Building, Selling, and Adopting AI in 2025

Our Top Takeaways from the 2025 AI Agent Conference

The AI-native wave is here, and it’s not slowing down. At the 2025 AI Agent Conference in NYC, AI investors, founders, engineers, and more shared their insights on staying sharp in an environment where “cutting edge” has a shelf life of about six weeks.

Here were our top takeaways for founders looking to build, scale, and survive in this new AI-native era:

1. Build Your AI Stack to Bend, Not Break

In AI, the only constant is change. The models, frameworks, and tools you use today might be outdated tomorrow. 

“What a lot of small to mid-sized teams are struggling with is how to set up and integrate different AI components and create an overall secure production environment, and then still stay cutting edge when the tools they’re using become obsolete,” said Misha Herscu, CEO of Cake.

💡 Takeaway: Don’t chase permanence – chase resilience. Architect for change by building a solid foundation and keeping the cutting-edge swappable. Separate stable infrastructure (like security, auth, and data pipelines) from fast-moving AI components (LLMs, vector stores, orchestration).

2. Hiring? Assess for AI-Tool Proficiency – Not Just Coding Chops

Modern engineering talent looks different. It’s not just about how well someone can code – it's also about how well they can leverage tools like Copilot, Cursor, or Replit to move faster.

💡 Takeaway: Update your interview process. Watch how candidates actually use AI dev tools. The most effective engineers are the ones who can wield AI as an accelerant – not just write code from scratch.

3. When Building AI, Internal AI Teams Need Breathing Room to Stay Ahead

In tech teams, there's typically a tension between short-term shipping and long-term exploration. The most competitive teams are intentionally carving out space for a dedicated AI team that doesn’t deal with day-to-day customer noise, but instead tracks papers, runs experiments, and attends conferences to bake insights into the product roadmap. 

“Isolating our AI team – not necessarily in a silo, but allowing them to focus on longer-term roadmaps – has completely changed the way we’ve been able to think about AI. Our AI team gets a little bit more space intellectually,” said Ali Hussain, CEO of Tabs. 

💡 Takeaway: Create an “exploration buffer.” Don’t force every engineer onto the shipping treadmill. Dedicate a slice of the team to chasing what’s next, and give them the runway to validate, integrate, and operationalize their findings.

4. Go-to-Market Motions for AI Must Match Buyer Maturity

With buyers currently at various stages in the AI adoption cycle, selling AI today is not one-size-fits-all. Some buyers want to move fast and break things. Others – especially in regulated or risk-averse sectors – need hand-holding, sandboxes, and slower onboarding to build trust.

💡 Takeaway: Calibrate your sales motion to the sophistication and risk tolerance of your buyer. Involve stakeholders early. Where necessary, guide them through technical validation, procurement, and change management with patience.

5. Proof-of-Concepts (POCs) Must Be More Than Experiments

While sandboxes and demos are important in selling, many AI tools get stuck in endless pilot purgatory – experimental run rate revenue (ERR) without durable ARR. Cassie Young from Primary VC explained: “People are very into trying things, but not all of those solutions will stick. This makes our jobs as investors very hard in telling what will lead to durable revenue.” 

Startups are breaking out of this by building full-cycle systems and tightly scoping POCs around real business value. The goal isn’t a "wow" – it’s a win that sticks.

💡 Takeaway: Define measurable outcomes upfront. Start with a clear pain point, nail the initial use case, and convert quickly to production. Build full-cycle, production-ready solutions with memory, feedback loops, and tight workflow integration. And once you’re in, optimize for stickiness by embedding into workflows that are hard to rip out.

6. AI Agents Aren’t Plug-and-Play – They Need High-Quality Context and Structure

As Arjun Prakash, CEO & Co-Founder of Distyl AI, said: “Don’t expect the AI to know what you didn’t bother to explain.” You can’t just drop an AI agent into yo

ur business and expect magic. That’s like hiring a brilliant human with no context, no team, and no clue what they’re supposed to do. You need to define clear roles, workflows, and expectations for agents — just like you would for a new hire.

💡 Takeaway: It’s not just about providing AI more data, it’s about proving AI high-quality data. That means cataloging your processes, decisions, and systems, and making them machine-readable. If your systems are a mess, AI will only amplify that chaos.