Architecting the Agentic Enterprise | Traversal, Kong, Pigment & Twelve Labs | RAISE 2026

Panel Context & Focus

• The discussion centers on transitioning AI agents from experimental POCs to reliable, autonomous enterprise production in 2026.
• Moderator Kathy Gao (Sapphire Ventures, $11B AUM) frames the core question as "Can agents do anything useful reliably?" rather than "Can they do anything?"
• Five founders represent distinct agentic domains:
  • Anish (Traversal): Agentic Site Reliability Engineer (SRE) for Fortune 500; focuses on autonomous system healing.
  • Carl (Kong): AI connectivity infrastructure for secure, scalable autonomous operations.
  • Eleanor Crespo (Pigment): AI business planning platform unifying teams and agents for governance.
  • Soyoung (12Labs): Multimodal video understanding models for media and security.
  • Shubo (Axiom Math): Mathematical superintelligence for formally proving enterprise system properties.

Pitfalls in POC-to-Production Transition

• Infrastructure & Governance: Enterprises require trusted, governed backends to manage access, validation, and human-in-the-loop thresholds across thousands of users.
• Auditability in High-Stakes Environments: Critical industries (finance, security) demand immediate audit trails to validate agent actions, as errors carry reputational and regulatory risks.
• Security & Actionability: Security teams resist autonomous actions that could destabilize systems; moving from "recommendations" to "system healing" triggers significant friction.
• Cost & Token Management: Daily production use causes significant token spend growth, forcing companies to calculate value and control usage permissions.
• Political Tensions: Discrepancies exist between buyers (seeking labor efficiency) and users (fear of displacement or lack of control).
• Non-Deterministic Failure Modes: Unlike traditional software (binary 500 errors), agents produce "200-type" errors where execution succeeds but outcomes are wrong, requiring full traceability to debug.
• Accountability Gaps: CIOs and CTOs struggle to assign blame for agent errors (e.g., incorrect loan approvals), often stalling deployment on core value-driving tasks.

Strategies for Reliability & Debugging

• Deterministic Layers: Using LLMs to generate code or mathematical proofs (deterministic outputs) rather than direct actions, enabling established debugging frameworks.
• Comparative Evaluation: Replacing absolute scoring models with stable pairwise comparisons to "hill climb" model parameters effectively.
• Human-in-the-Loop Optimization:
  • Low-Risk: Automate fully with full traceability (e.g., board recaps).
  • High-Risk: Maintain human oversight for strategic decisions (e.g., hiring plans, budgeting) where context is critical.
• Monitoring Systems: Systems must flag the 5-10% of failure cases for human review to avoid "alert fatigue," as humans cannot effectively monitor 95% success rates continuously.
• Context Management: Humans remain essential for providing the initial context required for agents to operate trustably.

Business Value & Adoption Statistics

• Gartner Data: 40% of AI initiatives launched in the last six months are projected to fail to deliver business value by 2027.
• McKinsey Data: While 88% of companies use AI, only 5% attribute value to their bottom line or EBITDA.
• Edge vs. Core: Current AI value is concentrated in "edge" use cases; true ROI requires deployment in "core" business operations (e.g., loan processing).
• Uber Case Study: Uber spent $3.4 billion in three months on AI with no realized value, highlighting the risk of unmonitored token spend.
• Cost Trajectory: Token costs are predicted to drop to near electricity costs, shifting the model from experimental edge to core value.

Organizational Structure & Roles

• Responsibility Shift: Digital natives assign agent responsibility to builders; enterprises shift responsibility to the business functions where value is created.
• Emergent "Empiricists": A new class of users (power users) emerges, treating agent interaction as research to discover unanticipated use cases.
• Transformation Roles: Temporary "AI transformation" roles (e.g., AI CFO) are currently accelerating adoption but are expected to disappear as workflows become standard.
• Decentralized Ownership: Long-term ownership of AI systems is expected to migrate to end-user business functions closest to the workflow context.

Pricing & Value Capture Models

• ROI Dual-Lens: Value must be measured via both productivity gains (e.g., 80% efficiency) and decision quality (optimizing margins/supply chain).
• Model Stratification: CFOs will split spend between expensive "frontier" models for high-value tasks and cheaper open-weight models for routine work.
• Outcome-Based Pricing: Shift away from raw token consumption toward fixed pricing or seat-based models guaranteeing specific business outcomes.
• Utility Metrics: Industries like chip verification will price based on "intelligence per dollar" or speed of outcome (e.g., faster verification vs. cost of testing).
• Commitment Models: Adoption of AWS-style large commitments drawn down across various use cases to manage costs.
• Sovereign Infrastructure: Lasting value lies in proprietary data infrastructure optimized for agent querying (indexing, caching), as standard APIs are ill-suited for agent needs.