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
- Core Thesis: Enterprise AI failure is driven 90% by organizational inertia and legacy data issues rather than model selection or data prep alone.
- The World Economic Forum (January 2024) reports less than 20% of organizations consider themselves "data ready."
- Only 5% to 8% of enterprise AI projects successfully reach production; the vast majority stall at the Proof of Concept (POC) stage.
- Primary Organizational Barriers:
- Legacy Infrastructure: 70% of enterprise software is over 20 years old, creating layers of unharmonized data incompatible with modern AI.
- Workflow Misalignment: Organizations attempt to overlay AI onto broken workflows rather than re-engineering processes from the ground up.
- Ownership Deficit: Initiatives often lack "line leader" ownership tied to specific business KPIs (e.g., inventory turns), resulting in isolated technical pilots rather than business transformation.
- Inertia: Companies often wait for AI governance policies before starting data preparation, a mistake that significantly delays deployment.
- Data Strategy Shifts:
- From Training to Inference: The primary data quality challenge has shifted from "training time" curation to "inference time" context retrieval and verification.
- Retrieval Importance: High-quality retrieval is now the critical lever for cost efficiency; better context reduces token usage during inference while improving answer quality.
- Global Namespace: Unifying data silos via a global namespace (rather than ETL copying) allows for policy application (HIPAA, GDPR) and auditability without creating new storage silos.
- Context Graphing: Enterprises often lack "context graphs" to access unstructured data (documents, audio, employee knowledge) necessary for agent functionality.
- Infrastructure & Deployment Realities:
- Compute Split: Global compute allocation has flipped from 80% training/20% inference to 80% inference/20% training within six months.
- Orchestration Needs: Inference requires dynamic data orchestration to move data to where the model is, unlike training where data is local to GPUs.
- Workload Segmentation: Inference is bifurcating into low-latency use cases (e.g., voice AI) requiring sub-second APIs versus high-throughput use cases (e.g., embeddings) requiring batch processing.
- Sovereignty Solutions: Data sovereignty is increasingly addressed by deploying models and indexes on-premise or within specific client clouds (e.g., LinkUp's deployable index) rather than relying on public cloud indexes.
- Quality & Accountability Standards:
- Human Baseline: Agents must meet or exceed human performance levels; organizations often struggle to define "good" for long-form outputs (e.g., legal memos) due to a lack of documented scoring rubrics.
- Responsibility: Humans retain ultimate liability; agents cannot be "fired," necessitating clear human-in-the-loop escalation protocols.
- Validation: Successful deployments require statistical baselines and back-testing capabilities to prove ROI, a hurdle for generative agents compared to traditional machine learning models like credit underwriting.
- Vendor & Customer Segmentation:
- AI Natives (e.g., Notion, Polymarket): Move extremely fast, prioritize performance/accuracy benchmarks, and integrate within days.
- Legacy Enterprises (e.g., Banks): Prioritize security, privacy, data residency, and SLAs; performance is only ~10% of the decision matrix.
- Adoption Speed: Enterprises are adopting vendors faster than historical norms due to acute operational needs, yet transformation cycles remain 10–15 years.
- Future Outlook (5-Year Horizon):
- Continued Bottlenecks: Data quality and organizational change will remain the primary constraints; the conversation will persist as a 2nd or 3rd inning issue.
- Emerging Constraints: The bottleneck is shifting from data and silicon toward energy supply and physical infrastructure limitations.
- Agent Maturity: Agents are evolving from single-task tools to long-running, 24/7 autonomous systems with significant real-world impact, requiring a foundation of verified facts and provenance.