Ori, PolyAI, Cresta, Writer & Fireworks AI: Is Enterprise Defensibility Possible in an AI's World?

Core Thesis: AI as the New SaaS and Services Layer

• The panel agrees that software is collapsing into "agentic" surface areas, shifting from traditional SaaS to AI-first solutions.
• Executive decision-making has pivoted: CFOs are rejecting new SaaS purchases and headcount in favor of building capabilities with AI.
• Agentic AI is breaking down silos between systems and teams, allowing agents to touch revenue directly and rebuild tooling landscapes.

Pricing and Delivery Model Disruption

• Seat-based pricing models are collapsing due to customer demand for outcomes rather than software access ("shelfware").
• Customers are shifting toward outcome-based pricing, refusing to pay for unused capacity or future hiring expectations.
• Delivery models are evolving from "handing over software" to "forward-deployed" product managers or engineers who build specific outcomes for the client.
• Specific example: Voice agents (Cresta) can replace thousands of human agents, rendering seat counts irrelevant if the agent performs the work.

Defensibility and Moats

• Consensus that model architecture (even from Frontier Labs like OpenAI) has little inherent defensibility; "no moat" exists in model construction alone.
• The primary moat is the "data flywheel": converting proprietary enterprise data into product improvements that make the model better.
• Success factors for agentic startups include the ability to integrate product analytics into the model to create a self-improving virtual cycle.
• Enterprise defensibility is bifurcated: competing against other enterprises for AI adoption speed, and competing against AI-native disruptors.
• Mere possession of proprietary data is insufficient; fine-tuning open models on private data has failed in cases where it produced worse results than base models.
• Successful cases (e.g., a top UK law firm) combined data with full organizational adoption programs (training 15,000+ lawyers) rather than just technical implementation.
• Lynn (Fireworks AI) notes that misalignment between product data and model development teams creates inefficiencies; closing this loop is critical for quality and cost.

Organizational Adoption and Culture

• Greenfield (new) opportunities are easier than brownfield (legacy) processes; panelists advise reinventing workflows rather than "shifting and lifting."
• Adoption requires a "full organizational strategy," not a separate "AI Guy" or siloed IT function.
• Brinks Home Security is cited as a success story for upfront communication with contact center agents, offering IT career paths to displaced workers.
• Contact center roles are being elevated from cost centers to strategic centers of intelligence, with agents managing "hive minds" and identifying systemic issues (e.g., overcharging) in real-time.
• The pace of AI is compressing career timelines; a contact center manager might present to a board within weeks, a process that previously took decades.

Infrastructure, Open Source, and Sovereignty

• Open Source vs. Proprietary: Fireworks AI and PyTorch alumni advocate for open-source models as a path to community lift, while simultaneously customizing these models with private data to treat them as IP.
• Buy vs. Build: Panelists argue "buy" is the superior strategy for enterprises due to the speed of change; internal build teams often fail to keep pace with the rate of innovation.
• Data Sovereignty: The acquisition of Scale AI by Meta highlights fears of model leakage and competitor data access.
• Sovereignty concerns are shifting from training data to generated data from agent interactions, which could expose sensitive internal communications if breached.
• May Habib (Writer) argues that existing data is less valuable than future models' ability to understand context, suggesting speed of cultural change is the true moat over data ownership.
• Nicola (PolyAI) advises enterprises to "stop thinking about defense" and focus on offense, as defensive postures lead to disruption.

Forward-Looking Statements and Predictions

• 2025 Outlook: Expected to be the year of production-level agentic AI deployments at scale.
• Cost Reduction: Infrastructure costs are expected to drop 10x, not 10%, enabling more viable business models for AI applications.
• Market Expansion: A "tip of the iceberg" of product-market fit applications is emerging; next year will see many more rise above the waterline as cost management improves.
• Revenue Autonomy: The panel predicts the rise of autonomous revenue generation agents within the enterprise.
• Uncertainty: Multiple panelists (May Habib, Lynn, Russell) expressed that predicting the specific trajectory of the technology is impossible due to the rapid, resetting nature of the paradigm.