality and users’ expectations).

Many founders I work with now insist on a working demo when evaluating a product idea. Prioritizing prototypes forces your team to engage with technical constraints and user experience early. Even if the prototype is held together with duct tape, it will advance your understanding far more than weeks of hypotheticals. In a space moving as fast as AI, it’s far better to test, learn, and iterate with real software than to polish a document about what you think will happen.

This culture shift can be summed up as “play with the product, then plan” (as opposed to “plan, then build”). It makes the development process far more iterative and experiment-driven. You can still have a roadmap, but it should evolve constantly based on what you learn from hands-on prototyping.

Admittedly, moving fast to prototype can create some chaos. Teams might find themselves juggling many experimental branches without a clear process for how prototypes are reviewed, stored, and graduated into production. (Where do all those mini-apps live – scattered in personal accounts or centralized? How do you document and share learnings from each experiment? How do you ensure these AI-driven features feel cohesive, not like a patchwork?) But these are good problems to have. It’s easier to add on structure than to fix a culture that lacks innovation.

You can add process and polish later, but you can’t afford to learn slowly now.

⚙️ 7. Build systems – the machine that builds the machine

If your underlying company machine (your org and processes) is well-constructed, AI will make it run even faster. If your machine is ad-hoc and chaotic, AI will accelerate the chaos. As a founder, your job isn’t just to build a great product; it’s to build a company that can repeatedly build great products.

In other words, you have to build the machine that builds the machine.

Mohit offered the analogy of driving in California vs. driving in India. In California, the system (traffic laws, enforcement, road design) makes it easy for everyone to drive safely – even a mediocre driver can get from A to B without much trouble. In parts of India with looser rules, even highly skilled drivers are constantly avoiding near-accidents. As CEO, you want to create the California freeway version of your business, not the free-for-all.

Mohit’s approach is to implement core systems early:

• Hiring: The first system he establishes is for hiring. He clearly defines the skills and qualities needed for each role, designs an interview process that rigorously tests for those traits, and trains every interviewer how to evaluate candidates consistently.

• Performance management: Within the first six months, Mohit rolls out a performance management cadence, including quarterly OKR reviews and regular check-ins. This creates a rhythm of accountability and feedback from the get-go. If someone isn’t performing or isn’t a cultural fit, you’ll notice by Q2 and can address it (or let them go) before a small issue spirals. It also forces you to articulate what success looks like (“what are our objectives this quarter and did we hit them?”), which many first-time founders avoid until it’s too late.
• Reference checks: Mohit approaches reference calls like a detective. He asks references to rate the candidate on a scale of 1-10 on multiple dimensions, and he forces specificity. “Would you enthusiastically hire this person again? Okay, what’s your rating? 6 out of 10? Interesting – why not a 7? and why not an 8?” By pinning them down to a number and then probing (“why not higher?”), he flushes out real reservations. If a reference isn’t glowing, he’ll dig for names of other people who experienced the candidate’s weaknesses and do blind reference checks with them. The process isn’t complete until he has a complete picture of the potential hire’s strengths and weaknesses.

With these examples in mind, ask yourself: What are the key processes in my business, and do we handle them ad hoc or systematically? If the answer is “mostly ad hoc” (which is normal for an early startup), think about the first lightweight system you can put in place. It could be as simple as a standing Monday meeting to review priorities (a system for focus), or a shared dashboard of metrics everyone updates weekly (a system for data visibility). Iterate on your company’s internal operating system as intentionally as you do on your product.

Even culture can be systematized to a degree. For example, if you want a culture of transparency and learning, institute rituals like a Friday wins/fails email or a monthly “what we learned” all-hands where people openly discuss mistakes and learnings. If you want a culture of ownership, create a process where every project has a directly responsible individual who updates the whole team on status. Small systems like these help make the desired behavior the default.

🏃‍♀️ 8. Learning velocity above all else

You can’t control the timing of the next research breakthrough, but you can control how quickly your team absorbs information and adjusts. The single biggest driver of long-term success is your startup’s rate of learning.

Jonathan puts it bluntly: “In the race to superintelligence, there is no such thing as product-market fit. Every quarter you have to re-earn your position, because the models are getting smarter.” You might have something working with today’s models, but by the time GPT-6 comes, your solution will need to get better or more specialized, a competitor might do it cheaper, etc. The only sustainable advantage is learning and adapting faster than everyone else.

One way to speed up learning is to treat every “no” not as a stop sign but as a signal about what’s needed for a future “yes.” When a customer rejects you, follow up by asking under what conditions it could be a yes – get the prospect to articulate in their own words what’s missing or not convincing yet. Sometimes you’ll find a path forward. Even if not, you’ve learned what the market values. Encourage a culture where no failure is wasted – every outcome (good and bad) is mined for insight.

Another key to moving fast is developing a bias for action. Jeff Bezos distinguishes between “one-way door decisions” (hard or impossible to reverse) and “two-way door decisions” (easily reversible). In any scenario where you can course-correct if needed, it’s best to act fast and learn.

Turing’s experience in 2022 is a helpful case study. At the time, Turing’s core business was a marketplace for remote engineering talent. Then OpenAI invited them to help train what would become GPT-3. In the meeting, OpenAI asked how many engineers Turing could dedicate. Jonathan’s gut thought “maybe 100” (since most client projects only used a few engineers). But, in the moment, he said, “We can give you 1,000.” OpenAI accepted the offer – and in hindsight Jonathan laughs that he should have said “10,000.” That moment of conviction – essentially betting the company on a new opportunity – changed Turing’s trajectory. In the years since, that decision opened doors to work with nearly every frontier AI lab.

When you see a promising two-way door, don’t hesitate or hedge – walk through it. Your first move in a new direction doesn’t have to be perfect. The goal is to accelerate learning – which means creating as many opportunities as possible to refine your trajectory based on the data you generate.