In an era where AI has become the core engine of business transformation, the challenge for large organizations is not just adopting the technology, but breaking free from the “pilot trap” and scaling AI initiatives into real, value-generating impact.

Join us in exploring the key insights from the session “A Startup-Driven Framework for Prioritizing Enterprise AI Use Cases” at AI-VOLUTION, featuring perspectives from Marc Manara, Head of Startups at OpenAI, and Pailin Vichakul, CIO of SCB 10X. Discover practical ways of thinking that can help your organization prioritize and apply AI to drive real, measurable business outcomes.

"Core vs. Feature": A Key Indicator of Success in Product Development

A major obstacle for most enterprise organizations is the “legacy mindset” — seeing AI as just an add-on to make existing systems slightly better. This stands in contrast to startups, which aren’t burdened by legacy systems and can design products with AI as a core foundation from day one.

Marc highlights the key differences that drive success as follows:

• AI as Core: Most startups begin building their product from scratch with the model as the central heart of the business, not just an add-on. As a result, when OpenAI releases a new model (such as the reasoning models), these startups can immediately integrate the new capabilities into their Product-Market Fit. This often results in an immediate spike in engagement or conversion rates (e.g., free-to-paid conversion).
  
  
• AI as Feature: Large enterprises often try to "bolt AI onto an existing legacy platform." This usually ends up resulting in low-impact features or getting bogged down by technical limitations, causing them to move much slower than their new competitors.

The "Chatbot" Trap to Strategic, Problem-Driven AI Solutions

In the early days of the generative AI era, many organizations fell into the “chatbot trap” — rushing to build basic Q&A bots just to ride the wave and avoid FOMO, without truly considering business value. The result is often a tool that doesn’t really solve customer problems and fails to justify the investment.

However, Marc believes we are now entering a new phase where organizations are starting to find the right path. The focus is shifting away from superficial chatbot projects toward two deeper, high-impact patterns of AI adoption:

1. Internal Productivity: Moving from small pilot groups to putting AI tools in the hands of employees to act as "thought partners" or research analysts. This has now become "table stakes"—a baseline requirement for every organization.
   
   
2. External Use Cases: Moving beyond superficial chatbots to building AI into the core of customer-facing features or data processing pipelines. This delivers new value directly to customers in ways that are difficult for competitors to replicate.

The Expensive Lesson of Using AI the Wrong Way

The most common misconception is viewing AI as a genius employee who knows everything and can provide the perfect answer instantly without context. In reality, LLMs need context to perform well.

Marc cites a Classic Failure Mode found in both startups and enterprises: using models without context.

• The Fail: An executive opens ChatGPT and types, "Please draft a startup team strategy for 2026." The result will be a plan that looks good but is generic because the model doesn't know your key customers, headcount, or internal debates.
  
  
• The Fix: The correct approach is Context Engineering. You must ask: "I have this data and these ideas. Can you critique this argument? Can you add two more ideas to this list based on this analysis?" Bringing the right context to the model is where the magic happens.

From "Demo" to "Production": The Chasm You Must Cross

The world of AI has one iron rule: Coding a cool demo is easy, but going to production is hard.

Many organizations fail in this phase because they celebrate initial results in a controlled environment. However, when faced with real data and unpredictable user behavior, the system fails.

Marc recommends a framework to cross this hurdle, emphasizing "Engineering Rigor":

• Obsess over Evals: The best startups don't just look at the demo; they build Evals (Evaluations) to measure the model across every dimension.
  
  
• The 99% Rule: You might hit a 72% score on your evals in a demo, but you need to be at 99%+ for enterprise use. This requires rigorous tuning.
  
  
• Guardrails & Latency: Building guardrails for consumer-facing apps and optimizing for latency and caching requires high-level engineering precision.
  
  
• Model Fit: Choose the right model for the task. Marc mentions that builders must understand the "shape" and "personality" of each model (e.g., how GPT-5 might differ from GPT-4) to use them effectively.

Buy vs. Build: When "Startup" is the Shortcut for "Enterprise"

Large organizations often fall into the "Not Invented Here" trap. However, in the AI era, trying to build everything yourself can mean wasting time reinventing the wheel.

A key insight is that organizations don't always need to build everything themselves. Marc revealed that many enterprises are simply adopting startup technology to solve specific problems like customer support, contract processing, or threat detection. Adopting technology from specialized startups (like Harvey for legal workflows) is often the smarter and faster strategy for Digital Transformation than building non-core software in-house.

The Best Time to "Build"

Finally, the message Marc leaves us with is the mindset of being a builder. Whether you are a Founder or a Corporate Innovator, this is an incredible time to build something new. The key is to be open to feedback from real-world usage because even OpenAI's research team learns the limits and capabilities of their models through the deployment of these startups.

Taking an organization from Pilot to Scale isn't about who has more data, but about who can adapt and integrate models into the business's core metrics faster.

Watch the full session here: https://youtu.be/E9D0gIqymo4?si=yGHFX8NIFhpmxDA7

For those interested in deep-diving into more techniques, you can follow OpenAI's community at openai.com/startups