Conference Presentation, Panel
The Data Problem: What AI Actually Runs On | RAISE Summit 2026
RAISE SummitMatt Fitzpatrick, Philippe Mizrahi, Noam Salinger, Molly Presley, Travers Nisbet, Petrina Anne Steele, Petrina Steele, Emily Fortuna
- AI is projected to transition from its current "first inning" state, with a 10 to 15-year cycle anticipated for full enterprise adoption, while 5% to 8% of current projects reach production.
- The industry split has shifted to 80% inference and 20% training compute, driven by the need to handle legacy data and 20-year-old software rather than training new models from scratch.
- A primary strategic shift involves moving models to data rather than data to models to satisfy sovereign requirements, with indexed deployment on-premise or within cloud infrastructure for regulated industries.
- Retrieval and data quality verification are identified as core bottlenecks, requiring significant work beyond the model itself, particularly for agents relying on web search via LinkUp over the next 20 years.
- Enterprises face the challenge that less than one in five are considered data ready, with most useful data residing in unorganized human knowledge rather than centralized ERP systems.
- Organizations must navigate complex change management, security, and privacy discussions, as the deployment bar requires human-level precision (e.g., 99% accuracy) that is difficult to backtest against statistical baselines.
- Infrastructure constraints are evolving from data to silicon and energy, with supply chain shortages potentially emerging due to high-throughput and low-latency inference demands.
- Future agents are expected to operate 24-7 with increased autonomy, yet success requires maintaining human-in-the-loop workflows rather than relying on off-the-shelf solutions for complex tasks.
- High-quality retrieval is predicted to offer better cost efficiency by utilizing fewer tokens at inference time, while data provenance down to raw information chunks becomes critical for attributing confidence.
- Data preparation cannot wait for finalized AI policies, and building context graphs remains a significant hurdle for most organizations attempting to unlock value in large models.
- The market is characterized by a dichotomy where AI-native companies deploy overnight compared to large banks' long cycles, with barriers to success remaining rooted in data and organizational change five years out.
- Physical AI and synthetic data generated by models are expected to create increasingly complex access challenges, alongside the difficulty of defining "good" output for long-form documents via rubric design.