What do AI startups need to consider to successfully partner with enterprises today, and how should they be thinking about the future AI tech stack? During NY Tech Week, Techstars brought together portfolio companies, investors, and corporate innovation leaders to answer these questions. 

First, Kerty Levy held a fireside conversation with Paddy Srinivasan, CEO of DigitalOcean (Techstars 2012), on enterprise AI transformation. Then we hosted a panel, moderated by Babak Kia, Senior Lecturer at Boston University and Techstars All-Star Mentor, on the future of the AI tech stack, featuring AI experts and thought leaders, including: 

• Paddy Srinivasan, CEO of Digital Ocean

• Nick Lordi, Partner, PwC’s Responsible Artificial Intelligence Practice

• Rafael Barroso, Engineering Manager, Apple

• Waleed Atallah, Co-Founder & CEO, Mako

The conversations revealed what makes enterprise AI partnerships work right now and what innovation to look for in the future:

Target outsourced functions, not internal roles.

The smartest enterprise pitches target outsourced functions rather than employees. As Paddy explained, startups can replace the expensive outsourcing arrangements with the pitch, “You're spending $100 on offshoring this service. We can do it for $10."

The key is choosing functions where mistakes don't threaten core business operations. Paddy shared an example from a startup founder who chose accounts payable automation over accounts receivable because companies view these functions differently. Payment errors affect external vendors, while collection failures directly impact company revenue.

This insight goes beyond just business risk. It’s also about understanding how people react to perceived threats. Paddy advised founders not to “underestimate the cultural resistance from companies. Most humans have a protective, defensive nature. That's just how they're wired."

The advantage is that you’re “not really threatening anyone's job except for the outsourced service provider." This allows startups to avoid triggering the internal opposition that kills AI initiatives.

Build systems, not features.

Adding chatbots to existing software and calling it “AI transformation” is like putting lipstick on a pig and calling it a beauty queen. It’s not transformation. Paddy drew a sharp distinction between superficial AI additions and genuine transformation: "Delivering an AI-native application is very different from just slapping AI into an existing application."

Real AI transformation means reimagining entire workflows around AI capabilities. Instead of adding AI features to traditional customer support, companies winning enterprise deals are those that flip the model, so AI handles most tasks while humans provide oversight. 

If you're building features, you're already behind. Build systems.

Create competitive moats through multi-model architecture.

Single LLM strategies are dead on arrival. Winners are orchestrating specialized models:

• General frontier models (OpenAI, Anthropic) for reasoning

• Domain-specific models trained on industry regulations

• Real-time specialist models pulling live market data

“Very quickly we are going to see a chain-of-agents type of model,” Paddy explained. For example, an accounts payable system for shipping companies might use OpenAI for basic reasoning, a specialized model for accounting rules and SEC regulations, and a third for real-time shipping data.

This creates systems that are both more capable and harder to replicate through switching costs that single-model approaches can't match.

Compete on proprietary data, not public datasets.

Many AI companies have exhausted publicly available training data. As Waleed from Mako explained, data scarcity is "preventing existing models from being really good," but three new data frontiers are creating first-mover advantages:

1. Locked business data. Information that's publicly available but not widely indexed, like LinkedIn profiles, MLS listings, and enterprise databases that exist behind access barriers.

2. Real-world capture. Physical devices that collect data from the real world. Think cameras on robots, wearables, and IoT sensors.

3. Industry-specific synthesis: Combined text and language data with visual, audio, and behavioral information to create richer, context-aware datasets

In other words, stop competing for the same training data everyone else uses. Instead, map your vertical's data landscape. What industry databases require paid access? Which real-world sensors could capture behavioral data relevant to your use case? What proprietary datasets do your target customers already collect internally? Focus on sources that would take competitors 12 months or more to replicate.

Build on foundation models; don't compete with them.

Waleed also emphasized where startups should - and should not - focus: The most successful AI companies from the last 12 months are "using APIs and open source models. They're not investing in extremely expensive pre-training."

“Your strategy as a startup absolutely should not be ‘we're going to make a better LLM than OpenAI.’” The economics alone make competing with foundational models nearly impossible. 

Rafael from Apple's engineering team offered another path: “Find that verticalized piece of functionality that will be hard to replicate because of domain knowledge.” In other words, use OpenAI's models as building blocks, but compete on specialized expertise.

Lead with governance transparency.

Nick from PwC's responsible AI practice sees enterprise hesitation about AI adoption firsthand. "Many [enterprises] are reluctant about the AI journey because of concerns around their own data security and accuracy."

Enterprise buyers treat data handling, security protocols, privacy safeguards, and legal compliance as mandatory checkboxes before any technical evaluation begins. Beyond compliance, they worry about operational reliability: "If I'm an accountant using an AI agent to look for fraudulent journal entries, what's the risk that the agent might have bias and miss something?"

Smart founders recognize this as an opportunity. Rather than treating compliance as overhead, they can build what Nick called "transparency around governance and risk management" as core product differentiation. For enterprises evaluating AI solutions, governance capabilities often matter as much as technical performance.

Act now.

Paddy delivered the panel's most urgent message: "The winners and losers are going to be decided in the next 18 months. It's going to be really hard to generate value once you miss this window."

Why the urgency? Right now, "AI is in 'copilot mode' where humans stay in the lead…in 12 months, it's going to be flipped. AI will do most of the work and humans will assist AI."

This shift from human-led to AI-led workflows will fundamentally change how enterprises evaluate and adopt AI solutions. Paddy’s advice to startups is to establish enterprise relationships now while buyers are still experimenting. The companies that do this will own categories when the market consolidates.

Your next steps?

The 18-month window is real. Start with one outsourced function at your target enterprise. Build AI-native workflows, not feature additions. Plan for multi-model architecture from day one. Lead enterprise conversations with governance transparency, not just technical capabilities.

The companies that establish these early enterprise relationships will own categories when the market consolidates. Those who wait will be watching from the sidelines.

Read More